Plexus

Self-supervised learning has transformed high-content phenotypic screening, but almost entirely for static readouts — cell morphology and transcriptomic state. Neuronal activity is the hardest dynamic case: the phenotypes that matter in epilepsy and neurodegeneration are properties of a network of synaptically coupled cells, not of any neuron alone. Calcium-imaging screens have leaned instead on hand-engineered signal-processing features that capture only what someone thought to measure.

Plexus is a self-supervised transformer built specifically for this modality. Developed at UCSF by the Kampmann, Keiser and Yala labs with collaborators at GSK's Laboratory for Genomics Research and UC Berkeley, it encodes GCaMP6m calcium activity traces from human iPSC-derived neurons into single-cell embeddings that are contextualized by the surrounding network. Where a morphology profiler such as OpenPhenom treats each cell as an independent image crop, Plexus takes a subnetwork of neurons as a joint input, so synaptic phenotypes surface in the representation of every cell in it. The work appeared as a bioRxiv preprint in February 2025 and was published in Nature Machine Intelligence in December 2025.

The model is one half of an integrated platform: iPSC-derived excitatory neurons co-cultured with astrocytes, carrying both the GCaMP6m indicator and CRISPRi machinery, imaged at 25 Hz in arrayed 384-well plates.

#Key Features

• Network-contextualized single-cell embeddings: A subnetwork of 8 neurons passes jointly through the encoder; tokens of the cell of interest are mean-pooled into a 768-dimensional vector, averaged over 10 sampled subnetworks to give one embedding per cell.
• Permutation-invariant cell identity: Neurons have no canonical ordering, so Plexus ranks cells within a subnetwork by the L2 norm of their mean token and uses that rank to index into learned cell-identification embeddings, so any permutation yields the same representation.
• Masked reconstruction across cells and time: Half the tokens are masked, dropping temporal chunks of activity across the whole subnetwork for a decoder to reconstruct. This extension of the masked-autoencoder objective forces network information into the latents rather than allowing per-cell shortcuts.
• Ground-truth simulation benchmark: A multivariate self-regulating Hawkes process on Watts-Strogatz graphs generates eight activity phenotypes with controlled synaptic strength, excitability and self-regulation, rendered as synthetic GCaMP6m signals — the ground truth that in vitro screens lack.
• Fully open release: Model, trace-extraction and simulation code ship as three MIT-licensed repositories, with checkpoints, traces and embeddings on Zenodo under CC-BY-4.0.

#Technical Details

Each neuron's calcium trace is chunked into 24-frame windows (0.96 s at 25 Hz) and linearly projected to 768 dimensions, giving 50–100 tokens per cell. Tokens carry a learnable temporal position embedding plus the permutation-invariant cell embedding, with a CLS token and five register tokens appended to the concatenated subnetwork sequence. Training runs 1,000 epochs at a 50% masking ratio on a well-aware 80/20 split. Simulation models trained in 4–24 GPU-hours on a single 24 GB NVIDIA TITAN RTX; the screening model took 78 GPU-hours across four.

On linear probing of the eight-class simulated phenotype task, Plexus reaches 0.9729 ± 0.0027 accuracy with 16-cell subnetworks (0.964–0.969 for 4- and 8-cell) against 0.8901 ± 0.0025 for engineered features and 0.8568 ± 0.0093 for a single-cell MAE; a 1D CNN autoencoder and a pretrained time-series foundation model also fell below the engineered baseline. Elastic-net classifiers separating neuroactive treatments from controls reached AUROC 0.98–1.00 across 2 mM Mg²⁺, TeNT, 2 mM Ca²⁺ and TTX, versus 0.93–0.99 for engineered features. In a 52-gene CRISPRi screen, Plexus embeddings yielded 18× (WTC11) and 27× (patient line) more perturbations clearing an AUROC of 0.8. Nuclei of transduced cells were segmented with a fine-tuned Cellpose model.

#Applications

The platform targets genetic modifier screens against neuronal activity phenotypes. KCNQ2, a voltage-gated potassium channel implicated in childhood epilepsy, served as a positive control: a covariance-based analysis mapped its learned axis onto peak width at half maximum, recovering an interpretable biophysical signature from the embedding. Applied to isogenic MAPT V337M pairs modelling frontotemporal dementia, over 36,000 single-cell traces were projected onto OPLS disease and off-target axes, nominating MAT2A, BIN1 and GRIA2 as knockdowns that shift mutant-tau neurons toward the wild-type phenotype with minimal off-target movement. The approach suits any screen where cellular context shapes a dynamic single-cell readout.

#Impact

Plexus extends self-supervised phenotypic profiling from static morphology and transcriptomics to dynamic, network-level processes, and shows the gain comes from network awareness rather than deep learning alone: the single-cell MAE ablation underperforms hand-engineered features, while adding any contextualizing neighbours reverses that. Constraints remain: GCaMP6m reports a calcium proxy for a subset of neuronal function, arrayed screening trades throughput for phenotypic depth, and the platform measures spontaneous activity only. The encoder is also pretrained per dataset — separate checkpoints exist for the simulation, neuroactive and CRISPRi corpora — so a new screen means rerunning the self-supervised objective on its own traces rather than loading a universal model.