Biology & Medicine
27 gains3 Traces3 problems
Every measured gain and problem on the record for this topic, newest first. Updated .
One square per record, darker for stronger evidence:Evidence: 2 high27 moderate4 emerging
AI in biology and medical science
Most of the evidence here comes from benchmarks and lab tests rather than trials in patients, and the strongest results are in predicting protein shapes. TRV-2026-0203TRV-2026-0211 Authors and reviewers in other areas say their predictions still need confirmation in living systems or on wider data. TRV-2026-0288TRV-2026-2034
In the CASP14 blind test, AlphaFold matched experimental accuracy in a majority of cases, even when no similar structure was known. TRV-2026-0203 RoseTTAFold's authors report that it approaches AlphaFold2's accuracy and also models protein-protein complexes. TRV-2026-0211 The AlphaFold database's authors report it grew from 300,000 predicted structures in 2021 to over 214 million. TRV-2026-1205 On held-out questions about unseen proteins, Genolator scored an MCC of 0.959 on confirm-or-deny function questions, against 0.179 for GPT 4.1, in a benchmark test. TRV-2026-1133
Retrained on 7.8 million RNA sequence-structure pairs, SincFold scored worse on average than the published version, with training F1 of 0.983 against validation F1 of 0.420. TRV-2026-1980 A herbarium pipeline labeled 11.1 million of 22 million specimen records as having flowers, but only 2.9 million had the complete metadata researchers need. TRV-2026-1263 In a rice gene benchmark, the deep learning models overfit early and a tuned Random Forest did better. TRV-2026-1675 A review of AI bone scaffold design found small datasets and weak links between computational predictions and biological outcomes hold the field back. TRV-2026-0288
A study of lab cultures from three donors found AI reading Raman signals told exhausted T cells from active ones, but it was shown only in culture. TRV-2026-1084 A mock-task study with one expert and three trainees suggests AI hand tracking can show cell culture technique improving with feedback. TRV-2026-1519 The authors of one study rated 35 of 55 ChatGPT-4o hypotheses on fungus and cervical cancer as high quality; none were tested in a lab. TRV-2026-1343 SPROTify flagged 334 small-protein candidates in cancer data, and their function is unconfirmed. TRV-2026-2034
News reports describe 16 of 285 AI-designed virus genomes producing viable viruses that killed E. coli, including strains resistant to a natural virus. TRV-2026-0732 The same work draws a warning from commentators, who say governance to steer such a capability safely does not yet exist, though no misuse is reported. TRV-2026-0671TRV-2026-1040
This is the best-measured thread, with blind-test and benchmark results showing near-experimental accuracy for structure and strong scores on function questions. TRV-2026-0203TRV-2026-0211TRV-2026-1133
See these 7 recordsBenchmarks show sequence models doing well on curated data, while an RNA structure model retrained on millions of examples did not improve. TRV-2026-1927TRV-2026-1122TRV-2026-1980
See these 6 recordsClassifiers separate cell types and states with high accuracy in lab data, and the authors of several say clinical use is not yet shown. TRV-2026-2263TRV-2026-1084TRV-2026-2034
See these 5 recordsA Trace is one record with both a gain and a problem, so it counts on both sides: 30 records show a gain and 6 show a problem. 13 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 | 2 | 0 |
| ModerateA measured study or an official source | 25 | 5 |
| EmergingEarly, small, unreviewed or reported second-hand | 3 | 1 |
By type of evidence: 21 benchmark or lab tests; 4 news reports; 3 observational studies; 3 controlled comparisons; 1 set of official statistics; 1 commentary or narrative review. How evidence is scored
RoseTTAFold, a three-track neural network, predicts protein structures with accuracy approaching DeepMind's AlphaFold2 and generates accurate models of protein-protein complexes.
In the CASP14 blind test, AlphaFold predicted protein 3D structures from amino acid sequence alone, matching experimental accuracy in a majority of cases, even with no similar known structure.
The AlphaFold database's authors report it grew from 300,000 predicted protein structures in 2021 to over 214 million, with predictions now integrated into PDB, UniProt, Ensembl, InterPro and MobiDB.
Only 2.9 million of the 11.1 million flower-labeled records had the complete metadata needed for downstream research, and uncertainty remains about missed flowers (false negatives).
pxGPT did not clearly match human consistency, its sensitivity was uneven for some plant structures, and it is untested on other crops or imaging conditions.
Lab tests show some AI-designed viruses and enzymes work, though only a minority of designs did; commentators say safety governance is missing. TRV-2026-0732TRV-2026-0671TRV-2026-1657
See these 4 recordsLanguage models and image pipelines match experts on narrow extraction tasks, but gaps in metadata, uneven sensitivity and the need for human guidance remain. TRV-2026-1862TRV-2026-1263TRV-2026-2199TRV-2026-1820
See these 7 recordsEach record sits in one thread at most. A record with both a gain and a problem counts on both sides. 4 records fit none of these threads.
A review found AI-driven bone scaffold design is held back by small datasets, limited cross-platform validation and weak links between computational predictions and biological outcomes.
5 records
Showing the records in one thread of the summary: Extracting data and automating lab and field work. Show every record
Researchers built download-tracking infrastructure for the PRIDE proteomics archive and analysed 159.3 million download records from 2021 to 2025, covering 35,528 datasets. Their DeepLogBot framework used labels produced by two language models, Claude and Qwen3, to sort traffic into bots, institutional hubs and…
PROTEOMICS
Researchers tested seven commercial large language models on extracting structured gene, disease and phenotype information from 406 full-text papers covering 180 congenital heart disease genes.
Researchers built an agentic framework in which large language model agents fetch omics studies, extract article metadata, locate and download published data, run containerized quantification pipelines, and synthesize…
Researchers filmed one expert with ten years of experience and three trainees with about two years each performing a defined mock cell seeding task inside a biosafety cabinet.
Researchers built a workflow that applies large language models to the free-text comments attached to biodiversity observations on citizen science platforms.
ace