Nesso-1: Accelerating Open-Source Binding Affinity Predictions @valence_ai 1. Nesso-1 introduces an open-source, coarse-grained cofolding framework that predicts protein–ligand binding affinity in approximately one second on a single GPU. It is more than 10× faster than Boltz-2 on average and exceeds 20× speed-up for larger proteins, allowing up to 20× more compounds to be screened under the same compute budget. 2. Its central design choice is to remove the computationally expensive heavy-atom diffusion module. Instead of generating a complete atomistic complex, Nesso-1 learns token-level geometric relationships between ligand heavy atoms and protein residue centers, suggesting that full-atom poses are not always necessary for accurate affinity prediction. 3. The model uses a streamlined Pairformer trunk that updates only pairwise token representations. It removes MSA processing, incorporates pretrained ESM-2 protein embeddings, and predicts protein–ligand distance distributions that encode both coarse structural information and intrinsic uncertainty. 4. During recycling, Nesso-1 processes the full protein only in the first pass and subsequently retains residues predicted to lie within 22 Å of the ligand. Its affinity module further focuses on protein tokens within an expected 15 Å cutoff, substantially reducing computation without measurably compromising affinity accuracy. 5. The affinity module predicts both binding likelihood for hit identification and continuous affinity for hit-to-lead ranking. It is trained with classification and regression objectives, including an up-weighted intra-assay difference loss that emphasizes compound ranking within the same experiment while reducing sensitivity to biases between assay protocols. 6. On the public FEP+ 4, OpenFE, and CASP16 benchmarks, Nesso-1 achieves weighted Pearson correlations of 0.80, 0.66, and 0.68, respectively, outperforming Boltz-2 in all three settings. It also exceeds several simulation-based baselines while requiring seconds rather than the multiple GPU-hours associated with free-energy calculations. 7. The authors caution that these commonly used public benchmarks are chemically similar to BindingDB and ChEMBL training data and should therefore be considered largely in-distribution. They show that sequence-based leakage controls alone do not guarantee a realistic assessment of generalization to novel medicinal chemistry. 8. On the more challenging OpenBind benchmark, containing 494 compounds targeting the conformationally flexible EV-A71 2A protease, Nesso-1 achieves the lowest RMSE and highest Spearman correlation among the evaluated AI models. However, molecular weight remains a surprisingly competitive baseline, highlighting the difficulty of zero-shot affinity prediction from public training data. 9. Across 25 internal biochemical assays, Nesso-1 reaches a weighted Pearson correlation of 0.44, compared with 0.37 for Boltz-2 and 0.27 for molecular weight. Its performance remains relatively stable across different levels of chemical similarity to the training set, although substantial assay-to-assay variability confirms that real-world out-of-distribution prediction remains unresolved. 10. Nesso-1 also demonstrates meaningful selectivity in selected target families by distinguishing the potency of identical compounds between primary targets and related off-targets. The results suggest that coarse cofolding representations can capture target-dependent interactions rather than relying only on ligand-level properties, although performance declines for highly unfamiliar targets and chemistry. 11. The study argues that atomistic structural detail may sometimes introduce noisy, high-frequency information into affinity models. Nesso-1 therefore establishes a practical speed–accuracy baseline for high-throughput screening and large-scale fine-tuning, while acknowledging that full-atom modeling, solvent, entropy, and water-mediated effects may still be essential during later-stage lead optimization. 💻Code: github.com/recursionpharma/n… 📜Paper: valencelabs.com/wp-content/u… #Nesso1 #BindingAffinity #DrugDiscovery #ComputationalBiology #ProteinLigand #VirtualScreening #Cofolding #MachineLearning #OpenSourceAI