Tsinghua University / ByteDance
All-atom generative foundation model trained on 5M+ biomolecular complexes that designs small molecules, peptides, and nanobodies against a target site from one checkpoint.
AnewOmni is an all-atom generative foundation model for designing molecules that bind a specified protein target. Unlike modality-specific generators that are built and trained separately for small molecules, peptides, or antibodies, AnewOmni learns a single unified representation and uses one fixed checkpoint to design binders spanning the full range of molecular scales — from small chemical entities to large biologics — given only a target binding site. It was introduced in the 2026 bioRxiv preprint "Programming Biomolecular Interactions with All-Atom Generative Model" by Xiangzhe Kong and colleagues at Tsinghua University in collaboration with ByteDance's AI Drug Discovery group (Anew Labs / Anew Therapeutics).
The core idea is an "atom-to-block" latent space that decomposes any molecule into chemically meaningful building blocks at atomic resolution. Canonical units such as amino acids and nucleobases are treated as single blocks, while small molecules and non-canonical residues are fragmented into frequent substructures mined from chemical databases. By encoding atomic detail into these building blocks, the model learns transferable atomic interaction geometries that hold across modalities, allowing a binder to be assembled from a shared vocabulary regardless of whether the final molecule is a drug, a peptide, or a nanobody.
AnewOmni sits at the intersection of structure-based drug design and generative protein design, addressing a long-standing gap: most prior generative models specialize in a single molecular class. By unifying these tasks, AnewOmni enables cross-modality transfer learning, which is especially valuable for difficult or data-limited targets.
AnewOmni was trained jointly on more than 5 million biomolecular complexes spanning protein–small-molecule, protein–peptide, and protein–antibody interactions, enabling transferable interaction patterns across scales. The architecture combines an all-atom VAE for building-block compression with an E(3)-equivariant diffusion model operating on point clouds of building blocks, using classifier-free guidance for steerable generation. In wet-lab validation with only low-throughput screening, success rates ranged from 23% to 75% across modalities and targets. Against the historically undruggable KRAS G12D switch-II pocket, the model produced small-molecule inhibitors (IC50 of 24 and 36 µM), linear peptides (23% hit rate; best IC50 = 2.37 µM), cyclic peptides (35%), and nanobodies (75%; best Kd = 587 nM). For PCSK9, with no known binding site provided, it generated orthosteric peptides (57% hit rate; best Kd = 3.19 µM) and allosteric small molecules against a cryptic C-terminal pocket (30% hit rate; best Kd = 2.72 µM), with a crystal structure confirming a 0.92 Å RMSD between the generated and experimental binding pose and cellular validation of PCSK9 secretion inhibition. Code and a model checkpoint are released under an MIT license.
AnewOmni targets structure-based therapeutic discovery, where designers need a binder for a specific pocket but are uncertain which modality will work best. Because one model spans small molecules through biologics, medicinal chemists and biologics teams can explore drugs, peptides, and antibodies against the same site in parallel rather than committing to separate pipelines. The programmable graph prompts make it useful for constrained design problems — macrocyclization, covalent anchoring, scaffold growing, or incorporating non-canonical residues — and the binding-site conditioning supports allosteric and cryptic-pocket campaigns against targets like KRAS and PCSK9 that resist conventional approaches.
AnewOmni is presented as the first generative foundation model to demonstrate successful functional molecular design across all scales, from small chemical entities to large biologics, validated experimentally on two clinically important and historically difficult targets. By showing that a single atomic-resolution model can outperform modality-specific approaches and operate in data-limited regimes, it points toward general-purpose molecular reasoning engines for drug discovery. As a 2026 preprint with openly released code and weights, its real-world adoption and independent benchmarking remain to be established, and the reported success rates come from low-throughput validation on a limited set of targets.