AI Science Tracker

Broad Institute of MIT and Harvard

4 discoveries from 2022 to 2024

Discoveries

HiDDEN finds malignant-like cell signals in early multiple-myeloma precursor samples

HiDDEN refines sample-level case labels at single-cell resolution. In human bone-marrow single-cell RNA-seq data, its labels tracked manual neoplastic-cell annotations in mixed precursor samples and inferred malignant-like cells in three low-purity MGUS samples that the original manual analysis had labeled as containing only healthy cells. The inferred cells were enriched for previously described malignancy-associated gene signatures. In a separate mouse demyelination dataset, HiDDEN identified an early endothelial subpopulation with expression changes associated with angiogenesis, blood–brain-barrier breakdown and extracellular-matrix responses.

Technology: HiDDEN, a machine-learning framework for refining disease/condition labels at single-cell resolution

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: HiDDEN refines sample-level case labels at single-cell resolution.
  • Still unknown: The MGUS finding is a transcriptomic inference in previously collected samples, computationally supported by known gene signatures and sample-purity estimates; the paper does not report independent pathology confirmation, clinical follow-up, or validation as a diagnostic test. The endothelial subpopulation was found in a mouse lesion model; its association with blood–brain-barrier dysfunction does not establish causality.
  • AI role: The model inferred which individual cells in case samples were affected rather than assigning the sample-level case label to all cells. This enabled the paper’s reanalysis of human myeloma precursor single-cell data and mouse demyelination single-nucleus data, yielding the specified malignancy and endothelial-population findings.

Evidence sources: https://www.nature.com/articles/s41467-024-53666-8.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, 24 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.

Machine-guided synthetic regulatory DNA drives cell-type-specific expression across cell lines and vertebrates

The CODA platform used the Malinois deep-learning model and sequence-search algorithms to design synthetic cis-regulatory elements (CREs). In massively parallel reporter assays, 94.1% of the synthetic sequences met the authors’ cell-type-specificity criterion, versus 73.6% of model-selected natural sequences and 40.6% of chromatin-mark-selected natural sequences; synthetic sequences also had higher measured target/off-target separation across K562, HepG2 and SK-N-SH cells. In transgenic zebrafish, two of three liver-directed and two of three neuron-directed CREs showed the intended activity. One neuron-directed sequence, synN1, retained cortical and neuronal expression in mice, including postnatal neocortical layer 6 and subplate neurons.

Technology: CODA (Computational Optimization of DNA Activity) with Malinois deep neural network and AdaLead, simulated annealing, and Fast SeqProp sequence search/design algorithms

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: In massively parallel reporter assays, 94.1% of the synthetic sequences met the authors’ cell-type-specificity criterion, versus 73.6% of model-selected natural sequences and 40.6% of chromatin-mark-selected natural sequences; synthetic sequences also had higher measured target/off-target separation across K562, HepG2 and SK-N-SH cells.
  • Still unknown: The in-vitro work used three transformed human cell lines and reporter assays. In-vivo testing covered a small selected set: two of three liver-directed and two of three neuron-directed elements were active in zebrafish; mouse validation focused on neuronal synN1, with n=3 synN1 mice for postnatal cell-type analysis and n=6 for the 5-week cortical reporter assay. These are transgenic reporter results, not therapeutic delivery or evidence of clinical utility. Earlier synthetic enhancer studies existed; this paper’s counted result is its CODA/Malinois comparison against natural sequences plus selected vertebrate validation, not the general idea of AI-designed enhancers.
  • AI role: Malinois predicted MPRA activity of candidate DNA sequences in K562, HepG2 and SK-N-SH cells; CODA iteratively optimized sequences for specified cell-type specificity. Researchers synthesized and experimentally measured the designs by MPRA, then tested selected CREs with reporter constructs in zebrafish and mice. The paper attributes the novel sequences and their measured expression specificity to this model-guided design loop.

Evidence sources: https://www.nature.com/articles/s41586-024-08070-z

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, 180 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.

Machine learning generates active daunomycin-binding synthetic nucleic-acid polymers outside selected sequence families

The study combined in-vitro selection with a conditional variational autoencoder trained on selection data to generate high-affinity highly functionalized nucleic-acid polymer (HFNAP) sequences. Experimentally generated CVAE sequences bound daunomycin with reported dissociation constants of 9–26 nM despite having no apparent primary-sequence similarity to experimentally selected polymers; competition selections found selected generated variants performed as well as or better than sequences obtained after multiple rounds of high-stringency selection.

Technology: Conditional variational autoencoder (CVAE) trained on in-vitro selection-derived HFNAP sequence/fitness data

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: Experimentally generated CVAE sequences bound daunomycin with reported dissociation constants of 9–26 nM despite having no apparent primary-sequence similarity to experimentally selected polymers; competition selections found selected generated variants performed as well as or better than sequences obtained after multiple rounds of high-stringency selection.
  • Still unknown: The finding is target binding by synthetic HFNAPs, not therapeutic efficacy or cellular activity. Binding results are specific to daunomycin and reported assays; the authors describe sequences as having no apparent similarity rather than proving wholly unrelated functional mechanisms. Some competition performance was assessed by in-vitro selection rather than a clinical or organism-level test.
  • AI role: The authors used the CVAE to condition generation of novel HFNAP sequences on predicted daunomycin-binding fitness, then synthesized/tested generated sequences in binding and competition-selection assays. The machine-learning model generated the active sequences; wet-lab selection and competition assays established their binding activity.

Evidence sources: https://www.nature.com/articles/s41467-022-31955-4

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 79 of works of the same type, year and subfield (OpenAlex work metrics, 23 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.

Learned activity models improve viral diagnostic assay designs

Learned diagnostic-activity scores guided assay design across viral variation. Experiments on selected synthetic targets supported improved detection compared with the studied conservation-based designs. The larger cross-species design catalog is computational; it is not thousands of validated diagnostics. Patient-specimen sensitivity, clinical specificity and field performance were not established.

Technology: ADAPT learned sequence-activity classifier and regressor

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: Selected assay designs were tested on synthetic viral targets.
  • Still unknown: Clinical and environmental sample performance require further validation.
  • AI role: ADAPT learned sequence-activity classifier and regressor supported the stated study; experimental and interpretive work remained with researchers.

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC9287178/

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: 16 Nov 2026. Review scope: Primary article metadata, affiliations, selected design Methods, experimental Results and limitations; supplements not reviewed..

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.