Loading…
Drug-conditional adapter inside a frozen single-cell foundation model, predicting transcriptional responses to unseen drugs and cell lines.
No providers recorded yet. Browse all providers
Predicting how cells will respond transcriptionally to a compound they have never encountered is the modeling problem underneath virtual screening and mechanism-of-action work. It is hard for a specific reason: chemical perturbation datasets are small, spanning hundreds of molecules across a handful of cell lines, while the response space is tens of thousands of genes. Models trained from scratch on that data overfit, and the models that do generalize tend to encode cell lines as learned identifiers, which makes a genuinely new cell line impossible to score.
scDCA — the single-cell Drug-Conditional Adapter, developed at Genentech — takes a different route. Rather than training a perturbation model, it borrows one that already understands cell state. scGPT was pretrained on tens of millions of single cells covering many cell types, states, and disease annotations, but it has never seen a molecular structure; chemical perturbation lies outside its pretraining modality entirely. scDCA bridges that gap by inserting small adapter modules into each of scGPT's transformer blocks and conditioning them on molecule embeddings, while the original transformer weights stay frozen.
The design intent is precise. Passing a molecule embedding straight into scGPT's input would shift the input distribution away from anything seen during pretraining and invite overfitting on a few hundred compounds. Routing the chemistry through adapters instead keeps the backbone's input a gene expression profile, exactly as in pretraining, and confines all molecular conditioning to the small trainable component.
Each adapter comprises a molecular projection layer, a down-projection, a residual block, and an up-projection; the molecular projection is a two-layer network mapping the compound embedding to bias vectors for the two projections. Molecular embeddings come from a frozen ChemBERTa, which reads SMILES strings and was pretrained on 77 million PubChem compounds, so structural similarity enters as prior knowledge rather than being learned from the perturbation data. Input expression is log1p-normalized and tokenized with a feed-forward network rather than discretized, following scGPT's genetic-perturbation setup.
Experiments use the sciplex3 dataset: 649,340 cells across 3 human cancer cell lines (A549, MCF7, K562) perturbed with 188 drugs, with the analysis restricted to 2,000 drug-sensitive genes. Cells sharing a drug-cell-line combination are mean-aggregated into pseudobulk "metacells", so predictions are made at that resolution rather than per individual cell. Performance is reported as R² over the top 20 differentially expressed genes, averaged across five random splits. Against naive full fine-tuning of scGPT, scDCA scores 0.81 versus 0.81 on unseen drugs, 0.83 versus 0.78 on unseen drug-cell-line combinations, 0.88 versus 0.81 on few-shot unseen cell lines, and 0.82 versus 0.51 on zero-shot unseen cell lines — the gap widening exactly where generalization is hardest. It also outperforms chemCPA, BioLORD, and SAMS-VAE across all four tasks; chemCPA cannot attempt the zero-shot cell-line task at all, because its architecture depends on trained cell-line embeddings.
The direct use is in silico screening: ranking compounds by their predicted transcriptional effect before committing to a multiplexed perturbation experiment, and extending predictions to cell lines for which no perturbation data exists but a control profile does. That second capability matters for target work in disease-relevant lines that are expensive or impractical to screen. The adapter approach also fits groups with modest compute, since only a small module is trained on top of a checkpoint downloaded once.
scDCA is a compact demonstration that a single-cell foundation model can be extended to a modality it never saw in pretraining, and that doing so through conditioned adapters beats fine-tuning the whole network — decisively so out of distribution. Some limits should be read alongside the results. Evaluation runs on pseudobulk metacell aggregates per drug-cell-line pair, so the zero-shot cell-line claim is about aggregate response rather than single-cell resolution. The method is described as generally applicable to transformer-based single-cell foundation models, but only the scGPT instantiation is demonstrated, and neither code nor adapter weights have been released. The name is also shared with an unrelated cell-cell communication tool in Briefings in Bioinformatics; this entry documents Genentech's perturbation model.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.