Use Nesso-1 Online
Commercially Available Nesso-1 No-Code Web Server
Nesso-1: Fast Open-Source Binding Affinity Prediction with Coarse-Grained Cofolding
Nesso-1 is a coarse-grained cofolding framework for protein-ligand binding affinity prediction from Valence Labs, Recursion. It needs about one second per prediction on a single GPU, more than an order of magnitude faster than the leading open-source baseline, Boltz-2, while matching or surpassing its accuracy on the same benchmarks. Code and weights, trained strictly on public data, are open source at github.com/recursionpharma/nesso.
Predicting how tightly a compound binds a protein is central to target-based drug design. Docking is fast but depends on a suitable holo structure and can miss induced-fit changes, while free energy perturbation is accurate but computationally expensive and sensitive to setup. Cofolding models predict the protein-ligand complex from sequence and SMILES, and Boltz-2 added an affinity module on top of them.
The report argues that full-atom resolution makes each Boltz-2 prediction take about 20 seconds, which is prohibitive for large hit-identification campaigns, and may be unnecessary for a separate affinity module. Nesso-1 keeps the Boltz-2 training data and optimisation framework but removes much of the architecture, notably the atomistic diffusion module.
Property | Detail |
|---|---|
Task | Binding affinity prediction (and binder versus non-binder classification) for protein-ligand pairs |
Inputs | Protein sequence and ligand SMILES; protein embeddings from ESM-2 |
Speed | About 1 second per prediction on one GPU; more than 10x faster than Boltz-2 on average, above 20x on larger proteins |
Developers | Valence Labs, Recursion |
Availability | Open-source code and weights trained on public data |
Why Coarse-Grained Cofolding Speeds Up Affinity Prediction
Nesso-1 works on a token-level, coarse-grained representation of the complex rather than generating heavy-atom structures. This removes the costly generative step, in line with the approach of the closed-source TerraBind model, but Nesso-1 does not rely on proprietary ligand embeddings.
Speed: timed on a single H100 across three targets of roughly 300 to 1000 residues, excluding fixed per-target costs such as loading ESM-2 embeddings or computing MSAs for Boltz-2.
Throughput: the authors report that the speed-up lets them screen up to 20 times more compounds under the same compute budget.
Versus simulation: OpenFE and absolute free energy calculations are reported to take 6 to 12 GPU hours per prediction, so Nesso-1 is described as more than 10,000 times faster.
Fine-tuning: the lower cost makes rapid fine-tuning on continuous data streams practical.
Nesso-1 Benchmarks: In-Distribution and Out-of-Distribution Accuracy
Nesso-1 was tested on the benchmarks used in the Boltz-2 study (MF-PCBA, the FEP+ subsets including OpenFE, and CASP16), on the newer OpenBind affinity benchmark, and on 25 internal biochemical assays.
Benchmark | What it tests | Reported outcome |
|---|---|---|
MF-PCBA | Active versus inactive classification at screening scale | Matches or exceeds Boltz-2; well ahead of docking and ML baselines |
OpenFE (876 pairs), FEP+ 4 targets (87 compounds), CASP16 (140 complexes) | Ranking compounds within a series | Outperforms Boltz-2 and, in the report's figures, OpenFE and ABFE |
OpenBind | Harder, lower-similarity hit-to-lead data | Lowest RMSE and highest Spearman among AI models tested |
25 internal assays | Real-world medicinal chemistry | Outperforms Boltz-2 overall; limited on some assays |
Selectivity and Interpretation
The authors show examples where Nesso-1 separates the affinities of identical compounds against on-targets and related off-targets. They also caution that OpenFE and CASP16 are largely in-distribution for models trained on public data, so benchmark scores should not be read as equivalent to FEP performance.
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 Virtual Screening with Nesso-1 on Tamarind Bio
High-throughput virtual screening: score large compound sets against a target at roughly one second per compound.
Hit identification: use the classification head to separate likely actives from inactives.
Lead optimisation triage: rank analogs within a chemical series by predicted affinity.
Selectivity checks: compare predicted affinity of a compound against a target and related off-targets.
How to Use Nesso-1 on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select Nesso-1.
Enter the protein: provide the target amino acid sequence, as the model works from sequence.
Enter ligands: provide one or more compounds as SMILES strings.
Set options: adjust settings such as the number of recycling steps, if exposed.
Run the job: submit and wait for predictions.
Download results: retrieve predicted affinities and binding likelihoods per compound.
Rank and follow up: use scores to prioritise compounds, then validate top picks with docking, simulation or assays.
Parameters shown on Tamarind may differ from the open-source release, so refer to the tool page for what is exposed.
Things to Keep in Mind
Zero-shot generalisation to real-world medicinal chemistry remains hard; the report acknowledges assays where performance is limited.
On OpenBind, the improvement over a simple molecular weight ranking is modest.
Nesso-1 is primarily an affinity model; the lack of atomistic resolution may limit structural tasks where fine-grained interactions matter.
Many public benchmarks overlap with the training data, so reported scores may overstate performance on novel chemistry.
Source: Shenoy N, Errington D, Bengio E, Kapusniak K, Klaeser K, Pang YT, Radenkovic V, Tossou P, Bois T, Wedlake A, Di Giovanni F, Valence Labs, Recursion. "Nesso-1: Accelerating Open-Source Binding Affinity Predictions," technical report. Code and weights: https://github.com/recursionpharma/nesso