Biotech software news summary, 1 November

Today in brief

In a study of 50 analysts using Claude Code on a fictional farming-cooperative analysis, nearly all followed the agent’s randomly assigned preferences.1 A cheminformatics roundup covered assay-aware models, reinforcement learning for molecule generation and a protein–ligand docking dataset.2

AI’s influence on analysts

Nearly all 50 professional analysts in a study adopted an AI agent’s randomly assigned preferences while assessing an advisory program for a fictional farming cooperative.1 More than three-quarters repeated the agent’s reasoning in their explanations, sometimes nearly verbatim, while 15% mentioned AI.1 About half of the surveyed analysts later considered the final recommendation more their own idea than the AI’s.1

Computational chemistry studies

Assay-aware models were evaluated across 167 G protein-coupled receptors and more than 5,000 assays.2 When assays appeared in both training and test data, one-hot assay identifiers outperformed BioBERT embeddings. For unseen assays, embeddings offered a modest improvement.2 Leaving out whole assays reduced performance more than scaffold splitting did.2

A reinforcement-learning approach raised the share of generated KRAS samples that were unique hits in the Enamine REAL library from about 10% to about 25%.2 The approach aligned the pocket-conditioned DrugFlow model with the Enamine REAL and Freedom 4.0 libraries.2

DockBind contains more than 18,000 modeled protein–ligand complexes paired with experimental affinity measurements, covering 612 targets.2 In an independent cross-docking benchmark, the workflow retained roughly 80% of recovered poses.2 The dataset includes modeled poses and uneven target representation.2