The drug evidence disagrees with itself

The same drug, different answer

Drug discovery runs on perturbation data — measurements of how cells respond to a drug. A preprint posted this week tested six major drug-perturbation transcriptomic resources and found that while treated expression profiles were consistent, the drug-induced changes themselves were poorly reproducible across datasets, even under matched experimental conditions. The average response across cell lines — the drug main effect — proved a robust predictor of cell-specific interactions, but control-expression-only baselines such as Tahoe failed to resolve substantial drug-specific variation. m:…7891

The finding is awkward for bio-AI: the datasets that train drug-response models disagree about the very thing the models are meant to learn. A model trained on one resource and tested on another is not failing at biology. It is failing at a measurement that was never standardised. m:…7891

Trials contradict each other

Randomised trials disagree with each other too, and a preprint introduces DrugCon, a framework that finds and explains those contradictions at scale. It scored an F1 of 0.744 for detecting contradictory claim pairs in trial abstracts, beating existing natural-language and rule-based baselines. Three quarters of the contradictions — 75.5% — involved more than one study-level factor, most often differences in outcome measures (64.0%), populations (51.3%) and dosage or exposure duration (35.6%). m:…8354

Contradiction, in other words, is usually not a mistake. It is two trials answering different questions and being read as if they answered the same one. A tool that names the differing factor — the dose, the population, the endpoint — turns a confusing literature into a map of what is actually known. m:…8354

Genetics predicts a little

The same modesty applies to predicting who responds to antidepressants. A genome-wide meta-analysis of 62,879 people with depression, posted as a preprint, found that self-reported response to SSRIs has a SNP-based heritability of 3.1%, with no single genome-wide significant variant. Polygenic scores for ADHD, PTSD, autism and addiction were associated with lower odds of reported response, and gene-set analyses implicated GABAergic and glutamatergic neurotransmission, including the gene SLC17A6. m:…3780

Three percent heritability is not nothing, and it is not a test. It means genetics can tilt the odds slightly, but the largest part of whether someone reports responding to an SSRI sits outside the genome — in the drug, the dose, the diagnosis, and everything a questionnaire cannot capture. m:…3780

The audit is being automated

The week's fourth preprint offers the other side of the coin: automation that enforces rigour instead of cutting corners. MedCheckpoint, a multi-agent language-model system for postmarketing pharmacovigilance, grounds its analysis in traceable sources and locks in study design. Validated against a 2.7 million-patient simulated electronic health record benchmark, it outscored baseline coding agents on report quality — 4.81 to 4.88 against 4.25 to 4.55 and 4.02 to 4.29 — and was far more consistent run to run. It also refused misleading analyses of infeasible comparisons or unmeasurable outcomes rather than producing a confident wrong answer. m:…3779

The pattern across the four papers is one of a field learning to distrust its own inputs: drug-response data that does not reproduce, trials that contradict each other, genetic predictions that explain a little. The response is not less automation. It is automation aimed at the evidence itself — checking the dataset, naming the difference between trials, refusing the unanswerable question. m:…7891 m:…8354