AI drug discovery is getting sharper — on paper

Knowing how sure the model is

Drug discovery depends on knowing when a prediction can be trusted. A preprint benchmarks uncertainty quantification for ADMET prediction, a critical task in drug discovery. Ensemble MC dropout achieved the best average rank for predictive performance, Brier score and 90% interval score across 22 ADMET datasets. LightGBM, the simpler baseline, stayed competitive on performance and Brier score but performed poorly on specific endpoints. m:…1202

The warning in the paper is about evaluation design. Uncertainty performance is highly sensitive to how the evaluation is set up, and small-scale evaluations often fail to recover the broader performance trends across chemical partitions. A method that looks best on a small benchmark may not be best at all. m:…1202

The benchmark's design is the point: the ranking reflects performance across chemical partitions rather than a single split, and small-scale evaluations often fail to recover the broader performance trends. The authors' warning is aimed at researchers picking a method on the basis of a small test — the ranking can shift when the evaluation grows. m:…1202

Knowledge that evolves

KnowEvo, a second preprint, takes a different approach to the same pipeline problem: instead of prompting a model as one monolith, it represents scientific knowledge as modular, executable blocks of expertise, organised into strategy mutation trees and refined by a code-agent feedback loop in a sandbox. Forward allocation follows a Thompson-style rule; backward revision refines the blocks that underperform. m:…8177

On protein stability and solubility tasks, KnowEvo outperformed monolithic program evolution and prior LLM-guided evolution baselines. The claim is about structured, reusable knowledge beating one-shot prompting — but it is measured on benchmarks, and the paper reports no wet-lab test of the proteins its evolved strategies produce. m:…8177

Agents finding biology humans missed

The third preprint sends an LLM agent into proteomics. The framework automates the construction and comparison of tissue-specific protein co-abundance networks, pulling evidence from expression atlases, interaction databases and the literature, at a scale the authors say is beyond manual curation. It identified 1,825 conserved protein clusters, with 76% of pairs sharing co-abundance relationships beyond chance. m:…7598

The check that matters is independent replication: 379 of 793 testable clusters replicated in independent CPTAC proteomic datasets. That is replication in other data, not validation in a lab — and the framework is limited to the databases and literature sources built into it. Across all three papers the pattern holds: sharper tools, measured on benchmarks and independent datasets, with the wet-lab proof still to come. m:…7598

The payoff the authors point to is conserved biology across organ systems: clusters conserved across tissues suggest patterns that single-dataset analysis cannot see. The agent's job is to generate those hypotheses with evidence attached, not to prove them — the proof, as with the other two papers, is still ahead. m:…7598

The framework is designed to find patterns across tissues — comparing networks across tissues rather than studying one tissue at a time — that single-dataset analysis cannot reach. The authors present it as a scalable way to find shared biology across organ systems. That is the promise; the evidence in the paper is computational replication, and the lab work that would confirm a target is not part of this study. m:…7598