Predict single-cell responses to unseen perturbations

What this involves

LabCompass learns a velocity field that transports control cells to perturbed cells conditioned on a perturbation representation. Once trained, you can pass a control population through the forward model with any perturbation vector to simulate how those cells would respond under an unseen condition, dose, combination, or cellular context.

The model operates in a reduced representation space, typically principal components derived from log-normalised highly variable genes, pairing controls and perturbed cells across batches via optimal transport. To predict truly novel perturbations or compounds rather than measured conditions, perturbation vectors must carry informative biological or chemical features, such as molecular fingerprints or gene embeddings, rather than simple one-hot encodings.

Doing this work rigorously requires holding out context or perturbation data and benchmarking predictions against trivial baselines: the unperturbed control state and a linear mean shift observed across other contexts. On shared cellular responses, standard mean shifts can perform comparably or better, making rigorous baseline comparison essential before trusting forward predictions.

How it goes

1. 1Flow matching configurationInitialise LabCompass FlowMatching with optimal transport coupling, set the device to CPU, define neural velocity field dimensions, and supply validation splits to track training.
2. 2Velocity field trainingExecute FlowMatching training across multiple CPU threads using PyTorch, saving the model object to disk with FlowMatching.save once the step target is reached.
3. 3State prediction and gene reconstructionFeed held-out control principal components and condition vectors into the predict method, mapping predicted coordinates back to gene expression using the PCA loading matrix.
4. 4Baseline comparison and scoringCalculate no-change and mean-shift baselines on held-out cells, evaluating R-squared and change correlations on overall genes and the top 100 perturbed genes.

Start it here

Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.

What to watch for

• Flow matching models out of the box may fail to beat an empirical mean-shift baseline on shared perturbation responses across cell types.
• Responses unique to an unobserved cell context cannot be learned if perturbations are supplied only as one-hot condition vectors.
• R-squared across all genes is elevated because baseline expression dominates; use the correlation of predicted versus real changes on top responsive genes to distinguish model quality.
• LabCompass defaults to CUDA execution, so device_id must be explicitly set to CPU on non-GPU environments.
• The model save function raises a FileNotFoundError if the target directory does not exist, and raises a runtime error on subsequent writes unless overwrite is set to true.
• Never attach WandbCallback during execution, as it attempts external authentication and data transmission to Weights & Biases.

What it runs

Questions