Genesis Molecular AI

Genesis Molecular AI, previously called Genesis Therapeutics, develops AI for small-molecule drug discovery. Its Genesis Exploration of Molecular Space (GEMS) platform combines language models, diffusion models and physical simulation to generate and optimize molecules, integrating structure prediction with predictions of potency and other drug-like properties. The company develops medicines for both its own pipeline and pharmaceutical partners.[1][6][3][2]

Genesis introduced Pearl—“Placing Every Atom in the Right Location”—on October 28, 2025. The generative 3D foundation model performs cofolding: predicting the structure of a protein together with a bound molecule. This addresses a central difficulty in drug design: a target protein can change shape when a ligand binds, so predicting the ligand’s placement also requires accounting for protein flexibility. At launch, Genesis described Pearl as deployed across its internal pipeline and partnership programs, and credited NVIDIA subject-matter experts as co-authors of its technical report.[3][5]

Pearl trains on proprietary synthetic structures generated using physics. Its architecture includes an SO(3)-equivariant diffusion head, which encodes rotational geometry: rotating a molecule does not change its identity. In unconditional mode, Pearl takes a protein sequence and a ligand’s 2D chemical structure. Scientists can also condition predictions on a known binding pocket or experimental structure, allowing them to test specific structural hypotheses.[3]

Genesis distinguished its launch benchmarks from production deployment. The benchmark models used public Protein Data Bank structures with a September 30, 2021 cutoff; in practice, the company said it trained a larger model on more public structures and proprietary structures. Its reported improvements over AlphaFold 3 therefore concerned specified benchmark comparisons, rather than every deployed version of Pearl.[3]

By June 3, 2026, Genesis had extended production Pearl into the Pearl system. It generates diverse binding poses through inference-time scaling, steers them toward physically plausible structures, and ranks them using physics- and AI-based scoring. On OpenBind’s 802 follow-on complexes for the EV-A71 2A protease, Genesis reported 78% success, evaluated best-of-25 among its top-ranked poses: a complex counted as successful if at least one of those 25 poses jointly met ligand RMSD ≤2 Å, PoseBusters validity and LDDT-PLI ≥0.8 against the experimental structure. The zero-shot evaluation supplied a protein sequence, ligand SMILES and an unbound protein crystal structure as a template, without target-specific tuning or binding-site information.[4]

The target presents a concrete induced-fit challenge: a loop beside the binding site must rearrange to accommodate the ligand. Genesis reported that, for compound A71EV2A-x4519a, the Pearl system predicted that loop movement and placed the ligand within 0.28 Å RMSD of its crystallographic pose, while passing PoseBusters and the LDDT-PLI ≥0.8 criterion. All 25 top-ranked poses reportedly passed the strict sub-angstrom criterion. Genesis credited the OpenBind Consortium for the benchmark and crystallographic data, and NVIDIA for Triton and cuEquivariance support that accelerated inference.[4]

In the July 1, 2026 Latent Space interview, co-founder Evan Feinberg argued that conventional 2 Å pose thresholds can conceal errors that mislead medicinal chemists: “If your model is sitting at 1.8, 1.9 Angstrom RMSD, that’s slop, most likely.” Brandon Anderson’s companion article described Genesis’s internal SAPPHIRE system as reasoning about poses, forming hypotheses, reading literature, using internal tools and creating candidates for another iteration. Feinberg and CTO Sergey Edunov argued that improved structural accuracy had made such agentic discovery loops possible. Edunov had joined Genesis from Meta, where he led Llama 2 training and Llama 3 pretraining.[1]

Genesis’s June 10, 2026 distillation report introduced DeCAF-Pearl, a flow-map model that learns to jump between points along a diffusion trajectory instead of taking many small denoising steps. Two design decisions addressed difficulties specific to molecular modeling: expressing the flow map directly in noise-level space avoided unstable gradients, while predicting clean structures preserved the teacher model’s rigid-alignment loss. That alignment accounts for the fact that rotating or translating a protein does not change the protein.[5]

Genesis reported approximately 20× fewer diffusion-sampling model calls and approximately 5× faster full inference, retaining most of the teacher’s prediction quality. The teacher was Pearl-2026.1.dev, a development checkpoint trained with a December 2023 cutoff; evaluation used 196 post-2023 Runs N’ Poses structures, best-of-five. This distillation effort was separate from the production Pearl system evaluated on OpenBind. Genesis also developed DeCAF-Search to maintain candidate populations, look ahead with the flow map, refine structures, reintroduce noise and allocate compute through a selection rule.[5]

The distillation report acknowledged MIT collaborators Ron Shprints, Peter Holderrieth, Juno Nam, Rafael Gómez-Bombarelli and Tommi Jaakola; Carnegie Mellon University collaborator Nicholas Matthew Boffi; and Joey Bose from College London and Mila.[5]

These structure-prediction projects followed an earlier period of company formation, platform development and pharmaceutical collaborations under the Genesis Therapeutics name.[1][9][3]

Genesis was co-founded in 2019 by Evan Feinberg and Ben Sklaroff; a historical company release identified Sklaroff as VP of Engineering. The company spun out of Stanford University and incorporated in 2019. Vijay Pande led the Stanford group responsible for some of the deep-learning technologies used by Genesis.[8][9]

Genesis closed a $52 million Series A in 2020. In August 2023, C&EN reported that Genesis said it had closed a $200 million Series B, led by an unnamed US life-science investor and others including Andreessen Horowitz. Genesis said the money would support GEMS development and advance its small-molecule pipeline toward clinical development. By that time, it had worked with Eli Lilly and Company and Genentech, while the pipeline’s contents remained undisclosed.[9]

Genesis’s company timeline reports Gilead and Incyte partnerships in 2024–2025 with $65 million upfront across both deals. The February 20, 2025 Incyte agreement specified two initial Incyte-selected targets, a $30 million upfront payment to Genesis and an option for an additional target. GEMS would discover and optimize small molecules, with Incyte receiving exclusive rights for potential clinical development and commercialization. Genesis was eligible for up to $295 million per target if all development, regulatory and commercial milestones were achieved, plus tiered royalties on sales of approved collaboration products.[2][6]

The May 20, 2026 expansion announcement stated that Genesis was to receive $120 million, with potential success-based milestone payments and royalties. Genesis’s timeline describes the consideration as upfront and says the expanded collaboration added at least five targets and provided experimental data from Incyte for GEMS training. The collaboration thus supplied both additional discovery programs and data for model development.[7][2]

The February 2025 joint announcement reported that Genesis had raised more than $300 million, was headquartered in Burlingame, California, and operated a fully integrated laboratory in San Diego.[6]