AI Science Tracker

Harvard Medical School

9 discoveries from 2022 to 2026

Discoveries

Active learning improved whole-cell antibiotic screening and validation

A closed-loop active-learning campaign for Borrelia burgdorferi whole-cell screening increased the experimental hit rate five-fold, from 0.2% in investigator-selected plates to 1.0% in the AL-guided campaign. The same study separately reports prospective in-silico selection of diverse compounds across bacterial species, with 11.0% experimental validation of predicted hits and 53-fold enrichment over investigator-directed screening; those validated compounds showed the intended narrow-spectrum B. burgdorferi activity. These are assay and screening results, not clinical efficacy.

Technology: Active learning using a directed message-passing neural network and evidential Dirichlet head

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: In a closed-loop Borrelia burgdorferi campaign, active-learning-guided selection produced a 1.0% experimental screening hit rate versus 0.2% in investigator-selected plates. A separate prospective in-silico selection across bacterial species achieved 11.0% experimental validation of predicted hits (53-fold enrichment over investigator-directed screening); the paper reports intended narrow-spectrum activity for B. burgdorferi.
  • Still unknown: Preprint evidence is not independent replication or clinical testing. The cross-species hit-validation experiment is distinct from the closed-loop Borrelia hit-rate comparison. Direct full page retrieval was blocked; checked material was the primary indexed preprint and archived primary article text.
  • AI role: A directed message-passing neural network with an evidential Dirichlet head supported active-learning compound selection for whole-cell inhibition testing; selected compounds were then experimentally screened.

Evidence sources: https://www.biorxiv.org/content/10.64898/2026.07.04.736489v1 · https://pmc.ncbi.nlm.nih.gov/articles/PMC13344992/

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 date: 17 Nov 2026. Review scope: Primary bioRxiv indexed text, PubMed metadata/abstract and PMC indexed full-text excerpt; direct bioRxiv and PMC page loads were blocked..

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.

popEVE: diagnosing rare developmental disorders

Applied to about 30,000 patients with severe undiagnosed developmental disorders, it led to a diagnosis in about one-third of cases and identified variants in 123 previously unknown genes, 25 of which were independently confirmed by other labs. This is preliminary clinical use; further checks are needed before adoption, and the page gives no detailed metrics.

Technology: popEVE (variant-effect prediction model)

Publication: Peer-reviewed paper · Test context: prospective or independent check. 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: prospective or independent check
  • Tested: popEVE was tested on documented variants and case studies, and applied to about 30,000 undiagnosed patients, leading to a diagnosis in about one-third; 123 novel gene candidates were identified, and 25 were independently confirmed by other labs.
  • Still unknown: The press article says popEVE will need further verification for safety and accuracy before wide clinical adoption, and gives no accuracy metrics; it is not yet in routine clinical use.
  • AI role: A deep generative model extending EVE with a protein language model, producing a calibrated disease-likelihood score for each variant.

Evidence sources: https://hms.harvard.edu/news/new-artificial-intelligence-model-could-speed-rare-disease-diagnosis · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41588-025-02400-1&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.

Editorial assessment: Impact 4/5 · Validation 4/5 · 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.

TxGNN: repurposing existing drugs for rare diseases

The model covers about 8,000 drugs across 17,080 diseases and is nearly 50% better than competitors at identifying candidates and 35% more accurate on contraindications. It was tested on 1.2 million medical records and on three rare diseases excluded from training. Computational predictions only, with no laboratory or clinical confirmation.

Technology: TxGNN (graph neural network)

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 model was checked against about 8,000 drugs and 17,080 diseases, validated on 1.2 million medical records, and asked to find drugs for three rare diseases excluded from training; it was compared with leading drug-repurposing models.
  • Still unknown: Identified therapies would require further evaluation for dosing and delivery, and there is no laboratory or clinical confirmation.
  • AI role: Proposed existing drugs for thousands of diseases, predicted contraindications and side effects, and gave a rationale for each prediction.

Evidence sources: https://hms.harvard.edu/news/researchers-harness-ai-repurpose-existing-drugs-treatment-rare-diseases

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, 303 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.

CHIEF: detecting cancer and predicting survival from slides

Trained on 15 million patches and 60,000 slides and tested on more than 19,400 slides from 32 datasets and 24 hospitals: it detects cancer with about 94% accuracy across 15 datasets, predicts mutations in 54 genes with over 70% accuracy, and improves survival prediction by 8% across 17 institutions. Retrospective validation; not trained on rare diseases or premalignant tissue.

Technology: CHIEF (pathology foundation 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: Retrospective validation on more than 19,400 slides from 32 datasets at 24 hospitals; cancer detection was about 94% accurate across datasets.
  • Still unknown: Validation beyond these datasets is still needed, rare and non-cancerous conditions and pre-malignant tissue were not tested, and gene prediction accuracy varies by gene and cancer type.
  • AI role: Read digital tumor slides to detect cancer, predict gene mutations and survival, and generate heat maps of tissue features for pathologists.

Evidence sources: https://hms.harvard.edu/news/new-artificial-intelligence-tool-cancer

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, 685 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.

Machine-learning Treg framework identifies SMARCC1 as a previously unrecognized regulator

A framework combining single-cell ATAC-seq, motif analysis, topic modeling, machine learning and genetic variation identified SMARCC1 as a previously unrecognized controller of regulatory T-cell chromatin programs. CRISPR-Cas9 deletion of Smarcc1 in Tregs produced the predicted loss of accessibility in topic-3 regulatory regions, supporting the computationally inferred connection.

Technology: Integrated machine-learning framework using topic modeling and high-density genetic variation over single-cell ATAC-seq 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: A framework combining single-cell ATAC-seq, motif analysis, topic modeling, machine learning and genetic variation identified SMARCC1 as a previously unrecognized controller of regulatory T-cell chromatin programs.
  • Still unknown: CRISPR validation supported the predicted SMARCC1 effect for topic 3; topic 10 was not represented and the authors note possible redundancy with SMARCC2. The result concerns regulatory chromatin programs in Tregs, not clinical or therapeutic outcomes.
  • AI role: The framework organized co-varying open chromatin regions into Treg regulatory modules, inferred transcription-factor effects on those programs, and nominated SMARCC1 as a regulator of topic-3/10 accessibility. The team tested the prediction by CRISPR-Cas9 editing Tregs and profiling chromatin accessibility.

Evidence sources: https://cbdm.hms.harvard.edu/assets/Publications/2024pub/chowdharyPNAS.pdf

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, 32 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.

EVEscape: predicting viral variants before they emerge

Starting from January 2020, the tool predicted SARS-CoV-2 mutations with accuracy similar to experimental antibody-binding assays and ranked their prevalence better; it generalizes to HIV and influenza. Retrospective validation; prospective performance is not reported.

Technology: EVEscape

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: EVEscape was trained only on pre-2020 sequences and compared against high-throughput experimental scans of SARS-CoV-2 variation that later emerged, a retrospective validation; prospective performance is not reported.
  • Still unknown: The abstract states no limitations; the page says testing on Lassa and Nipah is ongoing and vaccine or therapy use against future variants is a future application.
  • AI role: A deep learning model of historical viral sequences, combined with biophysical and structural information, scored the escape potential of mutations.

Evidence sources: https://hms.harvard.edu/news/ai-tool-can-help-forecast-viral-outbreaks · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06617-0&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, 245 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 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 assesses heart allograft rejection from biopsy whole-slide images

A deep learning system analysing gigapixel endomyocardial biopsy slides detected allograft rejection with an AUC of 0.962, distinguished cellular and antibody-mediated rejection (AUCs 0.958 and 0.874) and graded rejection with an AUC of 0.833. In a human reader study it was non-inferior to conventional assessment and reduced interobserver variability and assessment time. Caveat: test cohorts from the US, Turkey and Switzerland are retrospective and the authors call for clinical trials.

Technology: Deep learning on whole-slide images for endomyocardial biopsies

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: Rejection detection (AUC 0.962) was measured on a curated US dataset and independent test cohorts from Turkey and Switzerland, and a human reader study found the system non-inferior to conventional assessment.
  • Still unknown: The authors state that clinical trials are needed to establish the efficacy of AI-assisted biopsy assessment.
  • AI role: Deep learning system analysed gigapixel endomyocardial biopsy whole-slide images to detect, subtype and grade rejection.

Evidence sources: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41591-022-01709-2&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 10% of works of the same type, year and subfield (OpenAlex work metrics, 105 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.

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