A new molecule. A newly identified star. A mathematical argument that a computer can check. A better way to decide which experiment deserves a scientist’s time.
These are some of the places where AI progress becomes scientific progress—and September offered plenty to explore.
Across biology, medicine, climate, chemistry, astronomy, and mathematics, researchers published findings and released tools that show how AI is becoming part of the research process. Sometimes it helps uncover something unfamiliar. Sometimes it turns a mountain of data into a manageable list of hypotheses. Sometimes its most useful contribution is showing where an existing method falls short.
This monthly guide gathers 56 research developments, tools, funding announcements, and partnerships released or announced in September 2026. The research itself may have happened earlier. Some papers previously appeared as preprints; others earned their place here through a September journal publication, institutional announcement, or public release.
We’ve kept the evidence visible throughout. Laboratory results, computational predictions, clinical validation, company reports, and research funding each tell a different kind of story.
Know something we missed? Submit a paper, research result, tool, or funding announcement through our submission form. Please include the original source, its publication or announcement date, and a brief explanation of AI’s contribution.
Scientific findings and predictions
The first group covers work that reports scientific observations, experiments, mathematical results, hypotheses, or computational predictions. Where a result still needs confirmation, that distinction matters.
Biology, medicine, and ecology
1. A different way for RNA to recognize DNA
September 17 · Two peer-reviewed Science papers; related preprints appeared April 27
Two studies described VIPR systems, whose guide RNA uses separated pieces of sequence to recognize DNA. AI-assisted genomic searching contributed to the investigation, although the institutional account did not identify a specific model.
Laboratory experiments demonstrated phage defense and gene silencing, while structural researchers mapped 21 molecular structures. The work gives genome-engineering teams another natural system to investigate for programmable gene control. Its biotechnology potential remains at the proof-of-concept stage. Read the studies on noncontiguous DNA recognition and geometric triplex formation.
2. StriMap searches for microbial triggers of immune responses
September 25 · Peer-reviewed Nature Communications article; preprint appeared March 31
StriMap combines protein-sequence features, language-model representations, predicted structures, and graph learning to model interactions between immune receptors and potential targets.
After screening 13 million bacterial peptides, researchers confirmed that microbial lookalikes could activate a receptor on immune cells associated with ankylosing spondylitis. One peptide was more common in people with inflammatory bowel disease, suggesting a possible shared trigger without establishing causation. The method could help autoimmunity and cancer-immunotherapy teams prioritize antigens for experiments. Read the StriMap study.
3. AI identifies Alzheimer’s-associated brain-cell patterns
September 23 · Peer-reviewed Nature Medicine article; preprint appeared November 2, 2024
Researchers used PASCode, a framework combining ensemble methods and graph neural networks, to analyze more than six million brain-cell nuclei from over 580 donors.
The analysis highlighted Alzheimer’s-associated microglia and astrocytes, including patterns linked to cognitive resilience. Those findings could help researchers investigate cell-specific pathways involved in disease and resilience. Because the work draws on post-mortem associations, it does not establish causal treatment targets or a patient-level diagnostic test. Read the PASCode paper.
4. Claude flags an unfamiliar enzyme system
September 23 · Anthropic research announcement and company preprint
Anthropic reported that approximately 950 Claude agents searched DNA-sequence data for 21 hours, identifying an unusual repeat pattern next to a reverse-transcriptase gene.
Follow-up work in Anthropic’s biology laboratory characterized an array-associated reverse transcriptase, or ART, system in bacteriophages. Experiments found that the repeat array is transcribed into short RNAs, but the system’s biological function remains unknown.
This is company-run research rather than an external grant. It offers an example of agent-assisted genome mining followed by laboratory investigation; medical and gene-editing applications remain speculative. Read Anthropic’s report.
5. CT scans help predict pancreatic cancer’s early spread
September 8 · Peer-reviewed Nature Communications article; related conference abstract appeared in 2025
A Mamba-based deep-learning model combined CT information from pancreatic tumors and the liver to predict early liver metastases in pancreatic ductal adenocarcinoma.
In a retrospective study of 1,063 patients, it identified people at higher risk of early spread. A matched, non-randomized analysis also linked that group with apparent benefit from treatment before surgery. The approach could support future risk stratification, but prospective studies are needed before it guides treatment decisions. Read the study.
6. Neural networks help explain why uncertainty disrupts decisions
September 10 · Peer-reviewed Nature Neuroscience article
Researchers compared recurrent neural-network models with human behavior, monkey behavior, brain recordings, and causal experiments to investigate the cost of task uncertainty.
The evidence supported an explanation in which uncertainty strengthens irrelevant visual features, causing them to interfere with information needed for a decision. This gives neuroscientists an experimentally supported account of how uncertainty can impair perception and flexible decision-making. AI served as a way to develop and test an explanation of biological behavior. Read the study.
7. Hakken proposes relationships for scientists to test
September 3 · arXiv preprint with reported laboratory follow-up
Hakken combines time-based knowledge graphs with language-model information to predict relationships missing from the scientific literature.
Researchers scored more than 1.5 million aging-related hypotheses. Experts selected three for cell experiments, and two—TP53–BAMBI and RAF1–TNF—received supporting evidence. The third did not.
For aging and drug-discovery researchers, the interesting contribution is helping choose experiments from a large hypothesis space. The results do not establish an anti-aging treatment. Read the Hakken preprint.
8. Aging clocks are tested inside a trial of an AI-designed drug
September 7 · Peer-reviewed Nature Biotechnology article
Researchers applied six protein-based aging clocks to stored samples from a 12-week trial of rentosertib, an AI-designed drug being studied for idiopathic pulmonary fibrosis.
Among 42 participants with usable samples, the clocks generally estimated lower biological age in treatment arms, with the most consistent signal around week four. However, the clocks could not separate effects on aging from effects of treating the disease.
The study explores how aging biomarkers might be incorporated into clinical trials. It does not demonstrate slower aging or longer life. Read the paper.
9. A genomic classifier maps possible microbial symbionts
September 2 · Nature Biotechnology research briefing; underlying paper appeared August 31
The symclatron classifier uses microbial-genome features to identify organisms likely to depend on a host.
A scan of 107,067 genomes flagged 14,070 potentially host-dependent bacteria or archaea. Predictions were less reliable for microbes that were evolutionarily distant from the organisms represented in the model’s evidence.
The resource could help microbiologists choose uncultured organisms and possible host relationships for follow-up. Its September milestone was the research briefing, following the underlying August research article.
10. Personalized brain-cell analysis reveals Alzheimer’s diversity
September 23 · Peer-reviewed Nature Communications article; preprint appeared November 2, 2024
A knowledge-guided graph neural network helped researchers construct donor-specific networks of cell interactions and gene regulation.
Analyzing more than 1,900 brains, the team identified person-specific molecular patterns and distinct Alzheimer’s trajectories. The work could help neuroscientists investigate why the disease develops differently across individuals.
These maps describe associations and molecular diversity. They do not yet establish causes or demonstrate clinical usefulness. Read the personalized single-cell study.
11. AI proposes overlooked mammal–virus relationships
September 17 · Peer-reviewed Nature Communications article
Dynamic Positive–Unlabeled learning combines a model of observation bias with a classifier trained on known mammal–virus relationships.
The model proposed roughly 8.6 times as many mammal–virus-family links as were recorded. It also ranked some confirmed relationships highly when researchers withheld them during training.
Wildlife-disease teams could use the predictions to direct surveillance toward poorly sampled combinations. The proposed links remain hypotheses until additional evidence confirms them. Read the study.
Earth, climate, agriculture, and ecology
12. Tessera maps smallholder crops in Senegal
September journal publication and September 29 Cambridge announcement; preprint appeared January 23
Researchers evaluated Tessera, a satellite-image foundation model, for identifying crops in Senegal’s groundnut basin.
Across three test years, the model classified crop types correctly about 84% of the time. In one cross-year test, it outperformed the next-best method by a reported 28%; label quality affected performance.
The work could help food-security and climate-planning teams obtain crop information where field surveys are expensive or infrequent. See Cambridge’s September announcement and the research paper.
13. Explainable AI investigates extreme heat
September 7 · Peer-reviewed Weather and Climate Dynamics article
Researchers combined an MLP and a ConvNeXt image model, then used SHAP explanations to investigate summer temperature extremes at six locations across Europe and North Africa.
Large-scale atmospheric patterns dominated the model’s explanations, while soil drying contributed more strongly at some northern locations. The results could help researchers decide which atmospheric and land conditions deserve attention in regional heat studies.
These are model attributions rather than independent proof of causation. Read the extreme-heat study.
14. Climate models face a warmer-world test
September 3 · Peer-reviewed Journal of Advances in Modeling Earth Systems article
Researchers tested ACE2-ERA5, NeuralGCM, and cBottle under uniform sea-surface warming, comparing their responses with NOAA’s physics-based GFDL AM4 model.
The learning-based models captured some important rainfall changes, but differed in radiation responses and warming over land. The study gives climate researchers a practical test of how models behave beyond conditions represented in their training data.
That is useful progress in its own right: understanding a model’s limits helps establish where it can be trusted. Read the comparison.
15. Machine learning estimates hydropower-reservoir emissions
September 3 · Peer-reviewed Global Biogeochemical Cycles article
Researchers used Random Forest, XGBoost, and LightGBM models trained on field studies and environmental predictors to estimate emissions from reservoirs lacking direct measurements.
The framework estimated approximately 304 million tonnes of CO₂-equivalent emissions annually, with methane contributing about 55%. Tropical regions and African sites emerged as potential hotspots.
The findings could guide additional monitoring and improve climate inventories. The total remains a model-based estimate that depends on the study’s data and scaling assumptions. Read the synthesis.
16. ChatGPT assists a shellfish-toxin forecasting workflow
September 3 · Research account in Limnology and Oceanography Letters
Researchers used ChatGPT-assisted coding and data processing to develop LSTM models forecasting toxin levels in Portuguese shellfish.
The models achieved roughly 87–88% accuracy at a lead time of about one week, with performance declining further into the future. The workflow could help aquatic scientists and shellfish-safety teams develop early-warning tools.
Its contribution is model development and forecasting assistance. It does not identify a new mechanism behind harmful algal blooms. Read the research account.
Chemistry, materials, and physical science
17. AI nominates possible altermagnetic materials
September 14 · Journal article; related software and data were public in June
An XGBoost screening model, followed by density-functional calculations, identified 15 metal-organic frameworks as possible altermagnets. Additional models and SHAP analysis helped interpret structural features associated with predicted spin behavior.
For quantum-materials researchers, the result supplies candidates and design clues for future investigation.
The candidates were not synthesized or experimentally measured. This is computational materials discovery, with physical confirmation still ahead. Read the study.
18. Robots and human-guided AI explore material phases
September 10 · Peer-reviewed PRX Intelligence article; preprint appeared January 13
SARA-H combines active learning, human guidance, robotic thin-film processing, and automated phase identification.
In laboratory experiments, the workflow mapped processing conditions that stabilized two bismuth–titanium oxide phases. It also found evidence that substituting bismuth could slow a phase transformation in titanium oxide.
Materials teams could use this approach to explore synthesis conditions more efficiently. The demonstrated advance is a laboratory research workflow, with industrial deployment still untested. Read the experimental study.
19. AI-selected polymers inhibit bacteria
September 21 · Journal publication and Stanford announcement; preprint appeared November 7, 2025
Researchers transferred knowledge from antimicrobial-peptide data to polymers, then used active learning to choose candidates for synthesis.
Ten selected polymers inhibited E. coli in laboratory tests. One was especially effective against biofilms and disrupted bacterial membranes.
The results offer preclinical leads and a useful method for working with limited molecular data. They do not demonstrate a human treatment or a reduction in antibiotic resistance. Read Stanford’s account and the journal paper.
20. Generative AI proposes crystals that conduct little heat
September 10 online publication; September 23 issue date
Researchers combined crystal graph neural networks, cross-property transfer learning, and CrystaLLM to search for low-thermal-conductivity materials.
The generative model proposed 10,000 crystal structures. Stability checks narrowed the search to 17 predicted candidates.
Such a pipeline could help materials scientists prioritize structures for thermoelectric research and synthesis. None of the proposed materials were made or measured in the study, so their practical performance remains unproven. Read the paper.
21. Active learning helps design polarized white-light materials
September 21 · Peer-reviewed Nature Communications article
A dual-loop active-learning system on the Al-Scientist robotic platform balanced two objectives: white-light color quality and circularly polarized emission.
AI-guided experiments produced materials emitting polarized white light across much of the visible spectrum. Researchers also used them in prototype organic LEDs.
The work gives optical-materials teams a laboratory-tested approach for optimizing competing properties. The prototypes have not been demonstrated as manufacturing-scale products. Read the study.
22. A chemistry robot finds an unfamiliar reaction route
September 15 institutional announcement; peer-reviewed paper appeared July 2
The Robowski chemistry robot explored 960 reaction conditions. Researchers used spectroscopy, chromatography, NMR, and mechanism-aware analysis to investigate the products and reaction pathways.
The experiments uncovered a previously unknown route to complex ring-shaped molecules. Some products also sorted and assembled in response to metals.
This illustrates how systematic automation can reveal chemistry within a large experimental search space. Practical applications remain to be demonstrated. Read the September announcement and the research paper.
23. Machine learning screens salts for heat storage
September 3 University of Amsterdam announcement; paper appeared June 5
Composition-based machine-learning methods, including Random Forest/Magpie and ReacRoost, screened 29 salt hydrates for thermochemical heat storage.
Laboratory measurements on three promising hydrates closely matched predicted heat-storage properties. The results suggest that models using chemical composition can help narrow the materials worth testing.
For energy researchers, that could reduce the experimental search burden. The study did not demonstrate a deployed heat-storage system. Read the September announcement and the paper.
Astronomy and space science
24. AI searches Gaia data for rare hot subdwarfs
September 3 research announcement; preprint appeared January 29 and journal paper April 1
Researchers applied a convolutional neural network to Gaia XP spectra to classify rare hot subdwarf stars.
The analysis connected variability with binary systems and reported binary fractions above 60% among the active subset. The enlarged catalog could help astronomers investigate how companion stars shape hot-subdwarf evolution.
Those relationships are inferred from survey patterns. The September milestone was a research announcement about the earlier paper.
25. Follow-up observations confirm 13 Wolf–Rayet stars
September 23 preprint; accepted manuscript published online September 26
Machine-learning selection from Gaia spectra helped prioritize candidate Wolf–Rayet stars, which astronomers then examined with NASA’s SPHEREx infrared observations.
Of 30 candidates, follow-up spectra confirmed 13 additional Wolf–Rayet stars in the Milky Way. The study also identified likely planetary-nebula central stars and helium-emission sources.
The result expands the census of rare massive stars. AI helped choose targets; spectroscopy supplied the confirmation. Read the paper.
26. A deep-learning search finds a possible nearby Y dwarf
September journal issue; earlier preprint was already public
SMDET searches time-resolved WISE images at the pixel level. Researchers checked its findings against archived Spitzer and other infrared data.
The search identified a red, fast-moving object missed in earlier searches. Its infrared colors suggest a potentially nearby, very cool Y dwarf, but no direct spectrum confirms that classification.
For astronomers, it is a promising example of recovering faint objects from crowded archival images. The object remains a candidate. Read the research record.
Mathematics and formal reasoning
27. OpenAI releases a Navier–Stokes singularity construction
September 8 · Company manuscript and Lean formalization
OpenAI reported that an internal model and agent system produced a construction in which a viscous fluid with smooth external forcing develops unbounded speed in finite time. The fluid begins at rest, and its total kinetic energy remains finite.
GPT-6 Astra produced the Lean formalization. This was in-house company research, with no independent peer-review or Clay acceptance established by the announcement.
Fluid mathematicians and proof-system researchers can examine the manuscript and announcement and released proof artifacts. The Neuron’s detailed coverage explores the claim and attribution questions.
28. A separate Euler manuscript addresses unforced fluid motion
September 8 · Company manuscript and Lean certificate
OpenAI also released a separate result concerning the three-dimensional Euler equations, which describe an idealized fluid without viscosity.
The company reported that nearly 100 agents worked for about 50 hours to produce a construction where smooth initial conditions develop a singularity without an external force. The manuscript came with a Lean certificate.
This is another in-house mathematical result for independent examination. It concerns idealized equations rather than the molecular behavior of real fluids. Read the Euler manuscript.
29. AI assists a proof about making irreversible choices
September 17 · Preprint; concurrent paper appeared September 16
Researchers released an AI-assisted proof of the strong secretary conjecture for linear matroids. The problem concerns selecting a valuable, compatible set of options when choices arrive sequentially and rejected opportunities cannot be recovered.
The authors described obtaining the proof through a conversation with ChatGPT-6 Astra. They also acknowledged that another team posted essentially the same result, using an essentially identical approach, the previous day.
Algorithm researchers can investigate the guarantee while recognizing the concurrent discovery. Read the AI-assisted manuscript and the earlier concurrent paper.
Research tools and methods
Scientific progress also depends on tools that make the next investigation easier. These entries primarily report models, reusable methods, benchmarks, or workflows.
Biology, medicine, and ecology
30. AlphaGenome Atlas makes billions of DNA predictions searchable
September 8 announcement; accompanying preprint appeared September 16
Google DeepMind released AlphaGenome Atlas, exposing precomputed predictions for roughly nine billion possible single-letter changes in the human genome.
The resource includes predicted regulatory effects and a variant-impact score. Human-genetics and rare-disease researchers can use it to prioritize variants before conducting functional experiments.
These are model predictions rather than experimental measurements. The advance is providing a large, searchable starting point for investigating genetic variation. Read the announcement.
31. An immune world model proposes therapeutic hypotheses
September 13 · arXiv preprint
Researchers described Agent Genesis and an intervention-conditioned Immune World Model for forecasting immune responses across cells, tissues, and patients.
The authors reported that the frozen model generalized to unseen immune perturbations. It also nominated IL-36γ plus SIRPα inhibition as a combination worth investigating.
The work could help tumor-immunology and immunotherapy teams prioritize experiments. The proposed combination remains a computational hypothesis requiring prospective laboratory and clinical testing. Read the preprint.
32. ProteinTalks models drug effects through changing protein measurements
September 9 · Peer-reviewed Nature article; preprint appeared June 26, 2025
ProteinTalks was trained on millions of time-resolved protein-measurement records to model how biological systems respond to interventions.
Researchers used it to predict drug effects and combinations, then checked predictions against patient-derived organoids and biopsy data. The approach could help drug-discovery teams rank combinations for additional testing.
It remains an in-silico screening aid, with clinical treatment-selection value unproven. Read the virtual-cell study.
33. Botanic-1 brings long-context models to plant genomes
September 4 · bioRxiv preprint
Botanic1 is a family of plant-genome language models with sequence windows reaching 128 kilobases, accompanied by an agent-based model-building workflow.
The authors reported stronger plant-genome prediction than the comparison models, including interpretable signals around coding boundaries and splice sites.
Plant-genomics researchers could use the models to prioritize sequence features and develop crop-research hypotheses. The reported evidence comes from benchmarks; the work did not include wet-lab validation. Read the preprint.
34. GLM-Prior helps build gene-regulation networks
September 14 · Peer-reviewed Nature Communications article; related preprint appeared in July
GLM-Prior uses a fine-tuned genomic language model to predict links between regulatory proteins and genes from DNA, then combines those predictions with single-cell gene-activity data.
Across six cell-line settings, its predicted links agreed with reference networks better than chance. It supplied the strongest starting point in four of five mammalian lines.
The method could help researchers build initial regulatory networks when matched experiments are unavailable. Those predicted networks still need biological validation. Read the study.
35. EAGLE looks for esophageal cancer on noncontrast CT scans
September 22 · Peer-reviewed Nature Medicine article
EAGLE combines localization, classification, and segmentation to detect esophageal cancer on chest CT scans taken without contrast.
Validation involved 80,612 people across 12 centers in three countries. In eight-center external testing, the study reported approximately 90% sensitivity for cancer and 98.5% specificity; performance differed by setting and lesion type.
The opportunity is extracting additional screening value from existing scans. Clinical oversight, follow-up pathways, and implementation studies remain important to establishing its usefulness. Read the study.
36. NucleicBERT learns patterns in RNA sequences
September 3 · Peer-reviewed Nature Machine Intelligence article; preprint appeared September 2, 2025
NucleicBERT is a transformer pretrained on approximately 30 million noncoding-RNA sequences.
It matched or exceeded comparison methods on several structure and function tests without relying on alignments of related sequences. Its learned representations also reflected known structural constraints.
RNA biologists and therapeutic-design teams gain a sequence-based tool for prioritizing hypotheses. Computational performance does not replace experimental checks of structure or function. Read the paper.
37. GPN-Star helps rank potentially important DNA variants
September 9 · Peer-reviewed Nature article; preprint appeared September 21, 2025
GPN-Star, a 200-million-parameter genomic language model, uses whole-genome alignments and species trees to learn patterns of functional constraint.
In computational tests, it improved ranking of disease-linked and noncoding DNA variants and strengthened rare-variant analyses.
Human geneticists could use those scores to select variants for functional testing. The results are predictive improvements rather than a new laboratory assay or a confirmed explanation for every flagged variant. Read the study.
38. IRIS reconstructs the signals cells received
September 8 · Peer-reviewed Nature Methods article; preprint appeared March 17, 2025
Researchers trained IRIS, a neural network, on human stem-cell signaling data and transferred it to mouse-embryo datasets.
The model used gene-activity patterns to infer cells’ signaling histories. Researchers also experimentally checked one predicted signal involved in lung-cell development.
Developmental biologists and stem-cell engineers could use the approach to narrow the signal combinations they test when generating cell types. Read the IRIS study.
39. Paper2Agent makes published methods executable
September 16 · Peer-reviewed Nature article; preprint appeared September 8, 2025
Paper2Agent converts papers, code, and data into tools that AI agents can reuse.
Of 100 computational biology papers tested, 74 were successfully converted, producing 599 proposed tools. Of those, 593 passed automated validation. Researchers also combined published evidence to nominate a psoriasis-associated gene for further study.
The method could make scientific workflows easier to inspect and apply to new data. Missing code, data, and working software environments remained obstacles. Read the Nature paper and The Neuron’s deeper coverage.
40. SCIGMA combines molecular measurements while tracking uncertainty
September 3 · Peer-reviewed Nature Genetics article; preprint appeared April 19
SCIGMA is a graph neural network designed to integrate spatial measurements from multiple molecular layers and platforms.
Across 19 datasets, it improved grouping of tissue regions and indicated where alignments were uncertain. Researchers could use it to combine RNA, protein, imaging, and chromatin information while retaining a view of where the model is less confident.
That uncertainty information is useful when integrated maps will guide subsequent biological investigation. Read the study.
41. DePass combines noisy measurements of cells and tissues
September 24 · Peer-reviewed Nature Cell Biology technical report
DePass uses graph learning to iteratively reduce noise and integrate paired molecular measurements.
Tests across six types of paired data showed improved integration. Applications mapped immune-cell neighborhoods and differences between tumors in colorectal tissue.
The framework could help spatial-biology and oncology teams reconstruct tissue organization from multiple measurements taken together. Its contribution is improved analysis of complex data rather than a demonstrated new treatment. Read the technical report.
42. A “virtual biotech” organizes drug-development evidence
September 17 · Peer-reviewed Science article; preprint appeared February 23
The Virtual Biotech brings together specialized AI agents focused on targets, safety, treatment types, and clinical development.
The system analyzed nearly 56,000 historical clinical trials and identified associations between cell-specific drug targets and trial outcomes. Cancer and ulcerative-colitis examples demonstrated computational analysis rather than new therapeutic experiments.
Drug-development teams could use agent groups to organize evidence and generate hypotheses. The associations do not establish causation, and proposed strategies still need testing. Read the paper.
Earth, climate, agriculture, and ecology
43. JADNet adds regional detail to weather fields
September 9 · Peer-reviewed AGU research article
JADNet, a U-Net model, converts broad 25-kilometer ERA5 weather fields into regional fields at four-kilometer resolution.
In case studies using 2014 Arabian Peninsula weather, it produced results in seconds and better matched the timing and spatial patterns of heat, heavy rain, and wind.
Regional forecasters and hazard researchers could use the method to explore detailed scenarios more quickly. The reported results concern the evaluated historical cases. Read the downscaling study.
44. WeedNet recognizes weeds at global and local scales
September 30 · Peer-reviewed Nature Communications article
WeedNet uses a self-supervised vision model, with an additional version fine-tuned for Iowa.
The global model classified 1,593 weed species at a reported 91% accuracy. The Iowa-specific model classified 84 local species at 97% accuracy.
Agronomists and precision-weeding developers could use the model to support scouting and more selective weed control. These are classification benchmarks; they do not by themselves demonstrate improved field outcomes or a new ecological finding. Read the study.
Chemistry, materials, and physical science
45. TDiMS makes molecular predictions easier to interpret
September 17 · Journal publication; earlier version appeared April 7, 2025
TDiMS represents molecules through distances between pairs of chemical substructures.
The descriptor improved predictions on tasks where distant parts of a molecule interact, while identifying which substructure pairs mattered to the prediction.
Chemists and materials researchers could use those interpretable features to investigate how molecular organization affects properties. The method helps connect a prediction with recognizable chemical features. Read IBM Research’s publication record.
Astronomy and space science
46. NASA and IBM release an open lunar foundation model
September 8 preprint; September 10 model release
The NASA–IBM Lunar Foundation Model combines 11 types of lunar imagery and maps. The release includes an open model checkpoint, the SomBench dataset, benchmark code, and fine-tuning tools.
Evaluations covered crater and volcanic-feature mapping, alongside estimates related to polar-ice stability.
Lunar researchers can reuse the model and datasets for exploratory analysis. The release establishes a research resource and benchmark results rather than a new lunar discovery. Read NASA’s announcement.
Mathematics and formal reasoning
47. Claude helps formalize Fermat’s Last Theorem
September 4 announcement; work log records completion August 17
Anthropic reported that Claude, Lean, Mathlib, and Prove2Me produced an end-to-end computer-checked proof of Fermat’s Last Theorem, following an existing mathematical proof.
Dozens of agents contributed, with approximately 13 million lines of Lean code reported. This was company research.
The advance is formal verification of known mathematics. Mathematicians and proof-assistant developers can inspect and build on the resulting artifact, while its wider impact on future verification remains to be established. Read Anthropic’s announcement.
AI company funding and research support
September’s activity also included grants, model access, computing infrastructure, partnerships, and company-supported science.
These forms of support should be read separately. A grant award, an announced commitment, an application ceiling, and an in-house research project each represent something different. Funding for a subproject also should not be added again when it is already included in a larger portfolio.
Funding and institutional partnerships
48. Public Data for Health supports shared life-sciences datasets
September 15 · OpenAI Foundation program announcement
The OpenAI Foundation, a separate nonprofit from OpenAI Group PBC, announced more than $125 million in initial grants for public life-sciences data.
The portfolio supports datasets and benchmarks for drug disposition, transporters, and blood–brain-barrier penetration; preservation of regulatory records from discontinued drug programs; and tumor and immune-response data for cancer-vaccine research. Read the Foundation announcement.
UNC independently announced a $40 million award for its Initiative for Generative Immunotherapy. That award is part of the portfolio, rather than an additional sum to add to its total. Read UNC’s recipient announcement.
A Nature report also identified $15 million for OpenADMET and $500,000 for CTD Commons. Those two amounts are secondary-reported figures, rather than amounts itemized in the reviewed official announcements. CTD Commons itself launched in January; the September milestone was Foundation support.
The practical aim is shared evidence that researchers beyond the original grantees can use. The announcement did not specify a complete payment or spending schedule, and the Foundation’s earlier $25 billion umbrella commitment is not counted again here.
49. A $60 million portfolio aims to bring forecasts to farmers
September 10 · OpenAI Foundation funding and scale-up announcement
The Foundation announced a $60 million commitment over three years to support weather and crop-disease forecasting for smallholder farmers.
The six-organization portfolio includes the University of Chicago, UC Berkeley, Precision Development, AIM for Scale through the University of Notre Dame, Digital Green, and CIMMYT.
The goal is to reach 100 million farmers across South and Southeast Asia and East Africa through channels including FarmerChat, SMS, and voice. Better forecasts could inform planting, harvesting, fertilizer use, and crop protection.
The reach and benefits are intended outcomes of the expanded portfolio, with evaluation still ahead. Read the program announcement.
50. OpenAI opens a research call on AI and teen development
September 8 · External research funding call
OpenAI Group PBC, distinct from the Foundation, opened a program with a total funding ceiling of up to $5 million and individual project awards of up to $1 million.
Independent teams can investigate how generative AI affects people ages 13–17, including use patterns, developmental outcomes, safeguards, and age-appropriate design. Researchers are not required to use OpenAI models.
Applications run through October 6, with selections expected by November 13. These are ceilings for an open call; no awards or findings were announced with it. Read the call and FAQ.
51. Novo Nordisk and Anthropic announce a drug-discovery collaboration
September 16 · Scientific R&D partnership announcement
Novo Nordisk and Anthropic announced plans to test Claude Science in selected research workflows and jointly develop solutions for drug-discovery challenges.
Anthropic supplies models and technical collaboration; Novo Nordisk’s scientists define the research problems and conduct the drug-discovery work. The announcement emphasized biological reasoning, human oversight, and data governance.
No funding value or specific disease targets were disclosed, and the partnership did not report a scientific discovery. Its significance lies in establishing a collaboration whose research outcomes can be followed. Read Novo Nordisk’s announcement.
52. AWS and Columbia expand biomedical research computing
September 10 · Research-computing partnership announcement
Amazon Web Services and Columbia University Irving Medical Center announced a collaboration providing cloud computing, AI and machine-learning resources, high-performance computing, technical training, and a secure research portal.
Support covers faculty projects, research startups, and early-career scholars. Intended applications include imaging analysis, biomarkers, health-outcome modeling, diagnostics, and individualized treatment protocols.
The partnership could help researchers analyze sensitive biomedical data with more computing and technical support. No per-project funding amount or new scientific result was disclosed. Read the announcement.
53. Google and Ohio State announce scientific research access
September 29 · Institutional infrastructure and research partnership
Ohio State announced a partnership involving Google Cloud, Google Public Sector, and Google DeepMind.
Researchers will gain access to more than 200 AI models, AI-optimized computing, Gemini Enterprise research agents, and scientific tools including AlphaFold, Google Earth Engine, and AlphaEvolve. The effort connects with Ohio State’s AI(X) Hub and Innovation District.
The partnership aims to support hypothesis generation and testing across biomedical engineering, Earth science, and other fields. The announcement did not disclose a grant total or a discrete scientific result. Read Ohio State’s announcement.
AI-company-supported research results
54. A complete male fruit-fly nervous-system map becomes public
September 3 · Cell paper and open dataset release
A team led by HHMI Janelia, with Google Research and other collaborators, released a wiring map of the male fruit fly’s brain and central nervous system containing more than 166,000 neurons and 125 million synaptic connections.
AI-assisted electron-microscopy segmentation, reconstruction, and computing contributed to the decade-long project. Google was a long-term research partner and co-author; no company funding amount was disclosed.
The open map enables studies of sensorimotor circuits and comparisons with female wiring. Read the project account and explore the dataset.
55. An electric-fish circuit reveals mechanisms of sensory prediction
September 2 · Peer-reviewed Nature article; corresponding preprint appeared in 2025
Researchers combined connectome mapping, electrophysiology, and computational modeling to investigate a sensory-prediction circuit in an electric fish.
The study connected inhibitory, disinhibitory, feedforward, and recurrent pathways with fast, accurate, noise-robust prediction and cancellation of sensory signals.
Google Research collaborated, co-authored, and contributed connectomics methods; no grant amount was disclosed. The work shows how a wiring map can become a mechanistic explanation when paired with physiological measurements and modeling. Read the study.
56. SynthID Bio tests watermarks for AI-generated proteins
September 30 · Peer-reviewed Nature paper and Google DeepMind announcement
SynthID Bio embeds detectable signals into designed protein sequences or predicted structures using ProteinMPNN and AlphaFold-based methods. Google DeepMind conducted the research, with Adaptyv Bio assisting in laboratory validation.
Watermarked binders targeting VEGF-A, SARS-CoV-2 spike RBD, and PD-L1 showed comparable binding affinity and hit rates to unwatermarked versions. Structure watermarking retained prediction accuracy, with near-perfect detection reported in the study.
The proof of concept could support biological-data provenance and screening. Operational deployment and resistance to deliberate removal remain open questions. Read the research announcement and Nature paper.
What we’ll be watching next
Taken together, September’s releases show AI becoming useful at several stages of research: choosing targets, designing experiments, interpreting measurements, checking proofs, and making published methods easier to reuse.
The follow-up questions are concrete. Which predicted materials get made? Which biological hypotheses survive further experiments? Which clinical tools improve care? Which funding commitments produce datasets that other researchers can actually use?
Those are the developments that will give future editions their substance.
And if your team publishes something we should be following, send it our way.
Related reading
- OpenAI’s Navier–Stokes Claim: The Proof, the AI, and the Fight — A closer examination of the mathematical claim, proof artifacts, and attribution questions.
- Stanford’s Paper2Agent Turns Research Papers Into Working AI Agents — More on making scientific methods executable and reusable.
