The science we build on.

Introduces ESM3, a generative model that reasons over protein sequence, structure and function together. The authors used it to design a working fluorescent protein that is distant from any known natural sequence.

Hayes T, et al. Simulating 500 million years of evolution with a language model. Science 387:850–858 (2025). doi:10.1126/science.ads0018(opens in a new tab)

Shows that a language model trained only on protein sequences learns enough to predict 3D structure from a single sequence (ESMFold), fast enough to fold more than 600 million metagenomic proteins.

Lin Z, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379:1123–1130 (2023). doi:10.1126/science.ade2574(opens in a new tab)

AlphaFold predicted protein structures with accuracy competitive with experiment in the CASP14 assessment, the clearest sign that learned models can crack a core problem in biology.

Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589 (2021). doi:10.1038/s41586-021-03819-2(opens in a new tab)

Biohub's announcement of the ESMC and ESMFold2 models and the ESM Atlas. Per the announcement, the models are MIT-licensed and the atlas is available as open data on AWS.

Biohub. ESMC, ESMFold2 and ESM Atlas: a world model of protein biology (May 27, 2026). Source(opens in a new tab)

Geneformer: a transformer pretrained on about 30 million single-cell transcriptomes that, with little task-specific data, predicts how gene networks respond to change and nominated candidate targets in cardiomyopathy.

Theodoris CV, et al. Transfer learning enables predictions in network biology. Nature 618:616–624 (2023). doi:10.1038/s41586-023-06139-9(opens in a new tab)

A generative model pretrained on over 33 million cells that transfers to cell-type annotation, batch integration, perturbation response prediction and gene network inference.

Cui H, et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 21:1470–1480 (2024). doi:10.1038/s41592-024-02201-0(opens in a new tab)

scFoundation: a 100-million-parameter model trained on roughly 50 million human single-cell transcriptomes, applied to expression enhancement, drug response and perturbation prediction.

Hao M, et al. Large-scale foundation model on single-cell transcriptomics. Nature Methods 21:1481–1491 (2024). doi:10.1038/s41592-024-02305-7(opens in a new tab)

Trained jointly on dissociated and spatially resolved single-cell data, Nicheformer learns cell representations that carry information about a cell's tissue neighborhood.

Tejada-Lapuerta A, Schaar AC, et al. Nicheformer: a foundation model for single-cell and spatial omics. Nature Methods (2025). doi:10.1038/s41592-025-02814-z(opens in a new tab)

Universal Cell Embeddings (UCE) maps cells from any tissue or species into one shared space without retraining, a step toward a common coordinate system for cell state. Preprint.

Rosen Y, et al. Universal cell embeddings: a foundation model for cell biology. bioRxiv (2023). doi:10.1101/2023.11.28.568918(opens in a new tab)

A perspective from a broad group of researchers setting out a roadmap for AI virtual cells that span molecules, cells and tissues, and the data needed to build them.

Bunne C, et al. How to build the virtual cell with artificial intelligence: priorities and opportunities. Cell 187(25):7045–7063 (2024). doi:10.1016/j.cell.2024.11.015(opens in a new tab)

A RIKEN robot paired with Bayesian optimization searched about 200 million possible culture conditions by testing 143 of them, and improved iPSC-derived retinal pigment epithelium production by 88% on pigmentation scores.

Kanda GN, et al. Robotic search for optimal cell culture in regenerative medicine. eLife 11:e77007 (2022). doi:10.7554/eLife.77007(opens in a new tab)

Single intestinal stem cells, given the right matrix and growth factors, organized themselves into gut-like structures. This was the founding demonstration of adult stem cell organoids.

Sato T, et al. Single Lgr5 stem cells build crypt-villus structures in vitro without a mesenchymal niche. Nature 459:262–265 (2009). doi:10.1038/nature07935(opens in a new tab)

A review of how organoids recapitulate organ development, what they can model, and where variability still limits them.

Lancaster MA, Knoblich JA. Organogenesis in a dish: modeling development and disease using organoid technologies. Science 345:1247125 (2014). doi:10.1126/science.1247125(opens in a new tab)

A microfluidic device with a stretching alveolar-capillary interface reproduced organ-level lung responses, some of which depended on breathing-like mechanical strain.

Huh D, et al. Reconstituting organ-level lung functions on a chip. Science 328:1662–1668 (2010). doi:10.1126/science.1188302(opens in a new tab)

A review of organ-on-chip systems as human-relevant models for disease, drug development and patient-specific testing.

Ingber DE. Human organs-on-chips for disease modelling, drug development and personalized medicine. Nature Reviews Genetics 23:467–491 (2022). doi:10.1038/s41576-022-00466-9(opens in a new tab)

Evaluates scGPT and Geneformer without fine-tuning and finds that, on several tasks, simpler baselines match or beat them.

Kedzierska KZ, Crawford L, Amini AP, Lu AX. Zero-shot evaluation reveals limitations of single-cell foundation models. Genome Biology 26 (2025). doi:10.1186/s13059-025-03574-x(opens in a new tab)

Removed the statutory requirement that a drug be tested in animals before human trials, and allows nonclinical tests such as cell-based assays, organ chips and computer models.

FDA Modernization Act 2.0 (signed December 2022). Source(opens in a new tab)

FDA set out a roadmap to reduce, refine and replace animal testing, starting with monoclonal antibodies, naming organoids, organ-on-chip systems and AI computational models as alternatives. The stated aim is for animal studies to become the exception within three to five years.

FDA press release, April 10, 2025: FDA Announces Plan to Phase Out Animal Testing Requirement for Monoclonal Antibodies and Other Drugs, with accompanying roadmap. Source(opens in a new tab)