Discovery of candidate therapeutic targets with Geneformer.

Geneformer enables context-aware gene network predictions from limited single-cell transcriptome data, using in silico perturbations to rank candidate therapeutic targets by embedding shifts.

• Why it matters: Mapping gene connections can reveal disease-disrupted networks, but target discovery is hard when data are scarce. Geneformer is designed to work with limited data by leveraging large-scale single-cell pretraining.
• What they did: The pipeline tokenizes raw gene expression counts into rank value encodings, then uses zero-shot embeddings, optional fine-tuning (single or multi-task), and in silico perturbation simulating gene repression or activation. It evaluates models with confusion matrices, macro F1 scores, and embedding analysis, and can use a quantized model for efficiency.
• The result: It outputs fine-tuned, context-specific predictive models and rank-ordered perturbation predictions that aim to induce target cell states, typically completing in under 2 days on a standard GPU workstation. This supports practical therapeutic target prioritization under moderate computational and coding requirements.

AI summary of the abstract · may contain errors · not medical advice · read the paper.