AI Science Tracker

Stanford University

21 discoveries from 2020 to 2026

Discoveries

Antibody design across observed and forecasted antigenic evolution

The Germinal Multi preprint abstract reports antibodies designed against several antigen variants at once. Experimental screens of tens of designs per campaign found binders to resistance-associated EGFR variants and neutralizers of historical SARS-CoV-2 variants, including variants withheld from design. Redesign of BD55-1205 restored neutralization of a forecast-guided engineered escape variant while retaining parental-variant activity. The escape result concerns an engineered variant, not a demonstrated future circulating strain. Review is limited to the abstract and original-server metadata.

Technology: Germinal Multi (computational antibody design framework); evolutionary forecasting model

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Tens of designs per campaign were screened experimentally; designs bound EGFR variants and neutralized SARS-CoV-2 variants, including variants withheld from the design process.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: Germinal Multi computationally optimized one antibody paratope against multiple antigen variants, and an evolutionary forecasting model guided construction of an escape variant and antibody redesign.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.64898%2F2026.11.10.765734%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.64898/2026.11.10.765734

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

The Virtual Biotech: a multi-agent AI framework for drug development decisions

The Virtual Biotech journal abstract reports a human-guided multi-agent analysis of 55,984 trials. Drugs targeting cell-type-specific genes were associated with a 48% greater likelihood of reaching market and 32% fewer adverse events. The system also proposed a lung-cancer strategy and inferred possible reasons for a terminated ulcerative-colitis trial. These are retrospective associations and hypotheses, not demonstrated treatment benefits. The date represents the 2026-10-22 journal version; the linked preprint, including the same reported association, first appeared on 2026-02-23.

Technology: Virtual Biotech multi-agent AI framework (more than 37,000 agents)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Retrospective trial-outcome/target-feature analysis across 55,984 trials, plus computational therapeutic and trial-failure hypotheses.
  • Still unknown: Review limited to the indexed abstract and publication metadata; full-paper methods and additional limitations were not assessed.
  • AI role: More than 37,000 AI agents annotated trial outcomes and integrated multimodal evidence to propose and analyze therapeutic decisions.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1126%2Fscience.aeg6779%22&resultType=core&format=json · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID%3APPR1222685&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ProVADA steers a protein generator to adapt proteins to non-native biological contexts

ProVADA steered a pretrained protein generator to produce active, drug-inhibitable cytosolic renin and deimmunized two murine antibody fragments while retaining antigen recognition. ESCAPE-seq measured reduced HLA-I peptide binding, and redesigned FMC63 variants retained CAR-T-cell killing. This summary is limited to the abstract of the October 18, 2026 version 2 preprint; the initial preprint dates to July 17, 2025.

Technology: ProVADA (Mixture-Adaptation Directed Annealing sampler steering a pretrained protein generator)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Renin activity/inhibitor response, antibody antigen recognition and HLA-I peptide binding, and FMC63 CAR-T-cell killing, as reported in the version 2 abstract.
  • Still unknown: Only the abstract, PMC version header and original-server corresponding-institution metadata were assessed; full author-affiliation list and full paper were not reviewed. Preserve ID.
  • AI role: A pretrained protein generator, steered by ProVADA's annealing sampler toward externally defined properties while preserving functional constraints.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC13742580/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A%2210.1101%2F2025.07.11.664238%22&resultType=core&format=json · https://api.biorxiv.org/details/biorxiv/10.1101/2025.07.11.664238

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning attributes post-2015 emissions to intensification of Europe’s recent heat extremes

A diffusion model trained on ten CMIP6 climate models estimated event-specific temperature changes conditional on observed meteorology. It found >95% probability that anthropogenic emissions since the 2015 Paris Agreement intensified Europe’s summer temperature extremes in 2021–2025; for the June 2025 heatwave, regional mean temperature is estimated to have been 0.34 °C lower under 2015 cumulative emissions, with >99% probability.

Technology: Generative diffusion model for spatially explicit temperature fields

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: A diffusion model trained on ten CMIP6 climate models estimated event-specific temperature changes conditional on observed meteorology.
  • Still unknown: This event-specific attribution is conditioned on observed large-scale meteorology and trained on ten CMIP6 models; it is not a direct experiment that removed emissions. The paper emphasizes uncertainty from internal variability/model spread and reports robust regional-mean effects, while grid-cell-level changes are more limited. The June 2025 value is a median modeled counterfactual estimate.
  • AI role: The model learned climate-model relationships between meteorological conditions, cumulative CO₂ emissions and daily temperature fields, then generated conditional counterfactual temperature distributions to attribute the incremental post-2015 emissions effect to recent European heat extremes.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026GL125079

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Paper2Agent integrates perturbation datasets to prioritize GPR137 at a psoriasis risk locus

Paper2Agent-generated agents used AlphaGenome to prioritize GPR137 as the top predicted gene affected by psoriasis-associated variant rs887314 in CD4+ T cells, then integrated that prediction with existing MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq data. A cross-screen correlation of gene-expression signatures supported GPR137 as the probable causal gene under stimulated conditions (Spearman rho 0.613 at 8 h and 0.630 at 48 h; FDR < 0.05), but not at rest. The paper presents this as activation-dependent computational evidence.

Technology: Paper2Agent-generated AlphaGenome, MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq agents, coordinated through an AI co-scientist

Family AI Co-Scientist: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Paper2Agent-generated agents used AlphaGenome to prioritize GPR137 as the top predicted gene affected by psoriasis-associated variant rs887314 in CD4+ T cells, then integrated that prediction with existing MPRA-coupled scCRISPRi and CD4+ T-cell Perturb-seq data.
  • Still unknown: The analysis reuses existing published datasets and does not report new wet-lab experiments for this psoriasis case. AlphaGenome supplied a model prediction; cross-dataset signature correlation supports, but does not establish, causal biology. The GPR137 concordance was significant only after stimulation, and the authors describe it as probable/activation-dependent. A human researcher selected which AI-proposed strategy to run. Primary paper source checked; no independent replication assessed.
  • AI role: The AlphaGenome paper agent scored and ranked candidate genes at rs887314. Under human-in-the-loop supervision, the AI co-scientist inspected two other paper agents' datasets, proposed ten validation strategies, and the researcher selected signature-correlation analysis. The resulting cross-screen, cross-modality analysis found significant concordance for GPR137 only under stimulated conditions.

Evidence sources: https://www.nature.com/articles/s41586-026-11044-y

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Claude's Cycles: a Hamiltonian cycle decomposition found by Claude

Knuth's note reports that Claude Opus 4.6 found an explicit decomposition of the arcs of the m^3-vertex digraph into three directed Hamiltonian cycles for odd m, an open problem Knuth posed. Knuth's proof of the construction was reportedly formally verified in Lean by Kim Morrison. Caveat: sources disagree on the even case, which an independent arXiv paper (2603.24708) and a GPT-5.4 Pro proof treat differently.

Technology: Claude Opus 4.6

Family Claude: profile · models page

Publication: Not annotated in this catalog · Test context: Not annotated in this catalog. Catalog categories, not an evidence grade.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: other. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-generated virtual spatial proteomics reveals lung-cancer tumor–immune niches associated with immunotherapy response

HEX, a pathology-image AI model, predicts spatial expression of 40 protein biomarkers from routine H&E slides. In retrospective NSCLC cohorts, combining HEX-derived virtual spatial proteomics with histology improved prognosis prediction and, in a separate 148-patient immune-checkpoint-inhibitor cohort, improved response prediction over H&E-only and virtual-proteomics-only models. The authors interpreted virtual maps as showing co-localized T-helper and cytotoxic T-cell niches among responders and immunosuppressive macrophage/neutrophil aggregates among non-responders.

Technology: HEX (H&E to protein expression), using a pathology foundation-model backbone; MICA (multimodal integration via co-attention)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: In retrospective NSCLC cohorts, combining HEX-derived virtual spatial proteomics with histology improved prognosis prediction and, in a separate 148-patient immune-checkpoint-inhibitor cohort, improved response prediction over H&E-only and virtual-proteomics-only models.
  • Still unknown: Retrospective observational cohorts; predictive associations do not establish causal mechanisms or improved patient outcomes from model-guided treatment. The 148-patient immunotherapy cohort is separate from the prognosis cohorts. The paper itself says extensive clinical validation remains necessary; do not describe the finding as a clinically validated diagnostic or prospective treatment-selection tool.
  • AI role: HEX predicts expression of 40 proteins and produces spatially resolved virtual proteomics from H&E images; MICA fuses the original image with AI-derived maps for outcome prediction. Researchers used predicted marker distributions to interpret therapy-associated tumor–immune spatial patterns. These are retrospective model analyses with cohort-based clinical associations, not prospective treatment evidence.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12823406/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AI-enhanced seismic catalog exposes a ring-fault system at Campi Flegrei

A retrained machine-learning workflow applied to continuous seismic records from 2022–2025 expanded the Campi Flegrei earthquake catalog from about 12,000 to more than 54,000 events. The more complete locations delineated a sharply defined, shallow caldera ring-fault system, illuminated previously unclear northern faults, and identified very shallow hybrid events likely associated with the hydrothermal system. The authors report no direct signature of upward magma migration in the analyzed data.

Technology: Retrained machine-learning seismic detection and phase-picking workflow developed at Stanford

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: A retrained machine-learning workflow applied to continuous seismic records from 2022–2025 expanded the Campi Flegrei earthquake catalog from about 12,000 to more than 54,000 events.
  • Still unknown: The observations cover 21 January 2022–20 March 2025. The inferred hydrothermal source is described as likely, and the mapped faults do not establish imminent eruption or earthquake timing. Full Science article text was access-restricted in this review; the original abstract and reporting institution's article were read.
  • AI role: Detected and located small earthquakes missed by the prior monitoring catalog, providing denser observations that made the ring-fault geometry and shallow seismic structures visible.

Evidence sources: https://doi.org/10.1126/science.adw9038

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 32 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Decoding inner speech with a brain-computer interface

In four people with severe motor and speech impairments, researchers found imagined-speech signals in the motor cortex and decoded them with machine learning. They also tested filters and a trigger phrase to limit unintended decoding. This is a proof of principle with implanted electrodes: free, rapid inner speech cannot be decoded accurately, and attempted speech remains more accurate.

Technology: Brain-computer interface and machine learning

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Imagined-speech signals were recorded from motor cortex in four participants with implanted electrodes and decoded in real time, with filters and a trigger phrase tested to limit unintended decoding.
  • Still unknown: The source states this is a proof of principle, free inner speech cannot yet be decoded accurately, and attempted speech remains more accurate.
  • AI role: Machine learning decoded the imagined-speech neural signals.

Evidence sources: https://news.stanford.edu/stories/2025/08/study-inner-speech-decoding-device-patients-paralysis · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1016/j.cell.2025.06.015&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 87 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep learning maps previously unknown disease signals in ECGs across two health systems

Researchers independently trained PheWASNet on ECG and electronic health-record data from Stanford Health Care and Cedars-Sinai to assess 1,243 disease phenotypes. In held-out patient splits at both centers, the study reported previously unrecognized ECG associations for disorders including respiratory failure, neutropenia and menstrual disorders (AUCs 0.86, 0.83 and 0.84 at Stanford, respectively). Cross-condition model-output analysis found that 35 of 37 strongly detectable non-cardiac phenotypes shared signals with only four other disease outputs, and simple ECG measurement models explained some findings.

Technology: PheWASNet, a multi-task deep-learning model based on a one-dimensional EfficientNet architecture

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: In held-out patient splits at both centers, the study reported previously unrecognized ECG associations for disorders including respiratory failure, neutropenia and menstrual disorders (AUCs 0.86, 0.83 and 0.84 at Stanford, respectively).
  • Still unknown: This retrospective EHR-linked study establishes model-detectable associations, not causal physiology, screening benefit or clinical utility. The disease labels are diagnostic codes associated with ECGs within 30 days; results may reflect comorbidity or care context. The authors caution that generalization to other health systems and populations is not assured. Some detectable signals were explained by simple conventional ECG measurements, and the paper does not claim that deep learning is required for every association.
  • AI role: The researchers trained separate deep-learning models on raw 12-lead ECG waveforms linked to EHR diagnoses at two medical centers, then used held-out patient test sets to identify disease phenotypes whose signals were detectable from ECGs. Cross-phenotype prediction comparisons and simple-model analyses were used to characterize shared and interpretable signals.

Evidence sources: https://academic.oup.com/ehjdh/article/6/4/595/8126967

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 5 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Regional deep-learning phase picking reveals 24 additional earthquakes in a Gorda Ridge swarm

A long-window regional phase picker and phase-association workflow detected 34 earthquakes in a 29 February 2024 swarm near the Gorda Ridge, offshore northern California: all 10 reported by the USGS plus 24 additional events. The detections were possible from stations roughly 300 km from the swarm, beyond the distance range represented by standard local-event pickers. This is one paper-level catalog result, not 24 separate discoveries.

Technology: SKYNET deep-learning P/S phase picker with GENIE graph-neural-network phase association

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: A long-window regional phase picker and phase-association workflow detected 34 earthquakes in a 29 February 2024 swarm near the Gorda Ridge, offshore northern California: all 10 reported by the USGS plus 24 additional events.
  • Still unknown: The application covers one swarm day and a specific offshore region. The paper reports 24 additional algorithmically detected events and estimates local magnitudes, but does not claim every event had independent analyst confirmation. SKYNET is designed to complement local pickers; its near-source performance is weaker than PhaseNet, and the result is not a new tectonic mechanism.
  • AI role: SKYNET picked long-duration regional seismic waveforms and GENIE associated the picks into events; their combined workflow supplied the new event detections in sparsely instrumented offshore seismicity.

Evidence sources: https://seismica.library.mcgill.ca/article/view/1431

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 15 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Mal-ID maps disease-associated immune-receptor patterns across infections and autoimmune conditions

Mal-ID analyzed B-cell and T-cell receptor repertoires from 593 people, including patients with COVID-19, HIV, systemic lupus erythematosus and type-1 diabetes, influenza-vaccination recipients and healthy controls. Its combined models classified held-out immune states and highlighted disease-associated receptor patterns; interpretation identified distinct autoreactivity features in lupus and type-1 diabetes and antigen-associated receptor sequences in COVID-19. The authors also evaluated classification on independent COVID-19 cohorts.

Technology: Mal-ID (MAchine Learning for Immunological Diagnosis), an ensemble using repertoire feature models and ESM-2 protein-language-model embeddings

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The authors also evaluated classification on independent COVID-19 cohorts.
  • Still unknown: This is a patient-sample diagnostic-validation and immune-biology study, not evidence of clinical deployment or benefit. Validation cohorts and disease-specific sample sizes are limited; external COVID-19 classification involved small cohorts and thresholds were tuned in some analyses. Additional disease-specific validation, threshold calibration, sequencing-depth assessment and interpretation alongside standard clinical assessment remain necessary. Its 2024 bioRxiv preprint is the same work, not a separate finding.
  • AI role: Three machine-learning representations of B-cell and T-cell receptor data were combined to classify immune states and prioritize disease-associated receptor sequences, gene segments and isotypes for biological interpretation.

Evidence sources: https://pubmed.ncbi.nlm.nih.gov/39977494/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 132 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Representation learning uncovers the extragalactic X-ray transient XRT 200515 in Chandra archives

A representation-learning pipeline searched archived Chandra event data and identified XRT 200515, a previously unnoticed extragalactic fast X-ray transient in the Large Magellanic Cloud. Its light curve has a sharp sub-10-second hard burst followed by an approximately 800-second oscillating tail with spectral softening. The paper discusses possible origins including a low-energy giant magnetar flare or a Type I burst from a faint low-mass X-ray binary; the source's physical nature is unresolved.

Technology: Unsupervised representation-learning pipeline using sparse autoencoders, dimensionality reduction and density-based clustering

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: A representation-learning pipeline searched archived Chandra event data and identified XRT 200515, a previously unnoticed extragalactic fast X-ray transient in the Large Magellanic Cloud.
  • Still unknown: The transient is a real source-reported observational event, but its astrophysical progenitor remains unresolved. The study also catalogs candidate flares and dips; those catalog candidates are not counted as separately confirmed discoveries.
  • AI role: Converted Chandra photon event files into time-energy representations, learned low-dimensional embeddings and clustered similar events; anomalous transient-dominated clusters surfaced XRT 200515 for subsequent human inspection and astrophysical analysis.

Evidence sources: https://arxiv.org/abs/2412.01150

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: at percentile 86 of works of the same type, year and subfield (OpenAlex work metrics, 20 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Generative agents built from interview transcripts simulate 1,000 people

LLM agents built from participants' full interview transcripts reproduced their General Social Survey answers at about 85% of the rate at which participants reproduced their own answers two weeks later, a normalized score. They also matched personality and economic-game results (normalized correlations of about 0.80 and 0.66) and replicated four of five behavioural experiments. Caveat: arXiv preprint, not peer reviewed, and the results are normalized rather than raw accuracy.

Technology: Generative agent architecture with LLM and full interview transcripts

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Agents built from interview and survey data of 1,052 Americans were scored on held-out General Social Survey items against participants' own two-week test-retest consistency, and on personality, economic-game and experiment replications.
  • Still unknown: Preprint, not peer reviewed; there is no dedicated limitations section, and combining interviews and surveys gave only modest gains over either source.
  • AI role: An LLM was the engine of agents that answer questions on behalf of individuals, grounded in their self-report data.

Evidence sources: https://arxiv.org/abs/2411.10109

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Structure-informed model selects antibody variants active against SARS-CoV-2 escape variants

ESM-IF1 scored substitutions in antibody–antigen complexes, and researchers experimentally screened variants of two clinical antibodies. The journal article reports up to 25-fold neutralization improvement against BQ.1.1 and up to 37-fold affinity improvement against XBB.1.5; these are distinct assays/variants. The preprint abstract reports a different neutralization figure, so this record uses journal values and does not backdate them to preprint v1.

Technology: ESM-IF1 inverse-folding protein model

Family ESM / ESMFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: PMC-hosted primary article reports experimental screening and improved neutralization/affinity results for antibody variants.
  • Still unknown: Exact v1 result content and sequence-level overlap with earlier group studies not completely established; no clinical efficacy.
  • AI role: Structure-informed inverse-folding scores guided substitutions followed by antibody binding and neutralization tests.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11616794/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 16 Nov 2026. Review scope: Primary source checked for identity, date, model role and bounded reported result..

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Language models guide improvements to seven antibodies

Two rounds of laboratory evolution, using at most 20 variants per antibody, improved the affinity of four mature antibodies by up to sevenfold and three immature ones by up to 160-fold. Biochemical and pseudovirus tests; no demonstration of therapeutic benefit in humans.

Technology: Protein language models

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Language-model-guided variants of seven antibodies were made in two lab evolution rounds (at most 20 variants per antibody) and binding was measured by biolayer interferometry; four antibodies were also checked for thermostability, polyspecificity and pseudovirus neutralization.
  • Still unknown: The authors state that affinity gains are smaller than in vivo somatic hypermutation, the method improves existing binding rather than creating new binders, and it may work less well under other selection pressures; no therapeutic benefit in humans is shown.
  • AI role: General protein language models guided affinity maturation by choosing which antibody variants to test.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10869273/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41587-023-01763-2&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 481 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Machine learning reveals a nitrogen-assimilation strategy in Bacteroides

A microbiome metabolomics study used random-forest models to associate bacterial metabolite profiles with taxonomy and found a previously undescribed nitrogen-assimilation pattern in Bacteroides/Bacteroidetes. Follow-up growth experiments tested 60 Bacteroidetes strains with asparagine or glutamine as the sole nitrogen source and supported the predicted phenotype; the study also reports a distinct reduced-sulfur utilization result in B. thetaiotaomicron.

Technology: Random-forest models over metabolomics profiles

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Follow-up growth experiments tested 60 Bacteroidetes strains with asparagine or glutamine as the sole nitrogen source and supported the predicted phenotype; the study also reports a distinct reduced-sulfur utilization result in B. thetaiotaomicron.
  • Still unknown: The study's abstract reports machine-learning discovery of a new metabolism and the figure captions document follow-up culture tests; the article's full text is paywalled in this session, so the exact mechanistic pathway is not asserted. Mouse and culture evidence do not establish a human health effect.
  • AI role: The models identified metabolite features that distinguished taxa and highlighted a previously undescribed Bacteroides metabolic phenotype; researchers then tested bacterial growth under defined nitrogen sources. Comparative-genomics pathways remained candidate explanations rather than proven pathways.

Evidence sources: https://www.nature.com/articles/s41586-021-03707-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 417 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Fusing CT images and EHR data with deep learning improves pulmonary embolism detection

A deep multimodal network combining CT imaging with electronic health record data reached AUROC 0.947 (95% CI 0.946-0.948) on a held-out test set for pulmonary embolism. The abstract reports no sensitivity, specificity, study size or external validation sites, so the evidence is limited to retrospective internal testing.

Technology: Multimodal deep neural network fusing CT and EHR data

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Several fusion architectures were compared on a held-out test set, and the best late-fusion model reached AUROC 0.947 against imaging-only and EMR-only models.
  • Still unknown: The abstract reports no sensitivity, specificity, study size or external validation and states no limitations.
  • AI role: Deep neural networks classified pulmonary embolism from CT angiography volumes combined with EHR data.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41598-020-78888-w&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 263 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

PENet detects pulmonary embolism on volumetric CT angiography

PENet, a 3D deep-learning model for pulmonary embolism on CT angiography, reached AUROC 0.84 on an internal hold-out test and 0.85 on data from one unnamed external institution. The abstract does not report sensitivity, specificity or study size, and the external validation is limited to a single site.

Technology: PENet, 3D deep convolutional network on volumetric CT

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: PENet, a 3D CNN, was tested on a hold-out set from the training institution (AUROC 0.84) and on an external institution (AUROC 0.85), and was compared with other state-of-the-art 3D CNN models.
  • Still unknown: The abstract states no limitations; external validation covers one institution and sensitivity, specificity and study size are not reported.
  • AI role: A 77-layer 3D CNN, pretrained on Kinetics-600, detected pulmonary embolism on CT angiography and is proposed as a triage tool.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41746-020-0266-y&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 203 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

EchoNet-Dynamic: video-based AI for left ventricular ejection fraction

A deep learning model trained on echocardiogram videos segmented the left ventricle with a Dice coefficient of 0.92 and estimated ejection fraction with a mean absolute error of 4.1% (6.0% on an external dataset from another health system). Prospective evaluation with repeated human measurements found its variance comparable to or lower than that of human experts. Caveat: training data come from one institution's echocardiograms, with external testing on a single additional system.

Technology: EchoNet-Dynamic (video deep learning model)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Segmentation and ejection fraction were tested on held-out internal data (Dice 0.92, MAE 4.1%) and an external dataset from another health system (MAE 6.0%), and repeated human measurements were compared with model variance.
  • Still unknown: Not reported; the page states no limitations section.
  • AI role: Deep learning model trained on echocardiogram videos to segment the left ventricle and estimate ejection fraction.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-020-2145-8&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1292 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Closed-loop optimization of fast-charging battery protocols

A model estimating cycle life from the first 100 cycles, paired with Bayesian optimization, assessed 224 fast-charging protocols in 16 days; the authors estimate exhaustive search without early prediction would take over 500 days. It optimizes protocols, not new materials; only the abstract was reviewed.

Technology: Bayesian optimization with early prediction

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: 224 fast-charging protocols tested experimentally by measuring battery cycle life; high-cycle-life protocols identified in 16 days, versus an estimated 500+ days for exhaustive search without early prediction.
  • Still unknown: The abstract lists no explicit limitations; generalization to other problems is suggested but not discussed; only the abstract was read.
  • AI role: An early-prediction model estimated final cycle life from the first cycles, and Bayesian optimization chose which protocols to test next.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-020-1994-5&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1252 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Open the interactive profile for the timeline, or browse all organizations and model families.