TF target prioritisation

I worked on this with a small team at a weekend hackathon on agentic AI for drug discovery. The question was how to turn a long list of transcription factors linked to a trait into a short, defensible list of drug targets.

Each factor gets six lines of evidence: cell-type specificity, regulatory-network connectivity, available modalities, clinical precedence, target accessibility, and redundancy. Redundancy counts against it, because a paralog can take over its job when you hit the target. The scores are standardised and combined with weights, then rule-based flags such as undruggable or high redundancy point out concerns that a high rank would otherwise hide.

The weights start as expert priors. Given known drug-target successes and failures, a Bayesian step can refine them.

The code is private for now. The game uses made-up factors and weights, not the pipeline output.