Use AnewOmni Online
Commercially Available AnewOmni No-Code Web Server
AnewOmni: A Unified All-Atom Model for Programmable Molecular Design
AnewOmni is a unified generative framework, trained on more than 5 million biomolecular complexes, for designing molecules that bind a target site. Small molecules, peptides and antibodies have traditionally been designed with separate representations, rules and protocols. AnewOmni treats them in one framework. It assembles chemically meaningful building blocks at atomic resolution, which lets it transfer what it learns about molecular recognition across molecular scales.
The authors report that AnewOmni is the first to succeed in experimentally validated functional design across all scales, from small chemical entities to large biologics, within a single framework.
Why It Matters
Existing models tend to fall into two camps. Some represent large molecules as amino-acid tokens and small molecules as atom tokens, so atomic geometry, where the shared physics lies, is not modelled explicitly. Others treat everything as an atom graph, which captures fine detail but loses evolutionary structural priors. As a result, they are mostly limited to small molecules and short peptides. AnewOmni is designed to keep both atomic interaction geometry and modality-specific structure.
Key Capabilities
Cross-modality design: Small molecules, linear and cyclic peptides, and antibodies or nanobodies, from a single model.
Programmable graph prompts: Steer generation with chemical, topological and geometric constraints, without retraining.
Zero-shot generalisation: Binders for targets unseen in training, including DNA, RNA, glycans and phosphorylation sites.
Unsupervised likelihood ranking: Exact generative likelihood serves as a self-ranking score across modalities.
Experimentally validated: Binders confirmed against KRAS G12D and PCSK9, plus a crystal structure matching the designed pose.
How AnewOmni Works
AnewOmni has four main components.
Atom-to-block decomposition. Amino acids and nucleobases stay as standard units. Other chemistry is split into frequent fragments mined from ChEMBL, and rare substructures fall back to single atoms, so any chemistry is covered.
All-atom variational autoencoder. A reversible mapping between atom-level geometry and compact block-level latent points that keep atomic detail.
Latent diffusion model. Generation runs as a diffusion process over latent point clouds, conditioned on the atomic environment of the target binding site.
Exact likelihood estimation. An unsupervised, modality-agnostic score for ranking candidates.
Both the autoencoder and the diffusion model use an equivariant transformer. Training used 48 NVIDIA A800 GPUs.
Programmable Graph Prompts
Prompts attach chemical structures to latent points, or define how they connect. This lets users:
Design head-to-tail or disulfide cyclic peptides
Fix a sequence motif, such as a cell-permeability motif
Insert non-canonical amino acids from a user-defined library
Anchor a binder to a chosen residue on the target
Form a covalent bond to a specified target residue
Grow an existing scaffold, or design a linker between two fragments
Training Data
The model was trained on over 5 million binding-site and binder pairs. Small-molecule sources include BioLiP2, PDBBind v2020, CrossDocked2020 and the synthetic SIU dataset. Peptides came from PepBench and ProtFrag, and antibodies from SAbDab. A 40% sequence-identity cutoff separates training and test sets.
What the Benchmarks Show
Joint training helps: Across small molecules, peptides and antibodies, AnewOmni beat specialised baselines and its own single-modality variants on a composite in silico score.
Interaction recovery: 50% to 70% of native interactions were recovered in generated complexes. Cross-modality training added about 10% more recovery for small molecules and antibodies.
Nucleic acid targets: On 139 RNA/DNA targets, nearly all yielded at least one candidate above 50% of the native binding energy within 100 generations.
Zero-shot RNA discrimination: Generative likelihood separated binding from non-binding RNA ligands with an AUC-ROC of 0.77, without any RNA–small-molecule training data.
Large molecules: PoseBusters validity held up at molecular weights well above typical small molecules.
Experimental Validation
Success rates below come from low-throughput testing, without high-throughput display screening.
KRAS G12D switch II pocket
Modality | Result |
|---|---|
Small molecules | 2 of 3 synthesised compounds were hits, with IC50 of 24 µM and 36 µM. Similarity to MRTX1133 was only 0.12 to 0.15, meaning new scaffolds. |
Linear peptides | 7 of 30 achieved IC50 below 100 µM (23%). Best IC50 about 2.4 µM. |
Cyclic peptides | 7 of 20 achieved IC50 below 100 µM (35%). Best IC50 about 11 µM. No cyclic peptides were in the training data. |
Nanobodies | 3 of 4 designs bound with single-digit micromolar affinity (75%). A separate likelihood-selected design reached 587 nM. |
PCSK9
Orthosteric peptides: 4 of 7 peptides bound with Kd below 10 µM (57%), selected using generative likelihood alone. Best Kd was 3.19 µM.
Allosteric small molecules: AlphaFold3 predictions of PCSK9 with the clinical-stage compound AZD0780 converged on a cryptic C-terminal-domain site in 94.5% of 1,000 predictions. Designed compounds targeting it reached Kd of 2.72 µM and 2.98 µM.
Cellular activity: One compound raised LDLR expression at 100 µM to a level comparable to AZD0780 at 300 µM, and suppressed PCSK9 secretion.
Structural confirmation: A 2.70 Å crystal structure matched the designed binding pose with an RMSD of 0.92 Å.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with AnewOmni on Tamarind Bio
AnewOmni's pipelines pair generation with structure prediction and filtering. A platform like Tamarind can help teams:
Tackle "undruggable" targets: The paper demonstrates design against KRAS G12D, a target long regarded as undruggable, and against a large, flat PPI interface on PCSK9.
Find new binding sites: The PCSK9 work shows how cofolding predictions can reveal a cryptic allosteric pocket that then becomes a design target.
Compare modalities at one site: Generate small molecules, peptides and nanobodies against the same pocket in one framework.
Validate designs computationally: The antibody pipeline alternates AnewOmni CDR generation with AlphaFold3 complex prediction, and the paper notes that these post-filtering steps are computationally intensive.
How to Use AnewOmni on Tamarind Bio
Access the platform. Log in to tamarind.bio.
Prepare your target. Use a PDB or mmCIF structure of the target protein.
Define the binding site. Provide either a reference complex that includes a binder (the site is the residues within 10 Å of that binder), or a set of hotspot residues on the target alone. Use only one of the two modes. Hotspots should span the whole pocket, because the model defines the pocket as residues within 10 Å of them.
Choose a binder type. Options in the open-source release include small molecule, scaffold growth, linker design, covalent small molecule, linear peptide, disulfide cyclic peptide, head-to-tail cyclic peptide, and antibody CDR design (single or multiple CDRs). Antibody frameworks must use Chothia numbering.
Set size and generation parameters. For example, peptide length (4 to 24 residues in the open-source release) and the number of candidates per template. The released defaults generate 20 candidates per template.
Generate candidates. Run AnewOmni from the open-source repository (MIT license) or the official web server.
Filter for validity. The paper's small-molecule pipeline applies PAINS filters, PoseBusters checks, key-interaction constraints, Glide refinement and molecular dynamics. Peptides are filtered for geometry, clashes, hydrophobicity and L-amino acids.
Validate with cofolding. Check complexes with structure-prediction tools such as AlphaFold or Boltz-2 on Tamarind. The paper's antibody pipeline uses AlphaFold3 and, in the released code, Protenix or Boltz-2 backends.
Rank and select. Rank by generative likelihood, and by AlphaFold3 metrics for antibodies, then prioritise candidates for experimental testing.
Source: Kong et al., Programming Biomolecular Interactions with All-Atom Generative Model, bioRxiv, 2026. DOI: 10.64898/2026.03.12.711044