SynthMLM

Single-molecule localisation microscopy (SMLM) sits in an awkward regime for machine learning. One experiment yields millions of localisation coordinates, but the number of biologically distinct, well-annotated datasets in existence is far too small to train a data-hungry model on experimental images. The usual alternative — summarising a dataset with a hand-chosen statistic — fixes the question the analysis can answer before the data is examined.

SynthMLM inverts the problem. A parametric simulator builds localisation patterns from knobs that are interpretable biophysical quantities — fibre length, width, density and curvature; ring radius and linear density; cluster occupancy — and a convolutional network trained once on ten million of its images reads those knobs back out. Because the network's outputs are exactly the generator's inputs, the loop closes: descriptors inferred from an experiment feed back into the simulator to produce a matched synthetic companion dataset, or are interpolated between measured conditions to synthesise one nobody imaged.

The framework comes from Dylan Owen's laboratory at the University of Birmingham, written in Julia with Flux.jl, integrated into the group's nano-org SMLM data resource, and posted as a bioRxiv preprint in August 2026. It follows the same group's SMLM Contrastive Encoder, whose embeddings are used here to check that synthetic data lands where its real counterpart does. Only the analyser carries learned weights: the generator is explicit geometry, not a learned generative model.

#Key Features

• Descriptors as the representation: the network predicts 30 quantities — counts of fibres, rings, larger clusters and low-occupancy puncta, plus the mean and variance of 13 structural descriptors — each a physical measurement rather than a latent coordinate.
• Trained on simulation, applied unchanged: the training corpus deliberately spans mixed and even experimentally unobserved morphologies, and the checkpoint is applied to microtubules, receptor clusters and nuclear pore rings without refitting.
• Closed analysis-to-generation loop: because the regression targets are the simulator's own parameters, inferred distributions can be sampled to build companion datasets at either dataset level or image level.
• Interpolation between conditions: descriptor distributions from several measured drug doses can be interpolated to produce localisation data for an intermediate dose nobody acquired.
• Explicit rather than black-box generation: unlike adversarial, variational-autoencoder or diffusion approaches, every synthetic image arrives with the parameter set that produced it — which is what makes it usable as labelled ground truth.

#Technical Details

The image analysis model is a RegNetX-based convolutional network of approximately 27 million parameters performing multi-output regression. It accepts a 3 × 3 µm localisation pattern binned into a 100 × 100 image, with pixel intensities clamped at 40 and min-max scaled to [0, 1]; predictions are returned to physical units and aggregated across images into dataset-level distributions. Training used 10 million synthetic images with min-max normalised targets, AdamW with a OneCycle schedule and mean squared error loss, run for 20 epochs with model selection on 100,000 held-out images. The generator draws clusters as Gaussian point clouds, rings as fixed-radius circles whose centres are relaxed by a physics simulation to remove overlaps, and fibres as particles scattered along Bézier-curve scaffolds.

Agreement between real and synthetic data was quantified with the normalised Kolmogorov–Smirnov dissimilarity score used by nano-org, where a value above 1 marks two datasets as significantly different. Across every dataset on nano-org as of 1 May 2026, each paired against a 500-sub-image synthetic companion, the mean score was 0.676. On a nocodazole dose series in COS-7 cells the inferred descriptors tracked microtubule disruption — fewer, shorter, less dense fibres and more cluster-like structures — and an interpolated 0.5 µg/mL dataset occupied the same region of contrastive-embedding space as the real one, with the reduced spread expected from smoothing between descriptor distributions.

#Applications

The most direct use is benchmarking: SMLM cluster-detection and fibre-analysis algorithms can be run against data whose structural parameters are known by construction, with failure modes probed by sweeping a single descriptor. The generator also supplies labelled training data for other machine-learning workflows in nanoscale imaging. As an analysis tool, the descriptor distributions let imaging laboratories compare treatments, cell types or replicates in physical units, and interpolation fills in a dose series before microscope time is committed.

#Impact

SMLM holds far more raw localisation data than annotated data, and SynthMLM offers a route around that shortage that keeps the representation legible to a biologist. The authors are explicit about the boundary: the framework mimics experimental SMLM output, not the ground-truth positions of proteins in the sample, and its synthetic datasets are not a substitute for independent experiments. The network is a fixed-purpose estimator over three element classes in two dimensions rather than a general representation model, and exposes no embedding for downstream transfer. The software has not been released — the authors state it will appear as an open-source Julia package on GitHub when the article does — and the work remains a preprint.