AI Science Tracker

Chinese Academy of Sciences

9 discoveries from 2023 to 2026

Discoveries

Deep learning maps six high-potential deep-Earth scattering zones

Researchers used deep learning with iterative manual optimization to identify 174,929 PKP precursors in more than two million earthquake waveforms recorded from 1990 to 2024. The expanded global map connected previously isolated sampling regions and highlighted six high-potential scattering zones near the core–mantle boundary that the authors report had not previously been documented. These are model-inferred targets for later seismic imaging, not fully characterized structures.

Technology: Deep-learning PKP precursor detector with human iterative optimization

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: Researchers used deep learning with iterative manual optimization to identify 174,929 PKP precursors in more than two million earthquake waveforms recorded from 1990 to 2024.
  • Still unknown: The six features are high-potential scattering zones inferred from probability maps and proposed as targets for future multi-phase inversions; they are not precise outlines or independently characterized structures. The proposed thermochemical explanations are interpretations. Cross-validation with independent seismic phases supports multiple mapped ULVZs, but does not individually validate all six newly highlighted zones.
  • AI role: The deep-learning detector screened over two million earthquake waveforms to identify 174,929 weak PKP precursor signals; researchers iteratively optimized detections manually. This AI-enabled catalog increased global sampling density by an order of magnitude and supported the dual-probability map that identified six previously undocumented high-potential scattering zones.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025JB033195

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.

Deep learning guides higher-activity SpuFz1 nuclease variants

A DeepSequence-based predictor ranked SpuFz1 mutations without experimental activity labels; experimentally confirmed substitutions were combined into SpuFz1-v3. At the B2M locus in HEK293T cells, the article reports 35.52% indel formation versus 3.07% for wild type. Comparative sequence analysis, not this AI model, guided the separate ultrashort RNA design. The reported activity is assay- and locus-specific; off-target and clinical safety are not established.

Technology: DeepSequence-based Fanzor-Fitness Predictor

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: The primary source reports the study-specific measured result summarized here.
  • Still unknown: Independent replication and clinical or industrial efficacy are not established.
  • AI role: A learned sequence fitness predictor ranks candidate mutations; researchers assay single substitutions and combine successful variants.

Evidence sources: https://www.nature.com/articles/s41467-026-74624-6

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 indexed full text or abstract and publication metadata reviewed; scope stated in source_checks..

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.

Org-Mol prioritizes two experimental immersion-coolant candidates

A 3D transformer pretrained on molecular structures and fine-tuned for physical-property prediction screened millions of ester candidates. Researchers synthesized and characterized two selected candidates. One did not meet the preferred thermal-conductivity threshold. This is experimental candidate evaluation, not industrial deployment or environmental-safety validation.

Technology: Org-Mol 3D molecular transformer

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: The primary source reports the study-specific result summarized here.
  • Still unknown: Independent replication and broad deployment are not established by this review.
  • AI role: A pretrained molecular transformer predicts bulk properties to prioritize candidate esters; researchers assess feasibility, synthesize and measure two candidates.

Evidence sources: https://arxiv.org/abs/2501.09896

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 indexed article or abstract and bibliographic metadata reviewed; complete source scope is given in source_checks..

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-guided amidase engineering improves enantioselectivity

Random-forest models using chemistry and geometry descriptors were trained on 240 prior amidase datasets and used to guide substrate and variant selection. The primary article reports that variant I198F/W328F achieved an enantioselectivity E-value of 303.5, compared with 5.7 for wild type, for the reported substrate assay. Most experimental substrate predictions agreed, with identified exceptions. This bounded laboratory result does not establish broad performance across substrates or industrial application.

Technology: Random-forest classification models and descriptor-guided variant design

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: Primary paper reports the stated assay results.
  • Still unknown: Independent replication and exhaustive prior-art review not established.
  • AI role: Random-forest classification models and descriptor-guided variant design

Evidence sources: https://www.nature.com/articles/s41467-024-53048-0

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 publisher full text and bibliographic metadata reviewed for identity, AI role, dates, affiliations and bounded results..

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 identifies sediment-derived granites in the Hadean zircon record, supporting early subduction

A machine-learning classifier trained on trace-element fingerprints identified detrital zircon signatures from sediment-derived S-type granites as old as 4.24 billion years in the Jack Hills record. Across global detrital zircon data, the authors report long-term cyclic changes in S-type granite proportions and infer that exposed, weathered continental crust and subduction operated by about 4.2 Ga. The result bears on early Earth geodynamics, not a direct fossil or biological discovery.

Technology: Trace-element Support Vector Machine (TSVM) classifier

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 machine-learning classifier trained on trace-element fingerprints identified detrital zircon signatures from sediment-derived S-type granites as old as 4.24 billion years in the Jack Hills record.
  • Still unknown: The early-subduction interpretation is inferred from zircon trace elements and remains one interpretation of sparse Hadean geological evidence. The article reports that only the combination of early subaerial weathering and subduction explains the prevalence of the inferred S-type granites; it does not directly observe tectonics at 4.2 Ga.
  • AI role: Classified zircon as S-type or non-S-type from multielement geochemical fingerprints, allowing the authors to identify the oldest sediment-derived granite signal and quantify its temporal distribution across detrital-zircon records.

Evidence sources: https://www.pnas.org/doi/10.1073/pnas.2405160121

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, 25 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.

Deep learning reveals a new seismicity distribution along the Kabaw Fault in central Myanmar

A deep-learning-based catalog pipeline processed 18 months of seismic-array data and located 1,891 shallow earthquakes, more than twice the 849 events in the prior manual catalog. The denser catalog shows that shallow Kabaw Fault seismicity disappears south of about 22.8°N while deeper events occur west of the fault, and it fills part of a previously under-sampled seismicity gap; the authors interpret this contrast as potentially reflecting along-strike variation in deformation associated with oblique Indian-plate subduction.

Technology: DEPALORE pipeline using Earthquake Transformer, transfer-learned PickNet, REAL association, NonLinLoc location and HypoDD relocation

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 deep-learning-based catalog pipeline processed 18 months of seismic-array data and located 1,891 shallow earthquakes, more than twice the 849 events in the prior manual catalog.
  • Still unknown: The source reports catalog-derived patterns and offers a tentative tectonic explanation; the catalog contains 1,891 localized events and location uncertainties of several kilometers. The AI enabled detections and resolution, while geological interpretation remains the authors’ analysis.
  • AI role: Deep-learning detectors and phase-picking models extracted seismic arrivals from continuous array waveforms; association and location stages produced the higher-completeness catalog from which researchers interpreted fault patterns.

Evidence sources: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023GL105159

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, 16 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.

AI-assisted macaque cortex atlas reveals regional cell-type organization

AI-assisted segmentation and Spatial-ID annotation helped map 264 transcriptome-defined cell types across the macaque cortex. Regional cell-type composition tracked cortical hierarchy, and cross-species comparisons identified primate-specific types enriched in layer 4. This is the July 12, 2023 journal version; the atlas is observational, uses male monkeys, and has limited transcript capture per spatial cell.

Technology: AI-assisted cell segmentation; Spatial-ID (deep neural network and graph convolutional network)

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: New macaque tissue measurements and spatial maps were compared across sections and biological replicates; this supports a bounded anatomical result, not causal brain function.
  • Still unknown: Not assessed: independent replication and exhaustive supplementary material. The paper identifies limited transcript sensitivity and absence of female subjects.
  • AI role: Image segmentation identifies cells; Spatial-ID trains a DNN on nucleus profiles and a GCN incorporating spatial neighborhoods to assign spatial cell labels.

Evidence sources: https://www.sciencedirect.com/science/article/pii/S0092867423006797 · https://github.com/TencentAILabHealthcare/Cell_Segmentation_Fluorescence · https://bcas.cas.cn/issue/2023_3/202311/P020231112626745733563.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: 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 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.

AlphaFold-assisted structural clustering discovers new deaminase functions and compact base-editing enzymes

The team used AlphaFold2 structure predictions to cluster the deaminase family and found that most proteins in a DddA-like clade were not double-stranded DNA deaminases. The work characterized smaller single-strand-specific deaminases, developed a compact cytidine deaminase suitable for packaging a base editor into one AAV, and demonstrated robust editing by a clade member in soybean.

Technology: AlphaFold2-assisted structure-based protein-family clustering

Family AlphaFold: 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: The team used AlphaFold2 structure predictions to cluster the deaminase family and found that most proteins in a DddA-like clade were not double-stranded DNA deaminases.
  • Still unknown: This paper is an AI-assisted protein-discovery and engineering study; experimental base editing demonstrates the utility of selected deaminases, while therapeutic application remains prospective. The study was followed by a 2024 correction/erratum; the record references the original research article, DOI 10.1016/j.cell.2023.05.041, not the erratum.
  • AI role: AlphaFold2 predicted protein structures that enabled structure-similarity clustering of a full deaminase family. The clustering guided discovery of functional diversity and selection of enzymes, which the researchers characterized experimentally and used in cell, animal and plant base-editing tests.

Evidence sources: https://doi.org/10.1016/j.cell.2023.05.041

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, 254 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.

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