VLab4Mic: prediction of structural resolvability
in super-resolution microscopy
Damián Martínez1,2 , Bruno M. Saraiva1 , Tayla Shakespeare1 , Mark Bates3 , Dylan M. Owen4 , Christophe Leterrier5 , Mario Del
Rosario1,6,7, , and Ricardo Henriques1,
1

Instituto de Tecnologia Química e Biológica António Xavier, Universidade Nova de Lisboa, Oeiras, Portugal
2
Gulbenkian Institute for Molecular Medicine, Oeiras, Portugal
3
Department of Optical Nanoscopy, Institute for Nanophotonics, Göttingen, Germany
4
School of Infection, Inflammation and Immunology, College of Medicine and Health, and School of Mathematics, College of Engineering and Physical Sciences, University of
Birmingham, Birmingham, United Kingdom
5
Aix Marseille Université, CNRS, Institut de Neurophysiopathologie UMR7051, NeuroCyto, Marseille, France.
6
Faculty of Science and Engineering, Cell Biology, Åbo Akademi University, Turku, Finland
7
InFLAMES Research Flagship Centre, University of Turku, Turku, Finland

Determining whether a microscopy experiment can resolve a
specific feature of a protein assembly remains difficult because
researchers must balance imaging modality, labelling strategy,
and probe choice. We present VLab4Mic, a simulation platform
that predicts structural resolvability before experiments. Starting from atomic models from the PDB or AlphaFold predictions,
VLab4Mic places antibodies, nanobodies, chemical linkers, or
fluorescent proteins on epitopes, applies stochastic labelling and
steric constraints, and generates virtual samples for widefield,
confocal, AiryScan, Stimulated Emission Depletion (STED), and
Single-Molecule Localisation Microscopy (SMLM). Simulations
of nuclear pore complexes agree with experimental data across
all five modalities. In case studies, HIV capsid appearance depended strongly on orientation, and STED and SMLM separated
domed from flat clathrin lattices where confocal and AiryScan
did not. VLab4Mic therefore lets researchers judge which biological questions a given imaging configuration can answer before
spending time tuning parameters at the microscope.
super-resolution microscopy | microscopy simulation | structural biology |
fluorescence labelling | protein structure | experimental design | computational
modelling
Correspondence: (M. D. Rosario) mario.delrosariominina@abo.fi; (R. Henriques) r.henriques@itqb.unl.pt

Main
Super-resolution microscopy (SRM) has bridged the gap between cellular imaging and structural biology, enabling visualisation of macromolecular assemblies inside cells (2, 3). Although modern optics can localise a fluorophore with nanometre precision, high optical resolution alone does not guarantee
that the underlying biological structure is experimentally distinguishable. Recovering a biologically meaningful feature
further depends on probe geometry, linkage error, labelling
stochasticity, molecular organisation, and imaging modality.
The central challenge is therefore not simply achieving high
resolution, but determining whether a given experimental configuration can faithfully preserve biologically relevant structural information, a property we call structural resolvability.
This captures the ability to detect specific biological features
despite stochastic fluorophore binding and the physical dimensions of the probes used, presupposing that the structure is
detectable above noise; this sensitivity is governed by photon budget, labelling density, and detector noise (modelled in
Supp. Note 2).

Fig. 1. VLab4Mic workflow. a) Structure and probe selection. Users can
select or provide PDB/CIF structures from databases like PDB or AlphaFold. These
files contain atomic coordinates shown in grey. VLab4Mic parses this data to map
specific epitope sites in blue. A chosen structural model is used to build a probe.
For example, an antibody probe is shown with four conjugated fluorophores in
magenta. b) Epitope labelling. The platform models the binding of probes to
available epitopes. Users can model structure heterogeneity first to restrict available
binding sites (step 1). Probes are repositioned toward available epitopes (step 2).
Probe constraints may cause incomplete labelling (step 3). The final labelled particle
becomes a rigid entity. It can rotate (step 4) and translate (step 5) independently.
c) Virtual sample generation. Labelling is a stochastic process. A single probe
and structure combination can create multiple unique labelled structures. These
structures are arranged in space to build a virtual sample. d) Imaging simulation.
Virtual samples act as the ground truth for imaging simulations. The software
simulates modalities like widefield, confocal, AiryScan, STED, and SMLM. Additional
pipelines quantify imaging features, such as the ability to resolve substructures or
discriminate particles. Scale bars: 10 nm (a, b); 100 nm (d).

Choosing between antibodies, nanobodies, or chemical tags,
and matching them to an imaging modality, is usually a matter
of costly trial and error. Current simulation tools for SMLM
experiments have addressed critical aspects that affect the quality of acquired data such as label geometry, labelling efficiency
and fluorophore photophysics (Table S6), and can provide a
"ground truth" to benchmark experimental designs and algorxiv-maker |

Martínez et al.

|

June 22, 2026

|

1–31

Fig. 2. Comparison of VLab4Mic imaging simulations with experimental data. VLab4Mic imaging simulations based on available experimental images of Nuclear
Pore Complexes (1). For each modality, the positions of NPCs in the virtual sample were estimated as the local maxima peaks from its corresponding experimental
example (See Methods). Scale bars: 500 nm.

rithms (4–9). Other simulators instead start from molecular
kinetics, for example to test whether transcription-factor clustering reflects genuine condensates or imaging effects (10).
While specific studies have used microtubules (11) or nuclear
pore complexes (12) with structural detail, general-purpose
simulators typically oversimplify the biological sample, representing structures as an idealised geometric shape. This
disconnect means that questions critical to structural biologists, such as whether a structural change can be resolved with
a primary-secondary antibody pair or how structural integrity
affects the ability to measure pore diameter, cannot be reliably
answered in silico.
Starting from atomic-resolution models, VLab4Mic models
how molecular geometry, epitope accessibility, probe architecture, steric constraints, and stochastic labelling together
determine whether a biological structure stays experimentally
resolvable. This atomic input is essential for super-resolution
modalities, and remains informative for diffraction-limited
modalities by capturing how crowding and orientation effects
influence apparent intensity. By applying physically realistic labelling algorithms across widefield, confocal, AiryScan,
STED (13), and SMLM (14–16) modalities, VLab4Mic captures the steric hindrance and binding heterogeneity that dictate structural resolvability.
This framework separates resolvability from sensitivity: resolvability asks whether a specific biological feature, such
as a ring radius or a curvature transition, is recoverable from
the simulated image, whereas sensitivity governs general detectability above noise. The two are not orthogonal in practice
(image-similarity metrics like SSIM remain coupled to noise),
but VLab4Mic exposes them as separately controllable inputs
(probe geometry and labelling efficiency on the resolvability
side; photon budget and detector model on the sensitivity side),
enabling biological-feature hypothesis testing rather than mere
instrument tuning. Researchers can pose concrete structural
2

questions in silico to estimate whether functionally distinct
structural states are likely to remain experimentally distinguishable under a proposed experimental setup; for example, users can generate testable predictions regarding whether
domed and flat clathrin lattices are structurally resolvable. By
predicting how biology, chemistry, and physics shape the final
image, VLab4Mic supports experimental design in structural
cell biology.
Imaging experiments in VLab4Mic mirror the stages of experimental design: selecting a structure of interest, preparing
the sample by selecting a probe, choosing imaging techniques,
and analysing results. The workflow (Fig. 1) begins with the
selection of an atomic structure from the Protein Data Bank
(17) and the specification of a labelling strategy (antibody,
nanobody, or chemical linker), which together define target
binding sites and probe spatial properties (Fig. 1a). This approach generates a structural model with precise fluorophore
positioning, accounting for constraints imposed by both the
structure and labelling strategy. The labelling model incorporates stochastic elements such as incomplete binding and structural heterogeneity (Fig. 1b). Multiple independent copies of
labelled structures are distributed in three-dimensional space
to form a virtual sample. Imaging is then simulated through
a forward model that convolves the fluorophore distribution
with an effective post-reconstruction PSF per modality (widefield, confocal, AiryScan, STED; modality-specific σ in Table
S3), whereas SMLM is instead generated by sampling each
emitter from a Gaussian at its localisation precision (see Methods). VLab4Mic therefore produces the image a researcher
would obtain after any modality-specific reconstruction step,
not raw camera frames fed through a reconstruction pipeline
(Fig. 1c). A validation framework then systematically compares the simulated images to reference features to quantify
structural resolvability (Fig. 1d). The complete computational
pipeline is detailed in Fig. S1 (Sup. Video 1).
Martínez et al.

|

VLab4Mic

Each probe is represented by the spatial distribution of its
fluorophores (magenta) relative to the probe structure (grey)
and an anchor at the target binding site (Fig. S2), capturing
the epitope-to-fluorophore distance that sets the achievable
localisation precision (probe specifications in Table S1). For
diffraction-limited modalities (widefield, confocal, AiryScan)
this geometry effectively reduces to the probe centre, but
it becomes important for SMLM and other super-resolution
techniques, where probe-induced linkage error sets a floor on
the apparent feature size.
Probes are positioned on epitopes by a rigid-body algorithm
that orients each along the local epitope normal vector, with
a wobbling term modelling linker flexibility. Multiple copies
are placed to build a labelled structure (Fig. S8, Fig. S2; algorithm detailed in Methods). Primary-secondary probe pairs are
also supported, with secondaries binding the structure-bound
primaries under their own independent constraints to capture
the larger linkage error of indirect labelling (Fig. 3c, Fig. S6;
procedure in SI Methods).
To test the forward model against experimental data, we used
NPCs, a widely used nanoscale reference standard for quantitative super-resolution microscopy with experimentally validated dimensions. We generated virtual samples matching the
structure, epitope site, probe, and imaging parameters reported
by Thevathasan et al. (1) (see SI Methods). Because particle
positions were primed from the local maxima of the corresponding experimental images (seeded at the same positions
as the real individual NPC structures), this is a consistency
check on per-particle appearance rather than a prediction of the
field layout (Supp. Note 3). Across all five modalities, from
diffraction-limited widefield through super-resolved SMLM,
the simulated images reproduce the appearance of the experimental data (Fig. 2). Quantitatively, apparent NPC radii
measured by image-based circle-fitting on simulated and experimental SMLM data share the same bimodal distribution,
with a primary mode at approximately 52 nm and a secondary
mode at approximately 65 nm (Fig. S10). The primary mode
found in our radii estimations closely matches the reported
Nup96 radius of 53.7 ± 2.1 nm (1). This recovery should be
interpreted as a consistency check rather than an independent
validation of the NPC radius. The informative result is the
secondary mode: our simulated NPCs are identical in every
particle by construction, yet an image-based circle fit recovers the same second mode from the simulation as from the
experimental images, showing that the 65 nm mode does not
reflect a distinct NPC subpopulation (circle-fit procedure in
Supp. Note 3).
VLab4Mic can also sweep parameters automatically: a branching workflow generates all combinations of structural properties (labelling efficiency, structural completeness) and detection characteristics (imaging modality, microscope settings)
from user-defined ranges, scoring each against a reference and
summarising the space as heat maps (Fig. 3c,d; Sup. Video 2).
This replaces manual parameter testing with quantitative metrics for comparing experimental configurations (tested ranges
in Table S2; per-cell SSIM/Pearson grid in Fig. 3d).
Martínez et al.

|

VLab4Mic

Fig. 3. Parameterisation of Virtual Sample, Labelled Structure and Parameter
sweeps using HIV capsid model. a) Virtual sample parameterisation. Schematic
depicting how a labelled particle, such as HIV capsids, can be independently parameterised; for instance, with different orientations or positions relative to focal plane.
Magenta dots represent fluorophores. b) Axial orientation and position sweep.
SMLM image simulation of a virtual sample where a labelled particle’s axial orientation and axial position varies along the X and Y axis, respectively. c) Parameter
branching. The HIV capsid structure (PDB: 3J3Y, grey dots in atom locations) is
labelled with different probes targeting the p24 Capsid protein; epitope sites in blue,
fluorophores in magenta. Labelling efficiency (100%, 50%, 20%), structural integrity
(100%, 60%, 40%), and imaging modality (STED, SMLM, AiryScan) are varied
independently. A dotted line indicates the expected cone shape for structural integrity examples. Secondary labelling is also supported, with primary and secondary
probes independently parameterised. d) Parameter sweep. VLab4Mic sweeps
user-defined parameter ranges (here, labelling efficiency and structural integrity at
5 values from 0–100%) and compares each combination to a reference structure
via similarity metrics. High-similarity regions indicate experimentally recoverable
states; low-similarity regions indicate loss of structural distinguishability. Results
are summarised as a heat map table. e) Example sweep images. Representative
images from the parameter sweep in d).

Virtual samples can be systematically parameterised by particle position relative to the focal plane and by rotational orientation (Fig. 3a,b; Fig. S4; Fig. S5), enabling quantification of
how 3D sample geometry affects 2D image formation. This
addresses a recurring difficulty in comparing simulated and
experimental images: experimental imaging captures random
particle orientations and axial positions, whereas conventional
3

simulations often depict ideally oriented structures. By generating distributions of orientations and depths, VLab4Mic
produces image ensembles that reflect this experimental heterogeneity.
Accurate mapping of a structure is limited by the effective
labelling efficiency, which is a combination of the number
and disposition of the available targets, and the probe binding
affinity. The parameter sweep workflow quantifies the tradeoffs between labelling completeness and structural integrity
(Fig. 3e), reporting structural resolvability as the Structural
Similarity Index (SSIM) and Pearson correlation between
each simulated image and a fully-labelled, fully-intact reference (definitions in Supp. Note 3, Eq. 14 and Eq. 15; the
tested parameters for HIV-1 capsid sweep are described in
Supplementary Notes). We note that this reference is itself
simulated and so SSIM/Pearson here measure agreement with
the model’s idealised assumptions rather than with biological
ground truth; for biological-truth comparisons VLab4Mic supports task-specific metrics (radial distribution analysis for ring
structures, Fourier-space periodicity for lattices) discussed
below. At high labelling efficiency and complete structural
integrity, the simulated cone shape of the HIV-1 capsid recovers the reference (Fig. 3e top left). SSIM declines as either
parameter degrades: at 50% structural integrity with complete
labelling, the cone outline is preserved while internal density is distorted, and at jointly low labelling efficiency and
structural integrity the cone shape is no longer visually discernible (Fig. 3e bottom panels). Both labelling efficiency and
structural integrity therefore need weighing when designing
an experiment, since either alone can render a key feature
unresolvable. Structural integrity modelling algorithms and
validation are detailed in Supp. Note 1 and Fig. S3.
VLab4Mic applies to diverse biological structures and labelling strategies (Fig. S13), and makes it straightforward
to add custom structural models for resolvability assessment
(Fig. 4). As an example, we used ChimeraX (19) to assemble clathrin-coated structures (CCS) as either domed or flat
lattices (Fig. 4a), two conformations that coexist on cellular
membranes and are implicated in clathrin-mediated endocytosis (18). Because the exact arrangement of triskelia within
domes and plaques is not yet known (20), the models shown
here were assembled by us as illustrative examples, guided
by the PREM reference (Fig. 4b). Each CCS model was
equally labelled and simulated under four imaging modalities.
VLab4Mic predicts that confocal and AiryScan do not visibly
distinguish the two conformations, whereas STED and SMLM
resolve them as distinct geometries (Fig. 4).
VLab4Mic also accepts AlphaFold predictions as input (Fig.
S11; Sup. Video 3), which is particularly useful for targets lacking experimental structures where resolvability questions arise before crystallography or cryo-EM is attempted.
For structures where individual subunits need to be distinguished rather than averaged, VLab4Mic supports chainspecific multichannel labelling (Fig. S12). Using PCNA as a
protein nanoruler with the same residue targeted in each chain,
VLab4Mic predicts that SMLM with effective σ ≈ 2 nm does
4

Fig. 4. Building and testing custom structural models with ChimeraX and
VLab4Mic. Structures assembled in external software can be imported into
VLab4Mic to predict how distinct conformational states appear under different imaging modalities. a) Triskelion models. Triskelion building blocks used to create
composite models were derived from a related clathrin coat assembly (3 copies of
the heavy chain from the 3IYV model were selected and arranged to create each
triskelion). b) Composite CCS models. Clathrin Coated Structures (CCSs) assembled in ChimeraX as a dome (top) or flat lattice (bottom), placing copies of a
PDB to match a Platinum Replica Electron Microscopy (PREM) reference (18) (see
methods). Models were exported as CIF for VLab4Mic and are available on Zenodo
(see Data availability section). c) Overlay with reference PREM. Composite models
overlaid on the reference PREM image. d) Structure preview. Composite models
previewed in VLab4Mic. A small fraction of all the atoms is used for visualisation. e)
Simulations across modalities. VLab4Mic simulations of the dome and flat lattice
under four imaging modalities. Both models were labelled with the same antibody
probe targeting the light chain (see SI Methods). The dome and flat lattice are
predicted to be distinguishable in STED and SMLM but not in confocal or AiryScan.
Scale bars: 10 nm (a); 100 nm (b, c); 200 nm (d, e).

not resolve the 6 nm inter-site spacing, while a custom SMLM
modality with σ ≈ 1 nm does; multichannel labelling breaks
the trimer’s apparent symmetry and recovers per-chain topology in the simulation even when single-channel resolution is
limiting. These examples illustrate how VLab4Mic can predict whether a specific structural question is experimentally
tractable with a given imaging modality.
VLab4Mic provides a structure-agnostic forward model that
integrates structural biology data with realistic labelling and
imaging models, and complements other computational tools
(4, 5, 11, 12). Its key advance is the in silico capacity to
ask whether a structural feature can be resolved before any
experiment, testing the resolving power of different probe,
labelling and imaging settings.
The model reproduces the second mode of the apparent NPC
ring-radius distribution from a population of identical particles, so this mode does not reflect a distinct NPC subpopulation. The framework therefore captures key sources of
variability in structural super-resolution imaging (21). Scoring how labelling and acquisition parameters affect structural
fidelity also speaks to the wider problem of quantifying imaging artefacts (22, 23). By moving optimisation into a virtual
environment, VLab4Mic can reduce trial-and-error and unnecessary acquisition time.
Martínez et al.

|

VLab4Mic

VLab4Mic addresses practical experimental design questions:
which modality and labelling strategy can resolve a given feature, how much displacement a primary-secondary antibody
pair introduces, or when a design will fail outright, as the
orientation-dependent HIV-1 capsid case illustrates (Fig. 3).
It is not intended to replace experimental validation, but to
guide design toward configurations most likely to preserve
biologically meaningful structural information.
To stay tractable across all five modalities, the forward model
makes deliberate simplifications: structures are static, point
spread functions are analytical, and detector noise follows
a single camera-style model (quantum efficiency, stochastic
gain, Gaussian readout, and digitisation) applied to every
modality (Supp. Note 2). In SMLM, each emitter is rendered
at its localisation precision rather than as a series of blinking
events; in the diffraction-limited modalities (widefield, confocal, AiryScan), the fluorophores within a single probe fall well
inside the blur and are averaged to a single point. The model
is therefore most informative when image contrast is governed
by particle orientation and molecular crowding rather than by
photon statistics.
The scope of these claims is bounded by what the model takes
in versus what it produces. VLab4Mic assumes the structure,
probe geometry, labelling statistics, and detector parameters
are user-specified, and predicts the resulting image and which
features survive it; it does not infer unknown structure. A
match with an experimental image therefore confirms only
that the inputs and the image-formation model are mutually
consistent, not that either is biologically correct. The model
would be falsified if, for a target with independently known
ground truth, the predicted resolvability of a feature disagreed
with experiment; closing that loop prospectively would upgrade these consistency checks to predictive validation.
Because VLab4Mic generates synthetic images with known
atomic ground truth, it complements curated experimental
SMLM resources (24) and to emerging deep-learning analyses that require systematic benchmark generation, including contrastive-embedding methods for nanoscale spatialomics (25). It integrates with established analysis frameworks
(26, 27) and broader deep-learning workflows (28, 29). As
super-resolution microscopy moves toward molecular-scale
imaging (30–34), frameworks that test whether a structure is
not just visible but resolvable will matter more for experimental design and quantitative structural biology.
ABOUT THIS MANUSCRIPT
This manuscript was prepared using Rχ iv-Maker v1.22.0 (35).
DATA AVAILABILITY
Structural models are publicly available from the RCSB Protein Data Bank (https:
//www.rcsb.org): 7R5K (Nuclear Pore Complex), 3J3Y (HIV-1 capsid), 1XI5 (Clathrin),
3IYV (clathrin heavy chains used for creation of the dome and flat-lattice assemblies
in ChimeraX), 8GMO (T4 phage). Experimental NPC images and localisation tables
used for the consistency check against published NPC geometry are deposited in
BioStudies under accession S-BIAD8 (https://www.ebi.ac.uk/biostudies/bioimages/
studies/S-BIAD8) (36), associated with the original publication (1). Supplementary
data generated for this manuscript is available through Zenodo (https://doi.org/10.
5281/zenodo.20377069). All simulation parameters and configuration files used in
this manuscript are provided in the VLab4Mic repository examples directory.
CODE AVAILABILITY

Martínez et al.

|

VLab4Mic

VLab4Mic source code, installation instructions, API documentation, and example notebooks are available at https://github.com/HenriquesLab/VLab4Mic (v0.1.0)
under the MIT licence. VLab4Mic requires Python 3.10-3.13 and runs on macOS, Linux, and Windows. LabConstrictor executables are available at https:
//github.com/HenriquesLab/LabConstrictor-VLab4Mic. All code used to generate figures and analysis in this manuscript is provided in the repository examples directory.
AUTHOR CONTRIBUTIONS
Conceptualisation: D.M., B.M.S., M.D.R., R.H. Methodology: D.M., B.M.S., T.S.,
M.D.R., R.H. Software: D.M., B.M.S., T.S. Investigation: D.M., B.M.S., T.S. Formal
analysis: D.M. Resources: B.M.S. Data curation: D.M. Visualisation: D.M.
Writing - original draft: D.M., B.M.S., M.D.R., R.H. Writing - review editing:
D.M., B.M.S., T.S., M.B., D.M.O., C.L., M.D.R., R.H. Supervision: B.M.S., M.D.R.,
R.H. Project administration: B.M.S., M.D.R., R.H. Funding acquisition: R.H. All
authors read and approved the final manuscript.
ACKNOWLEDGEMENTS
R.H. acknowledges funding from the European Research Council (ERC) under
the European Union’s Horizon 2020 research and innovation programme (SelfDriving4DSR, grant agreement No. 101001332). R.H. is further supported by
the European Union through Horizon Europe (RT-SuperES, grant agreement No.
101099654); by a European Molecular Biology Organization (EMBO) Installation Grant (EMBO-2020-IG-4734); by a Chan Zuckerberg Initiative Visual Proteomics Imaging award (vpi-0000000044, https://doi.org/10.37921/743590vtudfp);
by a joint Wellcome, Chan Zuckerberg Initiative, and Kavli Foundation Essential
Open Source Software for Science Cycle 6 award (Wellcome 313383/Z/24/Z; CZI
EOSS6-0000000260); and by the La Caixa Foundation (CaixaResearch Health 2025,
VirusAwareScopes, HR25-00453). This work was supported by FCT – Fundação
para a Ciência e a Tecnologia, I.P., through the MOSTMICRO-ITQB RD Unit (DOI
10.54499/UID/04612/2025, UID/PRR/4612/2025) and the LS4FUTURE Associated
Laboratory (DOI 10.54499/LA/P/0087/2020). The Chan Zuckerberg Initiative awards
are made through the Chan Zuckerberg Initiative DAF, an advised fund of Silicon
Valley Community Foundation. D.M.O. acknowledges funding from the Biotechnology and Biological Sciences Research Council (BBSRC, grant BB/X018644/1).
M.D.R acknowledges the Center of Excellence IMMENS funded by the Research
Council of Finland (374180). C.L. acknowledges funding from the French National
Research Agency (ANR-24-CE13-7996 “MicRON”), and from Excellence Initiative
of Aix Marseille University - A∗MIDEX (DINoMIR AMIDEX AMX-22-RE-AB-137 to
C.L., NeuroMarseille AMX-19-IET-004). We would like to thank the Neuro-Cellular
Imaging Service and Nikon Center for Neuro-NanoImaging at INP, with funding from
CPER-FEDER (Plateforme NeuroTimone PA0014842), as well as the Institut NeuroMarseille (AMX-19-IET-004) and the Institut Marseille Imaging (AMX-19-IET-002) for
complementary equipment funding from Excellence Initiative of Aix-Marseille University – A∗MIDEX, a French “Investissements d’Avenir” program (ANR-11-IDEX-0001).
Funded by the European Union. Views and opinions expressed are, however, those
of the authors only and do not necessarily reflect those of the European Union or the
granting authority. Neither the European Union nor the granting authority can be held
responsible for them. The authors thank all members of the Henriques laboratory for
helpful discussions, Iván Hidalgo-Cenalmor and Guillaume Jacquemet for their help
with the LabConstrictor implementation, and Mike Heilemann for helpful feedback on
the manuscript, and Stéphane Vassilopoulos for feedback on the clathrin models.
COMPETING FINANCIAL INTERESTS
The authors declare no competing interests.
EXTENDED AUTHOR INFORMATION
• Damián Martínez:
0000-0002-5906-598X
• Bruno M. Saraiva:
0000-0002-9151-5477
• Tayla Shakespeare:
0000-0002-4159-0460
• Mark Bates:
0000-0003-0668-5277
• Dylan M. Owen:
0000-0002-5284-2782
• Christophe Leterrier:
0000-0002-2957-2032
• Mario Del Rosario:
0000-0002-0430-1463
• Ricardo Henriques:
0000-0002-2043-5234

Bibliography
1. Jervis Vermal Thevathasan, Maurice Kahnwald, Konstanty Cieśliński, Philipp Hoess, Sudheer Kumar Peneti, Manuel Reitberger, Daniel Heid, Krishna Chaitanya Kasuba, Sarah Janice
Hoerner, Yiming Li, Yu-Le Wu, Markus Mund, Ulf Matti, Pedro Matos Pereira, Ricardo Henriques, Bianca Nijmeijer, Moritz Kueblbeck, Vilma Jimenez Sabinina, Jan Ellenberg, and

5

2.

3.

4.

5.

6.

7.

8.

9.

10.
11.

12.

13.

14.

15.

16.

17.

18.

19.

20.

21.
22.

23.

24.

6

Jonas Ries. Nuclear pores as versatile reference standards for quantitative superresolution microscopy. Nature Methods, 16(10):1045–1053, October 2019. ISSN 1548-7105.
doi:10.1038/s41592-019-0574-9.
Lothar Schermelleh, Alexia Ferrand, Thomas Huser, Christian Eggeling, Markus Sauer, Oliver
Biehlmaier, and Gregor P. C. Drummen. Super-resolution microscopy demystified. Nature
Cell Biology, 21(1):72–84, January 2019. ISSN 1476-4679. doi:10.1038/s41556-018-0251-8.
Mickaël Lelek, Melina T. Gyparaki, Gerti Beliu, Florian Schueder, Juliette Griffié, Suliana
Manley, Ralf Jungmann, Markus Sauer, Melike Lakadamyali, and Christophe Zimmer. Singlemolecule localization microscopy. Nature Reviews Methods Primers, 1(1):39, June 2021.
ISSN 2662-8449. doi:10.1038/s43586-021-00038-x.
Dominique Bourgeois. Single molecule imaging simulations with advanced fluorophore
photophysics. Communications Biology, 6(1):53, January 2023. ISSN 2399-3642.
doi:10.1038/s42003-023-04432-x.
J. Griffié, T. A. Pham, C. Sieben, R. Lang, V. Cevher, S. Holden, M. Unser, S. Manley, and
D. Sage. Virtual-SMLM, a virtual environment for real-time interactive SMLM acquisition,
March 2020. doi:10.1101/2020.03.05.967893.
Tibor Novák, Tamás Gajdos, József Sinkó, Gábor Szabó, and Miklós Erdélyi. TestSTORM:
Versatile simulator software for multimodal super-resolution localization fluorescence microscopy. Scientific Reports, 7(1):951, April 2017. ISSN 2045-2322. doi:10.1038/s41598017-01122-7.
Varun Venkataramani, Frank Herrmannsdörfer, Mike Heilemann, and Thomas Kuner.
SuReSim: Simulating localization microscopy experiments from ground truth models. Nature
Methods, 13(4):319–321, April 2016. ISSN 1548-7105. doi:10.1038/nmeth.3775.
Jonas Ries. SMAP: A modular super-resolution microscopy analysis platform for SMLM data.
Nature Methods, 17(9):870–872, September 2020. ISSN 1548-7105. doi:10.1038/s41592020-0938-1.
Daniel Sage, Thanh-An Pham, Hazen Babcock, Tomas Lukes, Thomas Pengo, Jerry Chao,
Ramraj Velmurugan, Alex Herbert, Anurag Agrawal, Silvia Colabrese, Ann Wheeler, Anna
Archetti, Bernd Rieger, Raimund Ober, Guy M. Hagen, Jean-Baptiste Sibarita, Jonas Ries,
Ricardo Henriques, Michael Unser, and Seamus Holden. Super-resolution fight club: Assessment of 2D and 3D single-molecule localization microscopy software. Nature Methods,
16(5):387–395, May 2019. ISSN 1548-7105. doi:10.1038/s41592-019-0364-4.
Manya Kapoor and Mustafa Mir. Simulation of transcription factor clustering in nuclei from
molecular kinetics. bioRxiv, 2026. doi:10.64898/2026.06.04.730171. Preprint.
Susanna M. Früh, Ulf Matti, Philipp R. Spycher, Marina Rubini, Sebastian Lickert, Thomas Schlichthaerle, Ralf Jungmann, Viola Vogel, Jonas Ries, and Ingmar
Schoen. Site-Specifically-Labeled Antibodies for Super-Resolution Microscopy Reveal
In Situ Linkage Errors. ACS Nano, 15(7):12161–12170, July 2021. ISSN 1936-0851.
doi:10.1021/acsnano.1c03677.
Maria Theiss, Jean-Karim Hériché, Craig Russell, David Helekal, Alisdair Soppitt, Jonas
Ries, Jan Ellenberg, Alvis Brazma, and Virginie Uhlmann. Simulating structurally variable
nuclear pore complexes for microscopy. Bioinformatics, 39(10):btad587, October 2023. ISSN
1367-4811. doi:10.1093/bioinformatics/btad587.
Stefan W. Hell and Jan Wichmann. Breaking the diffraction resolution limit by stimulated
emission: Stimulated-emission-depletion fluorescence microscopy. Optics Letters, 19(11):
780–782, June 1994. ISSN 1539-4794. doi:10.1364/OL.19.000780.
Eric Betzig, George H. Patterson, Rachid Sougrat, O. Wolf Lindwasser, Scott Olenych,
Juan S. Bonifacino, Michael W. Davidson, Jennifer Lippincott-Schwartz, and Harald F. Hess.
Imaging Intracellular Fluorescent Proteins at Nanometer Resolution. Science, 313(5793):
1642–1645, September 2006. doi:10.1126/science.1127344.
Michael J. Rust, Mark Bates, and Xiaowei Zhuang. Sub-diffraction-limit imaging by stochastic
optical reconstruction microscopy (STORM). Nature Methods, 3(10):793–796, October 2006.
ISSN 1548-7105. doi:10.1038/nmeth929.
Samuel T. Hess, Thanu P. K. Girirajan, and Michael D. Mason. Ultra-High Resolution Imaging
by Fluorescence Photoactivation Localization Microscopy. Biophysical Journal, 91(11):4258–
4272, December 2006. ISSN 0006-3495, 1542-0086. doi:10.1529/biophysj.106.091116.
Helen M. Berman, John Westbrook, Zukang Feng, Gary Gilliland, T. N. Bhat, Helge Weissig,
Ilya N. Shindyalov, and Philip E. Bourne. The Protein Data Bank. Nucleic Acids Research,
28(1):235–242, January 2000. ISSN 0305-1048. doi:10.1093/nar/28.1.235.
Kazuki Obashi, Kem A. Sochacki, Marie-Paule Strub, and Justin W. Taraska. A conformational switch in clathrin light chain regulates lattice structure and endocytosis at the plasma
membrane of mammalian cells. Nature Communications, 14(1):732, February 2023. ISSN
2041-1723. doi:10.1038/s41467-023-36304-7.
Thomas D. Goddard, Conrad C. Huang, Elaine C. Meng, Eric F. Pettersen, Gregory S. Couch,
John H. Morris, and Thomas E. Ferrin. UCSF ChimeraX: Meeting modern challenges in
visualization and analysis. Protein Science: A Publication of the Protein Society, 27(1):
14–25, January 2018. ISSN 1469-896X. doi:10.1002/pro.3235.
Kem A. Sochacki, Bridgette L. Heine, Gideon J. Haber, John R. Jimah, Bijeta Prasai, Marco A. Alfonzo-Méndez, Aleah D. Roberts, Agila Somasundaram, Jenny E. Hinshaw, and Justin W. Taraska. The structure and spontaneous curvature of clathrin lattices at the plasma membrane. Developmental Cell, 56(8):1131–1146.e3, April 2021.
doi:10.1016/j.devcel.2021.03.017.
John S. H. Danial. Super-resolution microscopy for structural biology. Nature Methods, 22
(8):1636–1652, August 2025. ISSN 1548-7105. doi:10.1038/s41592-025-02731-1.
Siân Culley, David Albrecht, Caron Jacobs, Pedro Matos Pereira, Christophe Leterrier, Jason
Mercer, and Ricardo Henriques. Quantitative mapping and minimization of super-resolution
optical imaging artifacts. Nature Methods, 15(4):263–266, April 2018. ISSN 1548-7105.
doi:10.1038/nmeth.4605.
Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment:
From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):
600–612, 2004. doi:10.1109/TIP.2003.819861.
Surbhi Shirgill, Daniel J. Nieves, Jeremy Pike, Mohamed Ahmed, Helen Abbott, Mohammed
Baragilly, Karoll Savoye, Mark C. Wales, Adam Kaminer, Ruby Peters, Ricardo Henriques,
Steven F. Lee, Patrick Rubin-Delanchy, and Dylan M. Owen. Nano-org, a functional resource
for single-molecule localisation microscopy data. Nature Communications, 16:8674, 2025.
doi:10.1038/s41467-025-63674-x.

25. Surbhi Shirgill, L. G. Jensen, Daniel J. Nieves, Mark C. Wales, Adam Kaminer, Karoll Savoye,
Ruby Peters, Mike Heilemann, Ricardo Henriques, Steven F. Lee, Patrick Rubin-Delanchy,
and Dylan M. Owen. Nanoscale spatial-omics via contrastive embedding of single-molecule
localisation data. bioRxiv, 2025. doi:10.1101/2025.10.07.679170. Preprint.
26. Romain F. Laine, Kalina L. Tosheva, Nils Gustafsson, Robert D. M. Gray, Pedro Almada,
David Albrecht, Gail Tournaviti Risa, Fredrik Hurtig, Ann-Christin Lindås, Buzz Baum, Jason
Mercer, Christophe Leterrier, Pedro M. Pereira, Siân Culley, and Ricardo Henriques. NanoJ:
A high-performance open-source super-resolution microscopy toolbox. Journal of Physics
D: Applied Physics, 52(16):163001, February 2019. ISSN 0022-3727. doi:10.1088/13616463/ab0261.
27. Yu-Le Wu, Philipp Hoess, Aline Tschanz, Ulf Matti, Markus Mund, and Jonas Ries. Maximumlikelihood model fitting for quantitative analysis of SMLM data. Nature Methods, 20(1):
139–148, 2023. doi:10.1038/s41592-022-01676-z.
28. Lucas von Chamier, Romain F. Laine, Johanna Jukkala, Christoph Spahn, Daniel Krentzel,
Elias Nehme, Martina Lerche, Sara Hernández-Pérez, Pieta K. Mattila, Eleni Karinou, Seámus Holden, Ahmet Can Solak, Alexander Krull, Tim-Oliver Buchholz, Martin L. Jones,
Loic A. Royer, Christophe Leterrier, Yoav Shechtman, Florian Jug, Mike Heilemann, Guillaume Jacquemet, and Ricardo Henriques. Democratising deep learning for microscopy
with ZeroCostDL4Mic. Nature Communications, 12(1):2276, April 2021. ISSN 2041-1723.
doi:10.1038/s41467-021-22518-0.
29. Artur Speiser, Lucas-Raphael Müller, Philipp Hoess, Ulf Matti, Christopher J. Obara, Wesley R. Legant, Anna Kreshuk, Jakob H. Macke, Jonas Ries, and Srinivas C. Turaga. Deep
learning enables fast and dense single-molecule localization with high accuracy. Nature
Methods, 18(9):1082–1090, 2021. doi:10.1038/s41592-021-01236-x.
30. Sheng Liu, Philipp Hoess, and Jonas Ries. Super-Resolution Microscopy for Structural
Cell Biology. Annual Review of Biophysics, 51(1):301–326, May 2022. ISSN 1936-122X,
1936-1238. doi:10.1146/annurev-biophys-102521-112912.
31. Francisco Balzarotti, Yvan Eilers, Klaus C. Gwosch, Arvid H. Gynnå, Volker Westphal,
Fernando D. Stefani, Johan Elf, and Stefan W. Hell. Nanometer resolution imaging and
tracking of fluorescent molecules with minimal photon fluxes. Science, 355(6325):606–612,
February 2017. doi:10.1126/science.aak9913.
32. Kirti Prakash. At the molecular resolution with MINFLUX? Philosophical Transactions of the
Royal Society A: Mathematical, Physical and Engineering Sciences, 380(2220):20200145,
February 2022. doi:10.1098/rsta.2020.0145.
33. Susanne C. M. Reinhardt, Luciano A. Masullo, Isabelle Baudrexel, Philipp R. Steen, Rafal
Kowalewski, Alexandra S. Eklund, Sebastian Strauss, Eduard M. Unterauer, Thomas
Schlichthaerle, Maximilian T. Strauss, Christian Klein, and Ralf Jungmann. ångströmresolution fluorescence microscopy. Nature, 617(7962):711–716, May 2023. ISSN 14764687. doi:10.1038/s41586-023-05925-9.
34. Niels Radmacher, Alexey I. Chizhik, Oleksii Nevskyi, José Ignacio Gallea, Ingo Gregor, and
Jörg Enderlein. Molecular Level Super-Resolution Fluorescence Imaging. Annual Review of
Biophysics, 54(1):163–184, May 2025. ISSN 1936-122X, 1936-1238. doi:10.1146/annurevbiophys-071524-105321.
35. Bruno M. Saraiva, António D. Brito, Guillaume Jaquemet, and Ricardo Henriques. Rxivmaker: an automated template engine for streamlined scientific publications, 2025.
doi:10.48550/arXiv.2508.00836.
36. Jervis Vermal Thevathasan, Maurice Kahnwald, Konstanty Cieśliński, Philipp Hoess, Sudheer Kumar Peneti, Manuel Reitberger, Daniel Heid, Krishna Chaitanya Kasuba, Sarah Janice
Hoerner, Yiming Li, Yu-Le Wu, Markus Mund, Ulf Matti, Pedro Matos Pereira, Ricardo Henriques, Bianca Nijmeijer, Moritz Kueblbeck, Vilma Jimenez Sabinina, Jan Ellenberg, and
Jonas Ries. Nuclear pores as versatile reference standards for quantitative superresolution
microscopy (biostudies dataset deposit). https://www.ebi.ac.uk/biostudies/bioimages/studies/
S-BIAD8, 2019. BioStudies accession S-BIAD8.
37. Peter J. A. Cock, Tiago Antao, Jeffrey T. Chang, Brad A. Chapman, Cymon J. Cox, Andrew
Dalke, Iddo Friedberg, Thomas Hamelryck, Frank Kauff, Bartek Wilczynski, and Michiel
J. L. de Hoon. Biopython: Freely available Python tools for computational molecular biology and bioinformatics. Bioinformatics, 25(11):1422–1423, June 2009. ISSN 1367-4803.
doi:10.1093/bioinformatics/btp163.
38. Iván Hidalgo-Cenalmor, Marcela Xiomara Rivera Pineda, Bruno M. Saraiva, Ricardo Henriques, and Guillaume Jacquemet. Packaging jupyter notebooks as installable desktop apps
using labconstrictor, 2026. doi:10.48550/arXiv.2603.10704.
39. Bruno M. Saraiva, Iván Hidalgo-Cenalmor, António D. Brito, Damián Martínez, Tayla
Shakespeare, Guillaume Jacquemet, and Ricardo Henriques. EZInput: A CrossEnvironment Python Library for Easy UI Generation in Scientific Computing, January 2026.
doi:10.48550/arXiv.2601.08859.
40. Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm
for discovering clusters in large spatial databases with noise. In Proceedings of the Second
International Conference on Knowledge Discovery and Data Mining (KDD-96), pages 226–
231, 1996. doi:10.5555/3001460.3001507.
41. Mohamadreza Fazel, Kristin S. Grussmayer, Boris Ferdman, Aleksandra Radenovic,
Yoav Shechtman, Jörg Enderlein, and Steve Pressé. Fluorescence microscopy: A
statistics-optics perspective. Reviews of Modern Physics, 96(2):025003, June 2024.
doi:10.1103/RevModPhys.96.025003.
42. Bo Zhang, Josiane Zerubia, and Jean-Christophe Olivo-Marin. Gaussian approximations of
fluorescence microscope point-spread function models. Applied Optics, 46(10):1819–1829,
2007. doi:10.1364/AO.46.001819.
43. Shyamal Mosalaganti, Agnieszka Obarska-Kosinska, Marc Siggel, Reiya Taniguchi, Beata
Turoňová, Christian E. Zimmerli, Katarzyna Buczak, Florian H. Schmidt, Erica Margiotta,
Marie-Therese Mackmull, Wim J. H. Hagen, Gerhard Hummer, Jan Kosinski, and Martin
Beck. AI-based structure prediction empowers integrative structural analysis of human
nuclear pores. Science, 376(6598):eabm9506, June 2022. doi:10.1126/science.abm9506.
44. Alexander Fotin, Yifan Cheng, Nikolaus Grigorieff, Thomas Walz, Stephen C. Harrison, and
Tomas Kirchhausen. Structure of an auxilin-bound clathrin coat and its implications for the
mechanism of uncoating. Nature, 432(7017):649–653, December 2004. ISSN 1476-4687.
doi:10.1038/nature03078.
45. Jingen Zhu, Himanshu Batra, Neeti Ananthaswamy, Marthandan Mahalingam, Pan Tao,

Martínez et al.

|

VLab4Mic

46.

47.

48.

49.

Xiaorong Wu, Wenzheng Guo, Andrei Fokine, and Venigalla B. Rao. Design of bacteriophage
T4-based artificial viral vectors for human genome remodeling. Nature Communications, 14
(1):2928, May 2023. ISSN 2041-1723. doi:10.1038/s41467-023-38364-1.
Gongpu Zhao, Juan R. Perilla, Ernest L. Yufenyuy, Xin Meng, Bo Chen, Jiying Ning, Jinwoo
Ahn, Angela M. Gronenborn, Klaus Schulten, Christopher Aiken, and Peijun Zhang. Mature
HIV-1 capsid structure by cryo-electron microscopy and all-atom molecular dynamics. Nature,
497(7451):643–646, May 2013. ISSN 1476-4687. doi:10.1038/nature12162.
Jonas Ries, Christian Kaplan, Evgenia Platonova, Hooman Eghlidi, and Helge Ewers. A
simple, versatile method for GFP-based super-resolution microscopy via nanobodies. Nature
Methods, 9(6):582–584, June 2012. doi:10.1038/nmeth.1991.
Ingmar Schoen, Jonas Ries, Enrico Klotzsch, Helge Ewers, and Viola Vogel. BindingActivated Localization Microscopy of DNA Structures. Nano Letters, 11(9):4008–4011,
September 2011. doi:10.1021/nl2025954.
Dominic A. Helmerich, Made Budiarta, Danush Taban, Sören Doose, Gerti Beliu, and Markus
Sauer. PCNA as Protein-Based Nanoruler for Sub-10 nm Fluorescence Imaging. Advanced
Materials, 36(7):2310104, 2024. ISSN 1521-4095. doi:10.1002/adma.202310104.

Methods
implementation and availability. VLab4Mic
is implemented in Python 3.10-3.13 and available as an open-source package on GitHub
(github.com/HenriquesLab/VLab4Mic) under the MIT
licence. The core dependencies include BioPython for
PDB/CIF parsing (37), NumPy and SciPy for numerical
operations, and standard scientific Python libraries for image
processing and visualisation. VLab4Mic provides four usage
modes: (1) widget-based Jupyter notebooks requiring no
coding, (2) Python API for programmatic access, (3) Google
Colab notebooks, and (4) LabConstrictor (38) executables.
Zero-installation Google Colab notebooks let researchers
run VLab4Mic and optimise protocols without high-end
local hardware. VLab4Mic Jupyter notebooks are supported
through EZInput widgets for parameter selection (39). All
simulations presented in this manuscript can be reproduced
using the example scripts and notebooks provided in the
repository.
Software

Structural model generation. We obtain atomic-resolution

structures from the RCSB Protein Data Bank (17) in PDB or
mmCIF format (structural models used in this study are listed
in Table S5). VLab4Mic uses BioPython to parse these files
and extract atomic coordinates, chain information, and residue
identities (Fig. S7). Users specify target sites using flexible
selection syntax that can identify specific residues, chains, or
structural motifs.
Probe positioning. Each probe type (antibody, nanobody,

fluorescent protein, or chemical linker) is defined by a structural model showing the spatial distribution of fluorophore
positions relative to an anchor point. Probes are placed by a
rigid-body algorithm that makes the anchor coincident with
the target site and aligns the probe’s principal axis with the
local epitope normal (Fig. S8), ensuring physically realistic
orientation relative to the protein surface. The normal is calculated via the concentric shell approximation, approximating
each site’s local surface normal as an outward-facing vector
(Supp. Note 1, Surface Normal Calculation, Eq. 3). A wobbling parameter introduces angular variation by perturbing the
normal within a cone of specified opening angle, modelling
orientational flexibility and the resulting fluorophore localisation uncertainty (Fig. S2c; Supp. Note 1, Eq. 4). Multiple
Martínez et al.

|

VLab4Mic

independent copies of the probe are positioned on the epitopes
to generate a labelled structure (Fig. S2d).
Labelling efficiency and structural integrity. Labelling effi-

ciency is modelled as a Bernoulli process where each epitope
site has a probability of being labelled based on the specified efficiency parameter (Supp. Note 1, Eq. 1). Structural
integrity is modelled by stochastic removal of clustered subunits prior to positioning probes (Fig. S3; Supp. Note 1,
Eq. 5). This structural integrity model should be interpreted
as a phenomenological approximation of spatially correlated
structural incompleteness rather than a mechanistic representation of biological disassembly. The approach is intended
to emulate experimentally relevant perturbations such as partial fixation artefacts, incomplete assembly, lattice collapse, or
local molecular loss. The first step uses DBSCAN (40) to cluster epitopes into stable subunits defined by a short-distance
parameter. The second step selects a random breaking point,
from where neighbouring subunits will be removed. A largedistance clustering parameter limits the distance from which
two subunits are considered neighbours. An iterative process
removes neighbouring subunits until the number of remaining subunits closely matches the specified structural integrity
value (structural integrity clustering algorithms are described
in Supp. Note 1).
Virtual microscope implementation. The imaging simulation

treats each fluorophore as a point source emitting photons (41).
Image formation is simulated by convolving the effective PSF
model of each modality (42) with a discrete version of the
fluorophore distribution (Supp. Note 2, Equations Eq. 7 to
Eq. 9) and applying a noise model accounting for shot noise,
readout noise, and EMCCD gain (Supp. Note 2, Equations
Eq. 8 to Eq. 13). For SMLM, each emitter position is also
sampled from a Gaussian centred on its true coordinate, with
σ set to the modality’s localisation precision, mimicking the
accumulation of localisations across frames (Supp. Note 2).
The current noise model is calibrated for camera-based widefield and SMLM detection; consequently, confocal, STED,
and AiryScan images should be read as PSF-shape demonstrations rather than as photometrically faithful representations of
point-detector chains (such as PMT, APD, or GaAsP). Virtual
microscope parameters for all modalities are listed in Table
S3.
Parameter sweep framework. The parameter sweep function-

ality generates a full factorial design across user-specified parameter ranges. Each parameter combination defines a unique
simulation that is executed independently. We score each
simulation against a reference, either a ground-truth model
or a user-specified target, using structural similarity (SSIM)
and Pearson correlation (defined in Supp. Note 3, Equations
Eq. 14 and Eq. 15). User-defined metrics can be plugged in as
required. Results are aggregated into multidimensional tables
that can be visualised as heat maps or line plots.
Comparison against experimental NPC data. For the NPC

consistency check, we used structure 7R5K from the PDB
and selected C-terminal residues of Nup96 subunits as target
7

sites. Extended modality-specific parameters for the NPC virtual sample, PCNA site-specific labelling, and the ChimeraX
clathrin-coated-structure modelling protocol are detailed in
Supp. Note 4. Virtual microscope parameters, detection parameters, including pixel size, field of view, and noise characteristics, were specified to closely match reported experimental values where available (1). Virtual Samples for each
modality example were generated by placing multiple copies
of the labelled structure at positions corresponding to local
maxima in the experimental images, ensuring that the simulated sample closely mimics the spatial distribution of NPCs
observed experimentally (Supp. Note 1). For radius measurements, we implemented an image-based circle fit with
equal parameters to both simulated and experimental images
as described in Supp. Note 4.

8

Martínez et al.

|

VLab4Mic

Supplementary Information

VLab4Mic: prediction of structural resolvability
in super-resolution microscopy

Probe Type
Primary Antibody
Secondary Antibody
Nanobody
GFP-Nanobody
NHS-ester

Model ID
1HZH
1HZH
7MFV
6XZF
Linker

Format
CIF
CIF
CIF
CIF
NA

Probe Size (30)
˜15-20 nm
˜20 nm (primary+secondary)
˜2-4 nm
˜4-6 nm
NA

Wobble (11)
10°
15°
-

Sup. Table S1. Probe library and specifications. Specifications of the varying probe models available in VLab4Mic, detailing their local source models, linkage errors
(distance), and rotational flexibility (wobble) characteristics. Secondary antibody specifications reflect combined primary+secondary linkage distance. Wobble angles are
per-probe values applied at each binding step; the effective wobble of a primary/secondary chain compounds these and can approach 45° (see Structural Modelling and
Labelling note below). The geometry and positioning of these probes are detailed in Fig. S2.

Parameter
Labelling efficiency
structural integrity
Focal depth
Rotation angle

Type
Probe
Structure
Acq
Virtual

Tested Values
0.2-1.0
0.3 - 1.0
-200..+200 nm
0-180°

Figure
Fig. S3
Fig. S3
Fig. S4
Fig. S5

Sup. Table S2. Parameter ranges for sensitivity analysis. Range of values tested during parameter sweep experiments to assess the impact of labelling efficiency, structural
integrity, and optical constraints on imaging performance. Type categories: Probe (labelling-related parameters), Structure (sample integrity), Acq (acquisition/imaging
settings), Virtual (sample construction). Corresponding visualisations for these effects are referenced in the ’Figure’ column.

Modality
Widefield
Confocal
AiryScan
STED
SMLM

PSFxy σ (nm)
130
100
60
20
8

PSFz σ (nm)
400
300
230
20
8

Voxel (nm³)
10^3
10^3
10^3
5^3
5^3

Pixel size (nm)
100
70
40
15
5

Sup. Table S3. Optical parameters for virtual microscopes. Default acquisition settings for the different simulated imaging modalities, including Point Spread Function (PSF)
dimensions, voxel sizes, and Field of View (FOV). These optical properties govern the image formation process shown in Fig. S1.

Parameter
Structure
Labelling target
Strategy
Reference radius (literature)
Measured radii (this work, SMLM)
Effective resolution (PSF sigma in nm)

Value
7R5K (NPC)
Nup96 C-terminal "ELAVGSL" motif
Nup96-GFP + anti-GFP nanobody (indirect; PDB 6XZF)
53.7 ± 2.1 nm (1)
bimodal modes at ˜52 nm and ˜65 nm (Fig. S10)
Widefield = 120. Confocal = 100, AiryScan = 60, STED = 15, SMLM = 5

Sup. Table S4. Configuration for validation against experimental data. Parameters used in both the reference experimental dataset (Nup96-GFP labelled with anti-GFP
nanobody) and matching VLab4Mic simulations for validation. For the NPC SMLM validation specifically, the STED PSF σ deviates from the default in Table S3 (20 nm) and
was set to 15 nm to match the experimental dataset; this single-figure override is detailed in the NPC virtual-sample parameterisation methods.

10

Martínez et al.

|

VLab4Mic

Structure Name
Nuclear Pore Complex

PDB/CIF ID
7R5K

Description
Human NPC

Dimensions
˜110 nm

Symmetry
8-fold

Database
RCSB PDB

Clathrin Coat (full)

1XI5

˜40 nm

Icosahedral

RCSB PDB

Clathrin Triskelion

3IYV

˜30 nm

C3

RCSB PDB

T4 Bacteriophage

8GMO

Auxilin-bound
clathrin cage
Clathrin D6 coat
used to model
the dome and
flat-lattice CCS
(Fig. 4)
Capsid

˜100 nm

Icosahedral

RCSB PDB

HIV-1 Capsid

3J3Y

Capsid

˜60 nm

Fullerene (conical)

RCSB PDB

Reference
Mosalaganti et al.
2022 (43)
Fotin et al. 2004
(44)
—

Zhu et al. 2023
(45)
Zhao et al. 2013
(46)

Sup. Table S5. Macromolecular structural models used for simulations. This table lists the target biological structures, their source identifiers (PDB), physical
dimensions, and symmetry properties used as ground truth geometry in VLab4Mic. Clathrin appears at two granularities: PDB 1XI5 is the full auxilin-bound cage used for the
diverse-structures and structural-integrity demonstrations (Fig. S13b, Fig. S3c), while PDB 3IYV heavy chains were used to model the triskelion building blocks to assemble
the dome and flat-lattice models in ChimeraX for Fig. 4 (see "Clathrin Coated Structures model in ChimeraX" below). The extraction and processing of these structures are
illustrated in Fig. S7.

Feature

VLab4Mic

Input
Macromolecular complex incompleteness
Epitope positioning

Multiple epitope identities on structure
User-defined epitope sites (sequence,
residue, etc.)

SuReSim
Structure Definition
PDB ID or Example
PatPDB/CIF files
terns or surface
Yes
No
Epitopes
Parsing atomic Sampling
on
coordinates from model
PDB/CIF
Yes
No
Yes
No

Direct labelling
Site-specific labelling
Labelling-efficiency
Probe size
Probe orientation
Probe structure from PDB (Antibody, GFP,
SNAP-tag, etc..)

Yes
Yes
Yes
Yes
Yes
Yes

Multiple fluorophores
Advanced photophysics (photoconversion,
FP-maturation, spectral responses)

Yes
No

Virtual sample from experimental data
Expansion Microscopy
Imaging modalities

Custom modalities
3D-PSF
Multicolour imaging
Parameter sweep
Software availability

SMIS

CIR4MICS

TestSTORM

Example
Patterns
No

NA
No

Example
patterns or surface
No

Sampling
on model
or image
No
No

Pre-set positions
for NPCs

Sampling
model

Up to six labels
No

No
No

Yes
No
Yes
Yes
Yes
No

Yes
No
Yes
Yes
No
No

Yes
No
No
Yes
Yes
No

Fluorophores and photophysics
Yes
Yes
No
Yes

NA
NA

Yes
Polarisation,
Cross-talk

No
Yes
No

No
No
SMLM

No
No
No

No
Yes
Yes

No
Python (12)

Matlab (6)

Labelling and probes
Yes
No
Yes
Yes
Yes
No

Sample and imaging modalities
Yes
No
No
Yes
No
No
SMLM, STED, SMLM
Widefield
AiryScan, Confocal, Widefield
Yes
No
No
Yes
Yes
Yes
Yes
No
Yes
Miscellaneous
Yes
No
No
Python library. Java
com- Matlab (4)
Jupyter
Note- piled (7)
books.

on

Sup. Table S6. VLab4mic features compared against existing microscopy simulation tools.

Martínez et al.

|

VLab4Mic

11

Supp. Note 1: Structural Modelling and Labelling. The foundation of VLab4Mic’s simulation capability lies in its ability to

generate realistic, physically plausible models of labelled structures. Unlike simple point-source models often used in microscopy
simulations, VLab4Mic begins with the ground-truth atomic coordinates parsed from PDB or CIF files. This allows for the
definition of target sites (epitopes) with atomic precision, based on user specifications such as specific amino acid residues or
peptide sequences
Structure parsing and target sites. VLab4Mic leverages PDB/CIF models from databases, such as the RCSB Protein Data Bank
(RCSB PDB) or AlphaFold, for accurate modelling of probe placing. By starting from the actual molecular geometry (Fig. S7a),
the simulation can accurately capture the spatial relationships and constraints that exist in real biological samples. VLab4Mic
creates a structural model comprising the atoms within the model and methods to retrieve target sites and associated geometrical
properties. We use BioPython parsing tools to read the atoms in the file. However, some assemblies for macromolecular
complexes are defined as a single asymmetric unit (highlighted in blue) without explicit definition of the whole complex. For
these cases, VLab4Mic constructs the complete assembly model (grey) from the assembly operations in the structure file, as is
the case for the Nuclear Pore Complex (NPC) (Fig. S7a Top). From these structures, specific locations on the macromolecular
structure are defined as Target sites. Parsing of these sites is specific to the structure and independent of the probe model,
allowing for flexible combinations of targets and probes. We consider 2 main types of target definition, which are compatible
with PDB/CIF models (Fig. S7b). i. Protein sequence: an amino acid sequence that is compared against all peptide sequences
in the PDB/mmCIF. Only exact matches are retrieved as epitopes. The position of the epitope is calculated as the average
coordinate from the atoms that belong to that peptide sequence. ii. Residue: specified by a 3-letter code. VLab4Mic iterates over
all protein chains and retrieves all instances of the Residue, based on its alpha carbon. The position of the epitope is defined as
the coordinate of the alpha carbon atom. VLab4Mic allows for multiple probes to be defined for a single structure, each with its
own target specification (Fig. S7c). Additionally, the residue-based target allows for site-specific labelling, where the user can
define exact residue numbers to be targeted and the protein chain within the model Fig. S12.
Probe modelling. A probe describes the position of the fluorophores in 3D space relative to a structural model, such as an

antibody or GFP.
• Fluorophore positions. The exact positions of the fluorophores can be parsed from the atomic structure of the probe
model or as the chromophore coordinates of fluorescent proteins. Common conjugation sites for antibodies, such as lysine
residues, can be used to define these positions.
• Degree of labelling (DoL). To capture the variability in the number of fluorophores attached to a probe, VLab4Mic allows
users to specify a degree of labelling (DoL) parameter. This parameter defines the average number of fluorophores per
probe molecule, and the actual number of fluorophores for each probe instance is sampled from a Poisson distribution
with mean equal to the DoL. This allows for realistic variability in fluorophore density across different probes, which can
impact the brightness and localisation precision in the simulated images.
• Probe geometry. The spatial arrangement of fluorophores on the probe is determined by the probe model. Besides the
fluorophore positions, the probe model also defines the paratope site (the site with affinity for the epitope in the structure)
and an internal axis that defines the orientation of the probe relative to the target.
• Distance to epitope. The distance from the probe paratope to the epitope site can be adjusted with an offset parameter that
extends the probe anchor site along the probe axis prior to any probe positioning.
Probe Positioning and Orientation.

• Stochastic model for structure labelling. In experimental fluorescence microscopy, it is theoretically impossible for
every available epitope to be labelled. Factors such as antibody affinity, incomplete staining protocols, and competition for
binding sites result in fractional labelling efficiencies. VLab4Mic simulates this realistic variability through stochastic
labelling where each potential epitope site is independently labelled with a probe. Labelling is modelled as an iterative
process where each epitope site is evaluated for successful labelling (Bernoulli trial) and then subjected to steric constraints
based on the probes already placed.
• Bernoulli Labelling. For a given target structure, we assume that each potential epitope site i is independent. This
assumption simplifies accessibility effects, cooperative binding phenomena, and local environmental heterogeneity, which
are not explicitly modelled in the current implementation. The probability of any specific site being successfully labelled
is defined by the efficiency parameter εlabel . This is mathematically implemented as a Bernoulli trial for each site (Eq. 1):
Li ∼ Bernoulli(εlabel )
12

(1)
Martínez et al.

|

VLab4Mic

where Li = 1 indicates a successful label and Li = 0 indicates an unlabelled site. Per-site independence is assumed; the steric
clash term below enforces a minimum spacing between probes but does not encode cooperativity or accessibility-dependent
affinity. The effect of varying labelling efficiency on the resulting fluorophore distribution is shown in Fig. S9a.
Once a site is determined to be labelled, the position of the fluorophore relative to the target epitope is not random. Probes are
modelled as rigid bodies anchored to the target site, and their spatial orientation is determined by the local surface geometry of
the macromolecule. This correctly models the "linkage error" (the distance between the target protein and the actual emitter)
which is a critical factor in super-resolution accuracy (47).
• Steric Constraint Sampling. Physical probes, such as antibodies or even small organic dyes, occupy physical volume.
In dense epitope regions, it is physically impossible for two probes to occupy the same space or for a probe to bind if
its footprint is obstructed by neighbours. To account for these steric clashes, VLab4Mic enforces a minimum distance
constraint when placing probes in epitope sites. This minimum distance is estimated as the maximum pairwise-length
from the atoms within the probe model. A label at site i is only permitted if it satisfies a minimum distance dthreshold from
all other successfully labelled sites (Fig. S9b). The final distribution of fluorophores reflects the physical exclusion volume
of the probes and their efficiency of labelling an epitope.
• Probe Orientation. The probe’s principal axis is aligned with the vector normal to the protein surface at the binding site.
The final coordinates of the fluorophores is calculated by treating the probe as a rigid body and performing a rotation and
translation operation to align the probe axis with the epitope normal vector. We use the Rodrigues’ rotation formula (Eq. 2)
to compute this transformation:
protated = cos(θ )p + sin(θ )(k × p) + (1 − cos(θ ))(k · p)k

(2)

where p is the fluorophore coordinate, θ is the angle between the probe’s internal axis and the epitope normal, and k is the unit
vector along the rotation axis defined by the cross product of these two vectors.
• Surface Normal Calculation. To orient the probe correctly, the simulation must determine "which way is up" relative to
the protein surface. We approximate the local surface normal Ni at target site i by comparing the target position on the
original structure to a concentric, scaled-down version of the structure (scaling factor s < 1, default s = 0.95, centroid
preserved). The vector from the scaled-down position back to the original provides an outward-facing normal estimate
(Fig. S8, Eq. 3):
Ni = poriginal,i − pscaled,i

(3)

This concentric-scaling approximation is reliable for convex target distributions (e.g., outer rings of NPC, capsid exterior), but
degrades on locally concave or planar geometries (e.g., flat clathrin lattices, inward-facing epitopes); for those cases users can
supply pre-computed normals or substitute a local-neighbourhood method (e.g., PCA over neighbouring atoms).
• Angular Wobble. While probes are modelled as rigid bodies, the linkers connecting them (e.g., antibody hinges or
flexible chemical linkers) introduce rotational freedom. This flexibility, or "wobble," effectively blurs the fluorophore’s
position over time, degrading localisation precision (48). The wobble parameter should therefore be interpreted as an
effective orientational uncertainty rather than a direct physical measurement of probe dynamics. VLab4Mic models this by
introducing random angular perturbations to the probe’s ideal orientation, sampled to be isotropic over the spherical cap
defined by a user-specified maximum half-angle θmax (Eq. 4):
cos θwobble ∼ Uniform(cos θmax , 1),

φwobble ∼ Uniform(0, 2π)

(4)

The cos θ parameterisation ensures uniform sampling over solid angle within the cone, rather than concentrating samples near
the pole.
To accurately model linkage error, we recommend users approximate wobble cones based on the probe’s mechanical properties.
For instance, use tighter constraints for nanobodies ( 10°) compared to the flexible hinge regions of primary/secondary antibody
complexes ( 45°). This allows the user to simulate the specific localisation uncertainty introduced by their chosen labelling
strategy.
Primary and Secondary probe labelling. To model the commonly used primary and secondary antibody labelling strategy,
VLab4Mic allows users to define a two-step labelling process. Primary probes are placed on the target structure using the
stochastic labelling and probe positioning methods described above (Fig. S6a). In the second step, secondary antibodies are then
placed on the already positioned primary probes (Fig. S6b). The target site for the secondary antibody is defined as an epitope on
the primary antibody model. The placement of secondary antibodies is also subject to steric constraints, ensuring that they do not
overlap with each other or with the primary antibodies. When secondary labelling is used, the final fluorophore positions are
only the fluorophores on the secondary antibodies.

Martínez et al.

|

VLab4Mic

13

Structural integrity. Real biological samples are rarely perfect. They exhibit structural heterogeneity, damage from fixation,

incomplete assembly or epitope regions inaccessible to the probes. To test the robustness of imaging methods against such
imperfections, VLab4Mic explicitly models variations of structural integrity.
• Cluster-Based structural integrity (DBSCAN). Often, damage occurs in larger, contiguous patches, for example, a
missing sector of a nuclear pore complex ring. To emulate this large-scale structural damage, we employ Density-Based
Spatial Clustering of Applications with Noise(40) (DBSCAN). The first step groups epitopes into stable subunits defined
by a short-distance cluster parameter (Eq. 5). We note that this is a phenomenological abstraction rather than a biologically
validated damage model; the resulting degradation morphology is one of several plausible regimes, and users may substitute
alternative damage rules (for example, per-subunit dropout, patterned ablation, or accessibility-weighted removal) by
replacing the cluster-selection routine.
C = DBSCAN({epitopes}, dshort , 1)

(5)

where dshort defines the neighbourhood radius for epitopes to cluster into a subunit. Each subunit is then represented by the
average coordinate of the epitopes it contains. Then, a K-d Tree is built for these subunits, so that a distance parameter dlarge is
used to query neighbouring subunits.
Once grouped, a subunit is selected at random as the breaking point from which an iterative sample process will remove
neighbouring subunits. On each iteration, the breaking point is removed and a random subset of its neighbouring subunits
will also be removed. On the next iteration a new breaking point is selected from the subunits that are already removed. A
large-distance parameter limits the distance from which two subunits are considered neighbours. This process of removal of
subunits is repeated until the number of epitopes in the remaining subunits closely matches the specified structural integrity
value.
Simulating clustered integrity loss enables researchers to train and validate analysis pipelines, ensuring they can robustly
discriminate between structural integrity artefacts and genuine structural gaps in macromolecular assemblies.
Virtual Sample Construction. A virtual sample consists of multiple independent copies of these labelled structures distributed

in 3D space.
• Labelled Structure Placement. Labelled structures can be placed within the defined sample volume Vsample by specifying
their positions in 3D or by random placing. Crucially, this placement is subject to a non-overlap constraint to prevent
physical intersections between distinct molecular complexes (Eq. 6):
pi ∼ Uniform(Vsample ),

subject to ∥pi − p j ∥ > dmin ∀ j < i

(6)

• Labelled Structure Orientation. The orientation of each structure is defined by an orientation vector; when reorienting a
labelled structure, its central axis, and the rest of the structure, is aligned with the orientation vector. The labelled structure
orientation can be defined by the user or randomly assigned. If a random orientation is assigned, the orientation vector is
sampled uniformly from the unit sphere, ensuring an isotropic distribution of orientations across the sample. However,
users can also specify a bias in the orientation distribution by specifying a set of rotation angles (Euler angles).
• Labelled Structure planar rotation. Finally, a planar rotation can be applied to the labelled structure, which is a rotation
around the central axis of the structure. This is particularly useful for structures with a defined central axis, such as the
NPC, where the planar rotation will rotate the structure around its central axis.
Virtual Sample priming from input image. VLab4Mic allows users to "prime" a virtual sample with positions derived from a 2D

image. VLab4Mic supports two methods for extracting particle coordinates from an input image: (1) local maxima detection
and (2) random sampling from binary masks. The resulting virtual sample dimensions are calculated to match the physical
dimensions of the input image given its pixel size.
• Local maxima detection. This method identifies local intensity peaks in the input image, which are assumed to correspond
to individual particles or structures. The coordinates of these peaks are extracted and used as the positions for placing
labelled structures in the virtual sample.
• Random sampling from binary masks. A binary mask delineates regions of interest in the image where particles are
likely to be located. The selection of particle positions is done iteratively by randomly sampling coordinates within the
masked regions, constrained by a minimum distance to ensure non-overlapping placement of structures that can be derived
by the labelled structure dimensions or a user-defined parameter.

14

Martínez et al.

|

VLab4Mic

Supp. Note 2: Image Formation and Photophysics. Once a virtual sample is generated, VLab4Mic image formation process

follows a forward model based on an effective PSF model per modality. This effectively acts as a ’virtual microscope,’ translating
the exact 3D coordinates of fluorophores into pixelated images. This process involves modelling the diffraction-limited optics of
the system, the photophysical properties of the fluorophores, and the stochastic noise introduced by the detector.
Point Spread Functions. The primary determinant of image resolution in any optical system is the Point Spread Function (PSF),
which describes the system’s response to a point source. Image formation is mathematically modelled as the convolution of the
discrete fluorophore distribution with this PSF.

• 3D Elliptical Gaussian PSF. We approximate the effective PSF of a modality as a 3D elliptical Gaussian distribution
where lateral and axial width can be parameterised independently (Eq. 7):
!
(x − x0 )2 (y − y0 )2 (z − z0 )2
PSF(x, y, z) = I0 · exp −
−
−
(7)
2σx2
2σy2
2σz2
where σx , σy , σz define the lateral and axial resolution widths, and (x0 , y0 , z0 ) is the PSF centre coordinate.
The PSF is a discrete 3D array with sampling rate of 10 nm in each dimension for conventional modalities, while for SMLM and
STED, the sampling rate is 5 nm. Users can adjust this sampling rate to balance accuracy and computational efficiency.
Fluorophore Photophysics. We model the brightness of each fluorophore by defining the average photon emission rate per unit

time when the fluorophore is in the ’On’ state. VLab4Mic models the stochastic behaviour of photon arrival as a Poisson process,
which is a common assumption in fluorescence microscopy.
Scope of the photophysical model. Photon-emission rates and integration times drive shot noise and set the per-emitter
localisation count for SMLM (see Simulation of localisation microscopy modalities below). VLab4Mic does not explicitly model
fluorophore photoswitching kinetics (kon , ko f f , dark-state lifetimes, photobleaching) and does not derive blinking duty cycles or
exhaustion from first principles. The forward model targets the effective post-reconstruction image a researcher would obtain
after any modality-specific reconstruction step, rather than simulating raw camera frames followed by a reconstruction pipeline.
• Photon Emission (Shot Noise). For each emitting fluorophore, the number of photons detected during an integration time
∆t is a random variable drawn from a Poisson distribution with mean λ photons · ∆t, where λ photons is the average photon
emission rate (Eq. 8). This captures the inherent shot noise in fluorescence imaging, which arises from the discrete nature
of photon emission and detection:
N photons (On) ∼ Poisson(λ photons · ∆t)

(8)

where λ photons is the average photon emission rate.
Virtual sample discretisation. To perform convolution, the continuous fluorophore distribution must be discretised into a 3D

grid of voxels. The sampling rate of the virtual sample is defined by the sampling rate of the PSF of the imaging modality.
Before convolution, the number of photons emitted by each fluorophore is calculated based on the photophysical model and
integration time (exposure time), so that each voxel contains the total photon count from all fluorophores within it.
Simulation of localisation microscopy modalities. To simulate the effect of an emitter being localised multiple times across

different frames, as is the case in SMLM, we include a previous step where each fluorophore exact position is used to sample
new localisation positions from a Gaussian distribution centred on the fluorophore position and with a standard deviation defined
by the localisation precision of the modality. An average number of localisations per fluorophore can be defined by the user, and
the actual number of localisations for each fluorophore is sampled from a Poisson distribution with mean equal to this parameter.
Lateral uncertainty and axial uncertainty are sampled independently. The final image is the result of convolving the PSF of
the imaging modality with sampled localisation positions. This localisation simulation is set as default for SMLM, but can be
disabled to simulate a SMLM modality without localisation uncertainty. The Poisson per-emitter localisation count is appropriate
for PAINT-type acquisitions, in which transient binding events accumulate independently; PALM/STORM-style exponential
per-site distributions, where switching of a fixed fluorophore until photobleaching produces an exponential tail of localisations,
are not modelled.
Image formation via convolution and binning. To obtain the final image, the discretised fluorophore distribution is convolved
with the PSF of the imaging modality, integrated in the axial dimension and then binned to the pixel size of the detector.

• Convolution of virtual sample and PSF. The interaction between the optical system and the sample is linear and shiftinvariant. Therefore, the final intensity distribution I(x, y, z) is computed via Fast Fourier Transform (FFT) convolution of
the fluorophore spatial distribution F and the system PSF (Eq. 9):
Martínez et al.

|

VLab4Mic

15

I(x, y, z) = F ⊗ PSF =

Z Z Z

F(x′ , y′ , z′ ) · PSF(x − x′ , y − y′ , z − z′ ) dx′ dy′ dz′

(9)

This results in a 3D intensity array with the same sampling rate as the PSF.
• Depth of Field and Out-of-Focus Fluorophores. The intensity array is then projected onto a 2D plane by summing
along the axial dimension with respect to the focal plane in the sample. The depth of field of the microscope restricts the
planes above and below the focal plane that contribute to the image. This projection simulates the depth of field of the
microscope, where fluorophores that are out of focus contribute to the image with a broadened PSF in the axial dimension.
• Pixel Binning. Finally, the 2D intensity distribution is binned (with addition) to match the pixel size of the detector. This
step simulates the finite spatial sampling of the camera, where each pixel integrates the signal from a defined area of the
sample. The resulting 2D array represents the ideal, noise-free image formed by the virtual microscope.
Detector Noise Models. The final stage of image formation occurs at the detector, where photons are converted into digital

signals. VLab4Mic simulates detector characteristics by including quantum efficiency, amplification, and electronic readout noise
pixel-wise. A single camera-style model (quantum efficiency × stochastic gain × Gaussian readout × digitisation, see equations
below) is applied uniformly across modalities. This matches camera-based acquisitions (widefield, SMLM); point-detector
modalities (confocal, STED, AiryScan) whose physical detectors (PMT, APD, GaAsP) follow Poisson + dark-count statistics are
not modelled at the detector level, and their outputs should be read as PSF-shape and spatial-arrangement demonstrations.
• Detection Efficiency. Not every photon hitting the detector generates a signal. The probability of an incident photon
being converted into a photoelectron is determined by the detector’s Quantum Efficiency (QE). This loss is modelled as a
Binomial process (Eq. 10):
Ndetected ∼ Binomial(Nemitted , QE)

(10)

where Nemitted is the number of photons in the noise free image and QE is the quantum efficiency of the detector.
• Electron Multiplication. Detected photons are amplified before readout. This stochastic amplification process introduces
additional noise and is modelled using a Gamma distribution (Eq. 11):
Signalampli f ied ∼ Gamma(Ndetected , scale = g)

(11)

where g is the gain factor.
• Readout Noise and Baselevel. Finally, Gaussian readout noise, which is independent of the signal level is added to the
amplified signal (Eq. 12). A constant baseline offset (BaseLevel) can also be added which is usually done to prevent
negative pixel values:
2
Signalread = Signalampli f ied + N (0, σread
) + BaseLevel

(12)

• Digitisation. The analogue voltage signal resulting from detection, amplification and readout is then quantised into
discrete analogue-to-digital units (ADU), completing the image formation process (Eq. 13):
ADUcounts = ⌊ADU · Signalread ⌋

(13)

Multicolour Imaging. VLab4Mic supports multicolour imaging by allowing users to define multiple probes with distinct

fluorophores. Each fluorophore will generate a separate image channel for the same modality parameters. The final output is a
multi-channel image stack that can be analysed using standard multicolour microscopy techniques.

16

Martínez et al.

|

VLab4Mic

Supp. Note 3: Parameter sweep and validation. This section outlines the comprehensive workflow employed by VLab4Mic

for systematic parameter exploration and the quantitative metrics used to validate the accuracy of the simulations against
experimental ground truth.
VLab4Mic Computational Pipeline. The overall simulation pipeline integrates the structural modelling and image formation

components into a cohesive workflow. As illustrated in Fig. S1, the process begins with atomic structure parsing to define the
ground truth geometry. This is followed by the generation of a "virtual sample" (a populated 3D volume of labelled particles).
Finally, this virtual sample is effectively "imaged" by the forward model to produce multi-modality synthetic data (Widefield,
Confocal, STED, SMLM) from the exact same underlying sample definition.
Parameter Sweep Workflow. A robust validation framework requires more than single-instance comparisons. To systematically

explore the impact of experimental variables on imaging outcomes, VLab4Mic implements an automated parameter sweep
generator. This tool creates a factorial design of simulation parameters, varying factors such as labelling efficiency, probe type,
noise levels, and optical settings. It then executes batch simulations for each unique combination. This facilitates comprehensive
sensitivity analysis, allowing researchers to determine which experimental parameters are most critical for resolving specific
structural features.
Image Similarity Metrics. To objectively quantify the agreement between simulated images (Isim ) and experimental references
(Iexp ), we employ standard image quality metrics.

• Structural Similarity Index (SSIM). The SSIM is a perception-based model that considers image degradation as
perceived change in structural information. It measures similarity based on three comparisons: luminance, contrast, and
structure. Unlike simple mean squared error, SSIM is highly sensitive to the spatial structure of local image patterns,
making it ideal for microscopy validation (Eq. 14):
SSIM(Isim , Iexp ) =

(2µsim µexp + c1 )(2σsim,exp + c2 )
2 + µ 2 + c )(σ 2 + σ 2 + c )
(µsim
2
1
exp
exp
sim

(14)

where µ and σ represent the local mean and variance, and c1 , c2 are stabilising constants.
• Pearson Correlation. The Pearson correlation coefficient (ρ) evaluates the linear correlation between pixel intensities of
the two images. It provides a measure of how well the intensity variations in the simulation match those in the experiment,
independent of absolute brightness scaling (Eq. 15):
ρ(Isim , Iexp ) =

Cov(Isim , Iexp )
σsim · σexp

(15)

Comparison Methodology for parameter sweep. Direct comparison of simulated and reference images is often hindered by the

random placement of particles in experiments. To address this, VLab4Mic implements a "priming" function for validation. This
function extracts the exact particle coordinates from the reference sample. It then uses these coordinates to position the virtual
structures in the simulation. This ensures that the spatial arrangement of the simulated sample matches the reference image,
allowing for accurate computation of pixel-wise similarity metrics like SSIM.
Image masks. To further refine the comparison, we use the priming positions to generate a binary mask to limit the analysis
to regions where particles are present. By applying this mask to both the reference and simulated images before computing
similarity metrics, we focus the validation on relevant regions, reducing the influence of background noise and empty space.

Martínez et al.

|

VLab4Mic

17

Supp. Note 4: Extended Methods.
Virtual Sample parameterisation of Nuclear Pore Complexes. Virtual Samples for NPC simulations based on experimental data
(1) were generated by positioning independent copies of the NPC labelled with Nup96-GFP in a 1.5 µm x 1.5 µm FOV. We used
the PDB 7R5K as model for NPC (Table S4). The probe was based on a GFP with nanobody (PDB: 6XZF), with degree of
labelling of 2, epitope defined as the peptide sequence "ELAVGSL" and labelling efficiency of 0.4. The coordinates on where
to place the NPCs were obtained by extracting the local maxima peaks with modality-specific image parameters. Widefield:
pixelsize = 100 nm, sigma = 0.5, min distance = 100 nm, threshold = 0. Confocal: pixelsize = 70 nm, sigma = 1, min distance =
70 nm, threshold = 400. AiryScan: pixelsize = 40 nm, sigma = 0.5, min distance = 40 nm, threshold = 100. STED: pixelsize =
15 nm, sigma = 3, min distance = 90 nm, threshold = 3. SMLM: pixelsize = 5 nm, sigma = 6, min distance = 120 nm, threshold
= 10. Experimental data was obtained from available datasets (36). PSF width for Widefield was set as the mean PSF width
measured on the localisation table from the experimental data, and the effective SMLM PSF as the mean localisation precision
reported on the same table. PSF widths (sigma) were set as follows: Confocal = 100 nm, AiryScan = 60 nm, STED = 15 nm.
NPC radii quantification. VLab4Mic includes methods to allow users to prime a virtual sample with positions derived from an

experimentally acquired image. We used this functionality to create equivalent simulated images that allow us to reduce the bias
of structure placement within a virtual sample for image comparison (Fig. S10). We then measure the apparent radii of NPCs on
both simulated and real data to assess if VLab4Mic simulations allow us to recapitulate reported geometrical descriptors of the
NPC. Figure (Fig. S10a) shows an example of the STED imaging of NPCs and a VLab4Mic equivalent image simulation for the
same modality; the right panel shows their respective histograms for measured radii on each image using the same parameters.
Figure (Fig. S10b) shows a similar example but with SMLM images.
Two methods can be used to estimate an NPC radius from SMLM data. Thevathasan et al. (1) fit a circle directly to the table of
single-molecule localisation coordinates (a localisation-table fit), and report R = 53.7 ± 2.1 nm. We instead apply circle-fitting to
the rendered SMLM images, which is the operation available to a user who has access only to the reconstructed image and not to
the underlying localisation table. Under this image-based fitting we observe a bimodal histogram of fitted radii with a primary
mode around 52 nm (matching the localisation-table radius reported by Thevathasan) and a secondary mode around 65 nm. Both
modes are recovered by VLab4Mic on the simulated equivalent dataset, even though every NPC in the simulation is identical by
construction. This indicates the secondary mode is associated with the image-rendering and image-based fitting steps rather than
with biological variation in the sample. The agreement between the simulated and experimental distributions supports the use of
VLab4Mic to recapitulate image-based quantitative descriptions of the experimental data.
PCNA site-specific labelling. VLab4Mic allows for site-specific labelling by defining exact residue numbers to be targeted
within a protein chain. This feature is particularly useful for simulating experiments where precise labelling of specific amino
acids is required, such as labelling of the Proliferating Cell Nuclear Antigen (PCNA), to serve as a protein-based nanoruler (49).
Here, the residue SER-186 within each PCNA monomer is targeted, and the expected geometrical distribution of the fluorophores
is a triangle of 6nm separation between each target. VLab4Mic accurately models this site-specific labelling (Fig. S12a), where
the fluorophore positions (magenta) are shown relative to the PCNA structure (grey). The distances between the fluorophores in
our model closely match the expected 6 nm separation. Furthermore, VLab4Mic SMLM simulations with typical resolution
parameters (Fig. S12a) do not resolve the three sites as previously reported (49). However, VLab4Mic custom modalities enable
the user to explore the conditions required to resolve these closely spaced epitopes: at an effective resolution of 2 nm the three
sites remain unresolved as a single feature, whereas at 1 nm they are clearly separated (Fig. S12a).
Clathrin Coated Structures model in ChimeraX and target selection for VLab4Mic. Dome and flat clathrin coated structures were

modelled in ChimeraX by manually placing copies of triskelion derived from 3IYV. To guide the placement of the triskelions, an
image of a clathrin coated structure was loaded in the same session. The triskelions were placed and reoriented manually. The
resulting model was exported as a PDB file and used as input for VLab4Mic. For the probes to target both models, we chose an
arbitrary short sequence ("NNRIA") that could be parsed from both light chains (CLC) as these chains had been reported as
targets for probing CCS ((18)).
Parameter sweep with HIV-1 capsid. Parameters for HIV-1 capsid sweep were chosen as follows. Structure: "3J3Y". Probe

template: "anti-p24_primary_antibody_HIV". Number of sweep repetitions: 10. Labelling efficiency and structural integrity
parameters: min = 0, max= 1, steps = 0.25. Structural integrity distances were: small cluster = 20; large cluster = 100. Exposure
time = 0.001. All other parameters were used as default.
Interdependence of resolvability conditions. The minimal conditions to achieve structural resolvability are not independent:

probe geometry, labelling efficiency, structural integrity, and modality resolution interact, so no single parameter can be optimised
in isolation. For instance, a larger probe (a primary/secondary antibody pair rather than a nanobody) increases linkage error, and
recovering the same set of epitopes then requires a higher effective resolution; equivalently, a reduced labelling efficiency must
be compensated either by higher resolution that can disambiguate sparser localisations or by averaging over more particles. The
18

Martínez et al.

|

VLab4Mic

parameter sweep framework (Fig. 3) is designed to expose these trade-offs, letting the user identify, for a given target feature, the
joint operating envelope of probe, labelling, and acquisition parameters within which the feature remains resolvable.

Martínez et al.

|

VLab4Mic

19

a)

Structure atoms

PDB,
AlphaFold
or ...

HIV-1 capsid

... Clathrin Coat Pit

Nuclear Pore Complex

Source

Structures

b)

Probe atoms

PDB,
AlphaFold
or ...

GFP

Emitters

...

Nanobody

Antibody

Probes

Source
Structure atoms

c)

Target sites

Emitters

+
Selected structure and probe

Probe positioning

Virtual sample

Labelled structure

d)

Virtual sample
(continuous)

Virtual sample
PSF
(discrete) + activation

Activation volume

Activation frame

*
* Depth of Field

Virtual sample convolution with PSF

e)
Quantum efficiency
Signal amplification
Readout Noise
Digitalisation

Activation frame

Activation frame at
Detector pixel size

Noise model

Final image

Signal integration and image generation
Sup. Fig. S1. Computational pipeline for image formation simulation. a) Structure selection. VLab4Mic workflow starts by selecting a structure of interest. Here a model
for the macromolecular complex is based on PDB/CIF atomic models from available databases such as the PDB or AlphaFold. b) Probe specification. We specify a probe
model, such as a GFP or an antibody that is conjugated with fluorophores (magenta dots) and binds to specific sites in the structure, known as target sites (not shown). The
number of fluorophores shown in the antibody and nanobody models correspond to their degree of labelling. Scalebars = 2 nm. c) Labelled structure and virtual sample.
Having selected a structure and a probe, VLab4Mic models the labelling of the targets in the structure by positioning copies of the probe onto the target sites while following the
geometrical constraints of both models. This creates a general model for this particular labelling strategy. To construct the virtual sample, one or more copies of the labelled
structure are placed within a defined volume. Here, each labelled particle is independent of each other. d) Forward model and activation. VLab4Mic uses a forward model for
image generation to create an imaging simulation of the virtual sample. First, the virtual sample is discretised into a grid whose voxels are of the same dimensions as a 3D PSF
model. An activation model assigns a random number of photons to each emitter in the virtual sample, so that the intensity per voxel reflects the total amount of photons
emitted in that volume. The discretised virtual sample is then convolved with the 3D PSF to generate an activation volume. In a final step, an activation frame is generated by
integrating the signal of the activation volume over the depth of field specified by the imaging modality. e) Detection and noise. The activation frame represents the spatial
distribution of photons in the field of view that reach the detector. This image is binned (with addition) so that the new pixel size matches the detector pixel size, which now
represents the total photon intensity at detection. A modality-dependent noise model is applied to each pixel which includes efficiency of detection, signal amplification, readout
and digitalisation noise.

20

Martínez et al.

|

VLab4Mic

Sup. Fig. S2. Library of molecular probe models and positioning logic. a) Probe model. A probe is a model for the spatial distribution of emitters or fluorophores
(magenta), relative to the probe’s own model (grey). The probe model contains information of a target site for which it has affinity to bind (represented as an anchor at the
bottom). This probe is considered a rigid body that can be manipulated based on the anchor and an internal axis. VLab4Mic can model conventional labelling strategies such as
antibodies by using the antibody PDB to define putative emitter sites, such as Lysine residues used for NHS-ester fluorophore conjugation. The number of fluorophores shown
in the antibody and nanobody models correspond to their degree of labelling. The anchor point can be approximated by known paratope sequences or with a user-defined site.
b) Probe positioning. We model a probe binding its epitope by repositioning the probe as a rigid body to (1) match the probe anchor and a target site, and (2) align the probe
internal axis with the target normal vector. This process can include a wobbling step (right), in which the target normal is given a random perturbation within a user-defined cone
as depicted. c) Antibody probe example. Model for an antibody probe, depicted from its structural atoms (grey). Its target site (green) and target normal (vertical grey line) are
shown to illustrate probe positioning scenarios with no wobble (0 degrees, left), wobble within 30 (centre) or 45 degrees from the target normal. Scale bars in a) are 2 nm.

Sup. Fig. S3. Structural integrity modelling. VLab4Mic can emulate structural integrity as a proxy for epitope accessibility. For each structure, the target positions were used
as emitter positions. An example of a labelled particle is shown for different degrees of structural integrity (left panels). SMLM imaging simulations per particle and structural
integrity are shown on the right. a) HIV-1 capsid. Target sequence: SPRTLNA; short-range cluster = 20 Å, long-range clusters = 100 Å. b) Nuclear Pore Complex. Target
sequence: ELAVGSL; short-range cluster = 300 Å, long-range clusters = 600 Å. c) Clathrin Coated Pit. Target sequence: EQATETQ; short-range cluster = 100 Å, long-range
clusters = 200 Å.

Martínez et al.

|

VLab4Mic

21

Sup. Fig. S4. Impact of depth of field on 3D imaging outcomes. a) Virtual sample. Two nuclear pore complexes: NPC #1 located at the focus plane, NPC #2 placed 150
nm above the Z axis and opposite in the FOV. The virtual sample is shown from lateral and top perspectives. b) Simulated images per modality. Imaging simulations per
modality where the only parameter changed was the depth of field. An orange arrow indicates the depth of field at which a given modality collects signal from NPC #2.

Sup. Fig. S5. Construction and parameterisation of the 3D virtual sample. a) Sample space and placement. A virtual sample represents fluorophores in a 3D sample
space (XY plane highlighted in blue). A labelled structure (top) can be placed within the sample space by specifying relative positions within the field. The default placement is a
single labelled structure at the middle of the sample (middle), whereas setting multiple copies generates randomised, non-overlapping positions (bottom). b) Placement from a
binary mask. Labelled structures can be sampled from a binary mask, where each pixel represents equal probability for a position to occur within that area. c) Placement from
a greyscale image. Labelled structures can be placed by locating local maxima positions in a greyscale image. This approach approximates the input image with modelled
structures.

22

Martínez et al.

|

VLab4Mic

Sup. Fig. S6. Simulation of primary and secondary linkage error. a) Primary antibody labelling. (Left) Primary antibody with conjugation sites used to define the emitters
for this probe. (Middle) Model for the HIV-1 capsid labelled with a primary antibody, showing the target sites of the capsid, the emitters. (Right) SMLM image simulation of HIV-1
capsid labelled with a primary antibody. b) Primary + secondary nanobody. (Left) Primary antibody without conjugation sites for emitters, paired with a secondary probe
based on a Nanobody whose conjugation sites are used as emitter sites. (Middle) Model for the HIV-1 capsid labelled with the primary probe and a secondary directed to the
primary, showing the target sites of the capsid, the emitters. (Right) SMLM image simulation of HIV-1 capsid labelled with a primary and a secondary nanobody. c) Primary +
secondary antibody. (Left) Primary antibody without conjugation sites for emitters, paired with a secondary probe based on the same model for the primary antibody. (Middle)
Model for the HIV-1 capsid labelled with the primary probe and a secondary directed to the primary, showing the target sites of the capsid, the emitters, and an example SMLM
image. (Right) SMLM image simulation of HIV-1 capsid labelled with a primary and a secondary antibody. Scalebar = 100 nm. d) Secondary with no wobble. Secondary
probes are independent of primaries, which allows for independent parameterisation such as degree of wobble. Here, the secondary probe model is represented in grey without
fluorophores; primary is not shown. If the secondary is set to not include wobble, the orientation of secondaries will reflect the primary and epitope normal vector. This results in
neighbouring secondaries to have similar orientations. e) Secondary with wobble. Here the same secondary probe includes a wobble of 60º. Secondary probe is represented
in grey without fluorophores, primary is not shown. This results in randomised orientations of the secondary probes.

Martínez et al.

|

VLab4Mic

23

Sup. Fig. S7. Extraction of ground-truth geometry from atomic PDB models. A structure in VLab4Mic is the representation of an atomic model for the macromolecule of
interest. a) Structure parsing. VLab4Mic loads PDB files from databases such as the RCSB Protein Data Bank (RCSB PDB) or AlphaFold. We use BioPython parsing tools to
read the atoms in the file. For large assemblies, the atoms in the file represent an asymmetric unit (highlighted in blue), as is the case for the Nuclear Pore Complex (NPC) (top).
VLab4Mic creates the complete assembly model (grey) from the assembly operations in the structure file. b) Target parsing. A VLab4Mic structure provides methods to select
positions that match a specific amino-acid sequence or residue. For instance, it is possible to retrieve all Lysine residues found in the NPC structure (top left), or all positions
that match the sequence "ELAVGSL", a peptide motif found in the Nup96 protein. The same procedure can be applied to other structures such as the HIV-1 capsid model
(bottom): the example shows the Lysine residues in the HIV-1 capsid on the left, and positions that match the sequence "SPRTLNA", a peptide found in the HIV-1 capsid
monomer (CA). c) Multiple target selection. VLab4Mic makes it straightforward to retrieve multiple target sites, each based on its specific property (residues or sequences).
Example shows Nup96 and Nup107 C-terminals.

24

Martínez et al.

|

VLab4Mic

Sup. Fig. S8. Algorithm for probe positioning and surface orientation. a) Normal vector approximation. We approximate the normal vector direction by creating a
scaled-down version of the target sites and defining each normal vector as the vector from each scaled-target back to its corresponding target. b) Targets and normal vectors
across structures. Each target site (blue) on a structure has its own normal vector approximated as a vector that starts from the target site and points outwards (grey line
segment) from the structure. c) Probe positioning. Examples of different structures labelled with an antibody probe.

Martínez et al.

|

VLab4Mic

25

Sup. Fig. S9. Modelling binding heterogeneity and probe steric hindrance. a) Labelling efficiency. We model labelling efficiency during probe positioning by taking it as
the probability of success in a Bernoulli trial per target site. Shown with a linker probe while varying the labelling efficiency from 20% to 100%. In this example, probe binding
also takes into account proximity to neighbouring targets. b) Probe steric hindrance. We model steric constraints as the minimum distance required for neighbouring epitopes
to be available for labelling. For each probe, this distance is calculated from the maximum pairwise distance between its emitters. Shown are examples using no hindrance,
different fixed values for hindrance, and the estimated distance from the probe model itself (this calculation is used by default).

Sup. Fig. S10. Distribution of radii fit on simulated and experimental images. Image overlays show simulated and experimental data available (36) for a) STED or b)
SMLM, together with their histogram of fitted radii. Parameters were the same per modality. Scale bars: 250 nm.

26

Martínez et al.

|

VLab4Mic

Target sites on structure
(AF-P00520-F1-model_v6)

Virtual Sample
(expanded 10x)

Simulated SMLM Image

0

60
40

200
Nanometers

20
0
20
40

400
600
800

60
80
80 60 40 20 0 20 40 60
Angstroms

1000
500 nm

Sup. Fig. S11. Simulations of AlphaFold models. VLab4Mic can use AlphaFold models as structural inputs for simulations. Here we show an example of an AlphaFold
model for the Striatin-interacting protein 1. Left: target sites for Lysine residues (magenta) and atoms from the structure (grey). Middle: virtual sample with a single copy of the
labelled structure. The sample was expanded isotropically with a factor of 10 times. Right: SMLM imaging simulation of the virtual sample.

Martínez et al.

|

VLab4Mic

27

a)

b)

Sup. Fig. S12. Site-specific labelling and multicolour imaging. VLab4Mic allows users to define specific residue numbers and chains to be targeted within a protein
complex. a) Single-channel site-specific SMLM. Human homotrimeric proliferating cell nuclear antigen (PCNA) labelled with a probe targeting SER-186 in each chain of the
trimer. The leftmost panel shows the target sites for SER-186 residues (magenta) and atoms from the structure (grey). Using these target sites, we show example SMLM
imaging simulations of the site-specific labels with resolution of 2 nm and 1 nm. b) Multicolour site-specific SMLM. Three different probes are specified for multicolour
imaging: each was parameterised to target a different chain of the PCNA trimer at residue SER-186. This results in each chain being imaged in a different channel. For
visualisation purposes, the virtual sample was expanded isotropically with a factor of 5. The 3-colour SMLM imaging simulation is shown both as independent channels (each
one corresponding to a different chain in the homotrimeric antigen) and as a composite image.

28

Martínez et al.

|

VLab4Mic

a)

Nuclear Pore
GFP +
Nanobody

b)

Clathrin
Primary +
Secondary AB

c)

T4 Capsid
NHS-ester

d)

HIV-1 Capsid
Antibody

Targets
Emitters

Sup. Fig. S13. Diverse biological structures and labelling strategies. Four representative examples demonstrate versatility. a) Nuclear pore complex (7R5K).
GFP-nanobodies targeting Nup96. b) Clathrin-coated pit (1XI5). Primary-secondary antibody pair targeting clathrin heavy chains. c) T4 bacteriophage capsid (8GMO).
NHS-ester conjugation. d) HIV-1 capsid (3J3Y). Antibody labelling of p24. These examples encompass diverse structural scales (from 40 nm clathrin cages to 100 nm
phage capsids) and geometries: lattices (clathrin), icosahedral symmetry (T4 phage), conical fullerene capsid (HIV-1), and symmetric ring structures (nuclear pores). Each
labelling strategy introduces characteristic fluorophore displacement patterns. GFP-nanobodies provide tight labelling, while primary-secondary antibody pairs introduce larger
displacements that affect apparent structure size in super-resolution images. Scalebar on labelled structures is 50 nm; scalebar for image simulation is 250 nm.

Martínez et al.

|

VLab4Mic

29

Sup. Video 1. Complete workflow of a virtual microscopy experiment with VLab4Mic. (▶ Watch video) Example use of a VLab4Mic experiment for selection of structure,
probe, virtual sample parameters, imaging modalities and acquisition parameters to generate multimodal image simulations.

Sup. Video 2. Parameter sweep analysis with codeless notebooks. (▶ Watch video) Example use of the VLab4Mic parameter sweep functionality to generate extensive
simulation datasets from parameter combinations. The resulting dataset is analysed by comparing each parameter combination output against a reference standard.

30

Martínez et al.

|

VLab4Mic

Sup. Video 3. Use of AlphaFold-predicted models in VLab4Mic. (▶ Watch video) A model predicted by AlphaFold can be used as an input structure for VLab4Mic modelling
and simulations. Models from other databases or custom models made by users can also be used as input structures.

Martínez et al.

|

VLab4Mic

31