Drug Discovery
14 gains2 Traces8 problems
Every measured gain and problem on the record for this topic, newest first. Updated .
One square per record, darker for stronger evidence:Evidence: 0 high15 moderate9 emerging
AI in drug discovery
Most evidence on AI in drug discovery comes from lab benchmarks, cell assays and animal studies, with very little from patients. TRV-2026-0893TRV-2026-0940TRV-2026-2408 Authors of a Nature Reviews Drug Discovery Perspective say that after a decade, evidence of safer, more effective medicines reaching patients faster is disappointingly limited. TRV-2026-0733
A randomized, placebo-controlled trial in 51 healthy adults shows the AI-optimized COVID antibody GB-0669 caused no dose-limiting toxicities and had a 54-day half-life. TRV-2026-0225 In a controlled comparison, 85% of 100 AI-optimized antibiotic peptides were hits, and they outperformed their templates in two mouse models of Acinetobacter baumannii infection. TRV-2026-0193 On trial paperwork, the TrialScout tool linked registrations to published results with 92.5% sensitivity, and human error explained 123 of 200 reviewed disagreements with earlier matches. TRV-2026-0918
An AI literature screener scored an F1 of 0.684 on drug-drug queries, against 0.872 for drug-family queries, in a test on 7,751 articles. TRV-2026-1610 A GPT-4.1 tool reading consent forms matched experts on 92.1% of forms after tuning, which leaves a share of answers wrong. The authors note that hallucinations limit reliability in high-stakes settings. TRV-2026-1539 An AI tool that masks patient identifiers reached 88.48% average accuracy, which means some trial documents were not correctly desensitized. TRV-2026-1530 A review in the International Journal of Pharmaceutics says no AI-designed polymeric-lipid nanoparticle for breast cancer has entered a registered clinical trial. TRV-2026-1298
The anticancer prediction model was validated externally on only 10 FDA-approved compounds, and its ranked predictions for 1200 drugs are unconfirmed. TRV-2026-1723 TrialTriage matched the answer key on all 90 synthetic cases, but the authors say real-world performance is not shown. TRV-2026-0666 The AI-designed bird flu vaccine protected mice and ferrets in an animal study, and it still has to be tested in human clinical trials. TRV-2026-2408 An AI framework flagged 817 terminated trials that may have failed from unwanted drug spread in the body, though no drug has been shown revived. TRV-2026-2336
The strongest results are a small human safety trial of an AI-optimized antibody and animal studies of peptides, a vaccine and prostate cancer drugs. TRV-2026-0225TRV-2026-0193TRV-2026-2408TRV-2026-1644
See these 5 recordsModels rank well on benchmarks and turn up lab hits, but the work is preclinical and some methods are not built for large screens. TRV-2026-0940TRV-2026-1348TRV-2026-1663TRV-2026-1723TRV-2026-1954
See these 4 recordsLab tests plus machine learning flagged liver-toxic drugs in a 170-drug test, while reviews say black-box toxicity models struggle to win regulators' trust. TRV-2026-1295
A Trace is one record with both a gain and a problem, so it counts on both sides: 16 records show a gain and 10 show a problem. 20 context records are in the record list but not counted, because they report no measured gain or problem.
| Evidence | Gain side | Problem side |
|---|---|---|
| HighStrong study design, backed by at least two independent sources | 0 | 0 |
| ModerateA measured study or an official source | 12 | 3 |
| EmergingEarly, small, unreviewed or reported second-hand | 4 | 7 |
By type of evidence: 9 benchmark or lab tests; 6 controlled comparisons; 6 commentaries or narrative reviews; 1 news report; 1 investigation by reporters; 1 randomized trial. How evidence is scored
Against earlier human-coded matches, the TrialScout AI tool linked trial registrations to published results with 92.5% sensitivity and 81.2% specificity, and human error explained 123 of 200 reviewed disagreements.
In a check of 9,000 randomly chosen samples from 100,000 processed trial image records, an AI tool masked patient identifiers with 88.48% average accuracy, reaching 99.40% on serious adverse event documents.
In a placebo-controlled trial of 51 healthy adults, the AI-optimized antibody GB-0669 caused no dose-limiting toxicities, had a 54-day half-life, and showed dose-dependent virus neutralization in blood samples.
In a test on 7,751 articles, the AI literature screener's F1-score was 0.684 for drug-drug queries and 0.731 for single-drug queries, against 0.872 for drug-family queries.
With her liver failing, her skin was a ghostly apricot. The thoughts creep in as we are bombarded by artificial intelligence chief executives making bold proclamations such as “AI will cure cancer in five to 10 years” or “AI will cure cancer in our lifetimes”.
AI tools for matching results, screening patients, reading consent forms and masking identifiers do well in tests but still make errors. TRV-2026-0918TRV-2026-1530TRV-2026-1539TRV-2026-0666TRV-2026-2336
See these 4 recordsReviews and reporting say data problems, weak interpretability and few clinical trials leave the claim of better medicines largely unproven. TRV-2026-1172TRV-2026-0733TRV-2026-0893TRV-2026-1151TRV-2026-1298TRV-2026-0356
See these 5 recordsEach record sits in one thread at most. A record with both a gain and a problem counts on both sides. 3 records fit none of these threads.
The review says poor data quality, limited interpretability and poor generalizability still limit reproducible nanoparticle synthesis and clinical adoption of AI-driven nanomedicine.
4 records
Showing the records in one thread of the summary: Computational screening and benchmarks. Show every record
Researchers introduced qcGEM, a graph-based molecular embedding that builds quantum chemistry information into AI models for drug discovery. Pre-trained on 300 thousand molecules from the qcMol database, it was compared with 32 widely used molecular representations under a frozen-representation plus probe-network…
Advanced Science
Researchers built DeepFusion-ACSM, an ensemble that combines machine learning and deep learning models to predict whether small molecules have anticancer activity.
Researchers built a machine learning framework around a graph neural network classifier called AI2IMpred and used it to screen more than 7000 plant-derived compounds for the ability to interfere with the AI-2 quorum…
Researchers built a virtual screening funnel that combined a Bayesian Ridge regressor trained on ChEMBL CDK4/6 inhibitor data with structure-based docking and molecular dynamics.
ace