SMLM Contrastive Encoder

A single-molecule localisation microscopy (SMLM) experiment returns a list of coordinates, not a picture, and the conventional way to say something about that list is to choose a statistic — a Ripley's K curve, a DBSCAN cluster count, a Voronoi polygon area — and compare its histogram between two conditions. That choice is made before the data is examined, and one dataset at a time, so experiments summarised with different descriptors cannot be placed on a common axis.

This model replaces the hand-chosen descriptor with a learned coordinate. Each 3 × 3 µm region of interest is rendered as a 2D histogram of localisations and passed through a convolutional encoder that maps it to a 128-dimension vector in which Euclidean distance tracks structural similarity. Because every ROI from every experiment lands in the same space, clustering, dendrogram distances and trajectory analysis become ordinary vector operations. The authors frame this as nanoscale spatial-omics: localisation-level protein architecture treated as a measurable layer alongside genomics, transcriptomics and proteomics.

The encoder was developed at the University of Birmingham with collaborators in Denmark, Germany, Portugal and elsewhere in the UK, and released as a bioRxiv preprint in October 2025. It is trained once, on simulated point patterns alone, then applied unchanged to experimental data; the trained checkpoint ships inside the project's Python package.

#Key Features

• Simulation-only pretraining: positive pairs are two point patterns drawn from the same Perlin-noise generative parameters rather than two augmentations of one image, so the encoder learns structural invariance without ever seeing a labelled experiment.
• One shared latent space: every ROI, from any protein, cell type or instrument, receives a comparable 128-dimension vector, which is what makes cross-experiment comparison possible.
• Parameter-weighted contrastive loss: an adapted NT-Xent objective down-weights negative pairs in proportion to how similar their generating parameters were, so capacity is spent on genuinely distinct patterns.
• Built-in quality control: replicate batches and alternative localisation-fitting algorithms separate in the latent space, surfacing batch effects and workflow-dependent bias.
• Encoder-only inference: the projection head used during training is discarded, leaving a small convolutional encoder that runs on CPU.

#Technical Details

The encoder is four convolutional layers (3 × 3 and 2 × 2 kernels, feature depth rising from 1 to 8 to 16 channels) followed by a single fully connected layer, with Mish activations, batch and layer normalisation, dropout of 0.2, and average pooling before flattening; the output is a 128-dimension embedding. A three-layer projection head lifts that embedding to a 512-dimension output during training. Training used 15,000 pairs (30,000 point patterns) generated by the FractalPerlin2D function of the pyperlin package, split 80/20, optimised with Adam at a learning rate of 3 × 10⁻⁵, weight decay 1 × 10⁻⁶ and batch size 32, with plateau-based early stopping typically halting at 30–40 epochs.

On experimental SMLM of five proteins drawn from the Nano-org curated database — tubulin, actin, KIR2DL1, TIGIT and Lck — K-means clustering in the 128-dimension space reached an Adjusted Rand Index of 0.74 (95% CI 0.71–0.78), cleanly dividing fibrous from non-fibrous architectures; normalising every ROI to 100 localisations/µm² reduced this to 0.55 (0.51–0.59). A nocodazole dose–response on COS-7 microtubules gave an ARI of 0.49 (0.44–0.56), and resting versus 5- and 15-minute EGF stimulation of EGFR in HeLa cells gave 0.33 (0.30–0.37). Separation was weak for microtubules across COS-7, HeLa and HEK cells (ARI 0.25, 0.21–0.30) and weaker across fluorophores (0.20, 0.17–0.24), which the authors read as microtubule architecture being largely conserved across those lines rather than as a failure of the embedding. UMAP is used for visualisation only; all quantitative clustering is performed in the full 128-dimension space.

#Applications

The embedding suits any question that requires comparing nanoscale organisation rather than measuring it: clustering proteins by architecture, ranking the structural distance between treatment conditions, tracking a receptor's reorganisation along a drug dose series, or staging a process such as immunological synapse maturation. Because it also resolves purely technical variation, it doubles as a quality-control instrument for SMLM facilities — detecting a divergent replicate, or showing that a change of localisation-fitting algorithm has shifted the apparent structure of unchanged raw data. Integration with the Nano-org public data resource lets embeddings be computed over datasets contributed by other groups.

#Impact

Comparative analysis has been the missing capability in SMLM: the field holds large volumes of localisation data that no common representation could relate. By supplying a fixed encoder that any laboratory can apply to its own point clouds under an MIT license, this work makes nanoscale architecture something that can be pooled across studies — the precondition for the cell-atlasing the authors propose. The limits are stated plainly in the paper: the analysis is two-dimensional, the fixed 3 × 3 µm ROI will not suit every protein architecture, results remain sensitive to fluorophore properties and labelling strategy, and the work is a preprint awaiting peer review.