Latest material update · 29 October 2026
AI is contributing to drug development, molecular research, mathematical proofs and scientific tools. The evidence ranges from early company reports to peer-reviewed papers and public formal proofs. This tracker follows the result, AI's documented contribution and the work still needed to establish its value.
This is a curated record across scientific fields, not a complete survey of everything published. Our initial selected-source review was completed on 27 October 2026. Original research dates stay visible, including older foundation records.
For detailed mathematics coverage, see Kingy.ai's Mathematics & Science Breakthrough Tracker.
How to read the evidence
Each entry separates stage from evidence. A human trial can have only company-reported results; a peer-reviewed paper can describe a computational prediction. A formal proof checks a precisely encoded mathematical statement. Laboratory findings, trial milestones and regulatory approval answer different questions.
Latest developments
AI-assisted review shortens adverse-event annotation in an oncology study
Gallifant and colleagues evaluated an agentic language-model system that extracts six types of immune-related adverse event from oncology notes for human verification. In a randomized crossover study, 17 clinical research staff reviewed 20 unique notes, producing 316 evaluable timing observations. Median annotation took 428 seconds in the manual workflow and 242 seconds with the AI-assisted interface, a 43.5% reduction.
Limit: All study phases were within one health system, without external validation. The comparison also changed the review interface, so it does not isolate the model’s contribution. The separate May–July 2025 silent deployment performed less well than retrospective tests and did not assess patient outcomes. This does not establish better cancer outcomes or autonomous diagnosis.
Watch next: External validation and studies measuring effects on patient care. Nature Cancer paper, Study code and documentation, Original March preprint.
Sources checked 29 October 2026
Peer-reviewed EVENET study tests reusable particle-collision analysis
Hsu and colleagues describe EVENET, a shared AI model pretrained on 500 million simulated collision events and adapted to several particle-physics analysis tasks. The paper reports gains on selected comparisons and demonstrates transfer to experimental data by recovering the known upsilon meson in CMS Open Data.
Limit: Recovering a known resonance does not establish a new particle discovery. The authors leave substantial simulation–data mismodelling stress tests and a fully uncertainty-aware system for future work. Reported comparisons do not establish performance across every detector, dataset or analysis.
Watch next: Independent reproductions and broader tests of simulation-to-data transfer and systematic uncertainties. Nature Machine Intelligence paper, Original January preprint.
Sources checked 29 October 2026
Anthropic reports progress across six drug-development partnerships
Anthropic reports work with six organizations developing medicines. Examples include Claude-assisted CAR T genetic designs, antibody optimization, allergy-related protein designs and manufacturing-document review. BigHat describes seven experimental rounds optimizing two antibody fragments over about ten weeks.
Limit: These reports do not establish successful medicines or patient benefit. Anthropic and BigHat give different variant counts and binding-improvement figures, so we omit those numbers pending clarification. Document-review gains also need expert review and do not prove faster regulatory approval.
Watch next: Detailed experimental reports, reconciled measurements and progression into human studies. Anthropic research account, BigHat's account.
Sources checked 27 October 2026
Claude-designed proteins bind scorpion toxins
Anthropic reports 17 qualifying binders among 211 experimentally tested designs from a Claude-guided campaign. The result connects AI-generated protein designs with physical binding measurements.
Limit: Binding a toxin does not establish that a protein neutralizes it. Neutralization in a cellular environment remained untested in the account reviewed here. These designs are research candidates, not an antivenom.
Watch next: Neutralization experiments, broader biological testing and independent assessment. Anthropic's technical account.
Sources checked 27 October 2026
AI-guided search improves quantum-circuit resource trade-offs
Researchers used language models to propose quantum-circuit designs, with a separate evaluator checking and scoring the proposals. The paper reports lower Toffoli-gate and qubit counts in seven of eight tasks; the eighth improves modular-inversion trade-offs. Anthropic identifies Claude Mythos 5 in its account.
Limit: These are resource estimates for circuit designs. They do not demonstrate useful quantum advantage on hardware, and a reduction in one resource can require more of another.
Watch next: Independent checks, peer review and implementation evidence. Gu and colleagues' preprint, Anthropic's explanation.
Sources checked 27 October 2026
Recursion reports early REC-4881 human-trial results
Recursion reports a 47% median reduction in total gastrointestinal polyp burden after 12 weeks of REC-4881 in familial adenomatous polyposis. The company used its AI-enabled discovery approach to identify the therapeutic opportunity.
Limit: This is an investigational treatment in a small single-arm study, with a 10 September 2026 data cutoff. The release reports grade 3 treatment-related adverse events in four patients and four adverse-event-related discontinuations; none of those discontinuations was attributed to a grade 3 treatment-related event. The result does not establish approval, long-term benefit or that AI alone caused the outcome.
Watch next: Full study results, durability and safety follow-up. Recursion's release.
Sources checked 27 October 2026
An AI-assisted ultraviolet map fills gaps in the sky
With human direction, Claude Science helped gather, calibrate and combine historical ultraviolet surveys, then predict missing regions using other wavelengths. The current map labels 28% of its far-ultraviolet layer and 27% of its near-ultraviolet layer as predicted.
Limit: Predicted pixels are not observations and cannot establish a new astronomical source. The methods also warn that validation was generated by the same system. This entry records a scientific data tool, not a confirmed new discovery.
Watch next: Independent validation and clear separation of measured and predicted regions. Map and methods, project account.
Sources checked 27 October 2026
A preprint reports algorithmic improvements for 3SUM and shortest paths
Alman and Vassilevska Williams report fixed-polynomial exponent improvements for 3SUM and all-pairs shortest paths in specified integer models. Their manuscript credits Claude with discovering the core algorithm; the authors simplified, strengthened and extended it and take responsibility for the work.
Limit: The preprint remains provisional. Its asymptotic bounds do not establish a practical speedup for everyday workloads. The authors disclose that Anthropic offered compensation and provided Claude access. We inspected the public formalization documentation without replaying the proof checks.
Watch next: Specialist scrutiny, peer review and independent artifact checks. Manuscript and AI disclosure, formal artifact.
Sources checked 27 October 2026
AlphaGenome Atlas expands access to DNA-variant predictions
Google DeepMind reports releasing predictions for approximately nine billion possible single-nucleotide variants. Its account describes collaborators experimentally validating selected rare-disease findings. AI's role is to predict how DNA changes may affect biological activity and help prioritize investigation.
Limit: Selected validation does not establish the accuracy of every prediction. Atlas availability is not diagnostic approval or evidence that a treatment works.
Watch next: Independent evaluations, additional experimental validation and demonstrated uses in research. DeepMind's release and supporting work.
Sources checked 27 October 2026
Rentosertib advances into phase 3
Insilico reports that rentosertib, developed through its AI-driven discovery process, entered phase 3 testing for idiopathic pulmonary fibrosis. Its earlier randomized phase 2a study enrolled 71 patients and was published on 3 June 2025.
Limit: Entering phase 3 is a development milestone, not a successful phase 3 result or marketing approval. The 2025 paper remains older supporting evidence. We have not independently read the trial registry record in this review.
Watch next: Public trial details, recruitment and eventual controlled results. Phase 3 announcement, company updates, Nature Medicine phase 2a paper.
Sources checked 27 October 2026
Claude agents formalize Fermat's Last Theorem
Claude agents used Prove2Me and Claude Code to formalize an existing proof route drawing on Wiles, Taylor–Wiles and prior mathematical work. The project reports kernel, comparator and second-kernel checks.
Limit: The achievement concerns computer-checkable verification of established mathematics. It does not replace the original discovery or make AI the discoverer of Fermat's theorem. This review inspected the project's documentation rather than rerunning its proof build.
Watch next: Independent reproduction and assessment of the encoded statement and dependencies. Project account, proof repository.
Sources checked 27 October 2026
CLaC connects crystal structures with language
CLaC aligns crystal structures and natural-language queries to support materials retrieval and screening. Its graph and text encoders learn from paired descriptions and structures, including GPT-generated descriptions grounded in computed properties.
Limit: Improved retrieval does not establish that a new material was synthesized or a better battery was built. Dataset bias and errors in synthetic descriptions remain relevant. MIT's 18 October coverage is a news date, not the date of a new experimental result.
Watch next: Independent benchmarks and experimental follow-through. Nature Communications paper, MIT's account.
Sources checked 27 October 2026
Foundation record
AlphaFold 3 broadens molecular structure prediction
AlphaFold 3 predicts complexes involving proteins, nucleic acids, small molecules and other molecular components. Its diffusion-based approach extends AI structure modeling across interaction types, evaluated against structural data.
Limit: A predicted structure is not experimental confirmation of biological function, drug efficacy or the full dynamics of a cell. This is a dated foundation record, included to explain subsequent work.
Watch next: Independent evaluations and experimentally tested applications. Nature paper, publication record.
Sources checked 27 October 2026
Scope and updates
We track documented AI contributions across medicine and biology, mathematics and computing, physics and astronomy, chemistry and materials, and climate, Earth science and energy. A field's inclusion in this scope does not imply that every review finds a qualifying result in it.
Primary papers, proof artifacts, experiment records, trial registries and regulator decisions carry more weight than headlines. Company reports remain attributed. We record negative findings, failed programs, corrections and retractions alongside advances, and preserve uncertainty when sources conflict.
Entries change when evidence changes. Original event dates remain fixed; an unchanged source check does not become a new discovery or a fresh publication date. This page reports research progress and does not provide treatment advice.
Material update log
29 October 2026: Added two peer-reviewed publication milestones: AEGIS clinical-note curation and EVENET particle-collision analysis. Preserved preprint and underlying data dates; existing record qualifications remain unchanged. The tracker now contains 13 selected records.
27 October 2026: Created the initial curated tracker with 11 records. Distinguished computational predictions, laboratory findings, clinical milestones and formal verification; retained the original dates of older research.