Nesso-1

Nesso-1 is a coarse-grained cofolding model for protein–ligand binding affinity prediction, developed by Valence Labs, the research arm of Recursion. Given a protein amino-acid sequence and a ligand specified as a SMILES string, CCD code, or SDF file, it returns a binding-affinity scalar plus a binder/non-binder classification score, addressing the two workhorse tasks of early small-molecule discovery: triaging large virtual libraries (hit identification) and ranking close analogs within a chemical series (hit-to-lead).

The model's central design choice is subtraction. Cofolding systems such as Boltz-2 spend most of their compute predicting all-atom structure through a diffusion module, then read affinity off the resulting complex. Nesso-1 removes that diffusion module entirely, keeps only a pairwise token-level representation, and drops multiple sequence alignment inputs in favor of frozen pre-trained ESM-2 protein embeddings. What remains is a much lighter trunk that predicts affinity directly, at roughly one second per prediction on a single H100 GPU — more than 10× faster than Boltz-2 on average, and above 20× on larger proteins. It is an openly licensed counterpart to closed-source coarse-grained cofolding models such as TerraBind, trained entirely on public data and without proprietary ligand embeddings; code and weights are released under Apache-2.0.

#Key Features

• Diffusion-free cofolding: The heavy-atom diffusion module is removed, and the trunk updates only pairwise token representations over ligand heavy atoms and protein residue centers (Cβ, or Cα for glycine), yielding a large parameter reduction relative to Boltz-2.
• No MSA requirement: Rather than searching sequence databases at inference time, the model conditions on frozen ESM-2 embeddings, which removes the MSA pipeline from the critical path and compensates for the lighter trunk.
• Uncertainty-driven cropping: A normalized protein–ligand entropy over the predicted distogram acts as an intrinsic structure-uncertainty signal and gates cropping — full protein context on the first recycling pass, then only tokens within a predicted 22 Å of the ligand.
• Dual affinity head: A separate, smaller Pairformer predicts both binding likelihood for hit identification and continuous affinity for hit-to-lead, conditioned on the atom encoder, the frozen protein language model embeddings, and the pairwise representation masked to protein tokens within 15 Å of any ligand heavy atom.
• Assay-bias-aware regression: The Huber regression loss splits into absolute and relative-difference terms, with the relative term up-weighted so that intra-assay potency differences drive learning rather than inter-assay calibration offsets.

#Technical Details

The trunk follows the standard cofolding blueprint: an atom encoder embeds the protein sequence and ligand graph, and a Pairformer applies triangle attention and triangle multiplication to produce a pairwise token representation trained with categorical cross-entropy over distance bins. Structural training data comprises PDB crystal structures released before 30 September 2021, AlphaFold DB distillation targets, and a curated subset of distilled protein–ligand complexes from SAIR. Affinity curation mirrors Boltz-2: BindingDB and ChEMBL v34 for regression; PubChem 1.8.1, the CeMM Fragment Dataset, and MIDAS metabolite data for classification, with matched decoy generation and the same constraint removing training points sharing ≥90% protein sequence similarity with the FEP+ benchmark. Ki, Kd, IC50, and EC50 labels are treated interchangeably, and classification uses a focal loss.

On Boltz-2's own evaluation suite, Nesso-1 matches or surpasses it: highest average precision on MF-PCBA with comparable enrichment factor, plus competitive results on 876 FEP+/OpenFE pairs, an 87-compound four-kinase ABFE subset (CDK2, TYK2, JNK1, P38), and 140 CASP16 affinity complexes. It attains the highest accuracy among AI models on the out-of-distribution OpenBind benchmark and outperforms Boltz-2 across 25 internal biochemical assays. Pose quality is assessed on FoldBench, PoseBusters, and Runs N' Poses, where a full-context variant matches or slightly exceeds TerraBind.

#Applications

The model targets computational chemists and discovery teams who need affinity estimates at library scale. Its roughly one-second inference cost makes it practical for virtual screening campaigns where physics-based free energy perturbation is prohibitively expensive, and the classification head supports hit triage before any compound is synthesized. Because it requires no MSA search and no proprietary embeddings, it slots into open pipelines and can be fine-tuned on in-house assay data.

#Impact

Nesso-1 shows that much of the compute in cofolding affinity models goes to structural detail the affinity readout does not need, and that a permissively licensed model trained only on public data can match a leading open competitor while running an order of magnitude faster. The authors are candid about the limits. Zero-shot transfer to real medicinal chemistry programs remains difficult, with weak performance on some internal assays. The FEP+/OpenFE and CASP16 sets are largely in-distribution relative to BindingDB and ChEMBL, so those results should not be read as general superiority over free energy perturbation methods, a comparison the authors explicitly caution against. On OpenBind the margin over a simple molecular-weight ranking baseline is modest, even though it beats other AI models clearly. The coarse-grained representation also cannot resolve side-chain placement in the binding pocket, leaving open whether heavy-atom generative modeling is still required for precise cofolding.