Nesso-1: Accelerating Open-Source Binding Affinity Predictions

Why Binding Affinity?

We recently released Nesso-1, a model designed to predict binding affinity more efficiently than any existing cofolding-based open source model. Why focus on binding affinity? It’s a key problem in small molecule drug design. Specifically, the challenge is determining if a compound will bind to a specific target (and/or off-target), how strongly it will bind, and the concentration required to elicit a biological response. Once a disease target has been identified, the goal is often to inhibit or activate it, requiring the lowest possible dose to be efficacious whilst minimizing unwanted side-effects. While drug-discovery is a complex multi-optimization problem—for example, a compound binding to a target may lack key pharmacological properties or be too synthetically complex—many challenges become more relevant after strong binders have been found. Moreover, other risks affecting a drug candidate, such as off-target toxicity—stemming from a compound binding to other targets, with undesirable biological effects—could be significantly mitigated by being able to predict affinity at scale and accurately across the targetome.

Disrupting the physics-based Pareto front through AI and data

Depending on the stage of the drug-discovery pipeline, our constraints change drastically. During the initial phase of Hit-Identification (Hit-ID), we are looking for any hit – a compound that shows the desired activity. This is a broad search across a vast chemical space (often the Enamine REAL space) to collect diverse starting points that can later be refined into potent compounds with desirable pharmacological properties. Speed here is the main axis: if a screening tool takes too long to reject an inactive compound, it severely limits our exploration, which delays our campaigns or provides highly sub-optimal initial candidates. Eventually, we must turn hits into compounds with sufficiently strong binding affinity and very low affinity for related off-targets. In these later stages (Hit-to-Lead and Lead Optimization), the pool of compounds is smaller and structural details of the interactions in the pocket play an increasingly important role.

Traditionally, navigating the requirements of the different stages of early drug-discovery meant sliding along a physics-based Pareto front. We usually had to trade speed for physical accuracy. This led to a spectrum of methods operating at different computational trade-offs: structure-free approaches that execute in milliseconds, docking algorithms that treat proteins as rigid bodies to run in seconds, and simulation-based techniques that account for dynamics, solvation, and entropy at the cost of hours per run.

In the last few years, AI-models have started to disrupt this Pareto front, delivering unprecedented levels of both speed and accuracy. AlphaFold 2 revolutionized protein structure prediction. Since then, models have expanded to represent more complex biomolecular systems. One of the most notable developments is cofolding—where deep-learning models directly predict the stable, bound structure of a protein-ligand pair, entirely bypassing the trade-offs of classical docking. Building on this powerful new paradigm, a year ago MIT and Recursion announced Boltz-2, the first open-source model specifically designed to use these powerful joint structural representations for binding-affinity predictions. On public benchmarks, it showed that the accuracy of simulation-based methods can be approached at a fraction of the cost.

The challenges of AI-driven Structure-Based-Drug-Design

Cofolding models have immense potential for structure and affinity predictions. However, several challenges remain. Addressing them transparently will help build momentum sustainably.

• The Closed-Source Bottleneck: Since Boltz-2, most advances in cofolding and affinity prediction are still closed-source. This massively limits our ability to assess progress, further increasing the gap between the community and proprietary technologies.
• The Computational Cost of Cofolding: Cofolding is a powerful paradigm that enables handling protein-ligand interactions in a flexible way. However, the vastness of chemical space means current cofolding models may still be too expensive, limiting their use in early stages of the pipeline. The vision of one model to rule them all is likely ill-posed. Most compounds are not hits, thus running full-atom cofolding to reject them is a waste of compute.
• Human vs. AI Inductive Biases: We often approach the problem from a human-centric perspective. Structure is what a designer would look at and inspect. Physics is how we ground our predictions. But an AI-driven lab may have different pillars. For AI-powered designers, reasoning over actual 3D coordinates in Euclidean space can be suboptimal and whether physics needs to be integrated into our models or leveraged externally to assess proposals remains an open question.
• The long-lasting problem of public benchmarks: It is increasingly clear that common evaluation and training-split strategies are still affected by memorization issues and leakages. This significantly inflates the performance of AI models over physics-based computational approaches and distorts our perspective on progress.

What Nesso-1 unlocks, the remaining problems and the road ahead

Two weeks ago, our team at Valence and Recursion released Nesso-1, a coarse-grained cofolding model for predicting binding affinity of protein-ligand pairs. We share a diagram of the architecture below. The two main simplifications compared to existing cofolding baselines:

1. There is no generative module instantiated in full-heavy atom space, hence the structural representation learned is at the token-level inside the trunk (one residue = one token);
2. The trunk itself only operates over pairwise representations and does not have an MSA module, which results in a significantly smaller and more efficient network.

Compared to the leading open-source cofolding affinity model (Boltz-2), Nesso-1 is more than one order of magnitude faster and matches or surpasses their accuracy on a range of public and proprietary assays.

We built Nesso-1 specifically to address the challenges outlined above:

• Nesso-1 is open-source. Weights and model are accessible under a permissive license. We remain committed to open-sourcing and enabling the community to use and test our models freely. In light of the empirical results we shared in our technical report, Nesso-1 is the fastest and most accurate open-source affinity model publicly available. Nesso-1 is meant to be a foundational tool to keep building on and improving. Future updates and additional features are already scheduled and in-progress.

• Nesso-1 takes 1 second per prediction on a single GPU (H100). The coarse-graining approach (along with further tweaks to the classic formula) allows to run Nesso-1 up to targets of 800-900 residues taking 1 second per-prediction. To put this in perspective, on a budget of 12 H100 GPUs, we can now screen over 1 Million compounds every day at the accuracy of structure-based AI affinity models. Using Nesso-1 as a reward function when searching large chemical libraries is now not only feasible but affordable as well. Further acceleration to break the 1-second barrier is something we are actively working on.

• Removing full-heavy atom resolution at no accuracy cost. Nesso-1 builds on the premise that full heavy-atom resolution is a price we do not have to pay to match or surpass the accuracy of leading baselines. Nevertheless, the extent to which full atomistic resolution is needed still requires further investigation. Intuitively, there has to be a phase during a drug development where ignoring explicit atomistic interactions between target and ligand must hurt accuracy. Existing frameworks though, have not yet found a robust way of benefitting from this increased resolution. Whether this is purely a limitation stemming from the public data we have access to, or is affected by how these models are optimized, is a question we are aiming to address in the coming months. Importantly though, our report demonstrates that, despite the coarse-graining, information from the target is leveraged for downstream affinity predictions. In fact, we show that the model can achieve meaningful selectivity, when we evaluate its ability to predict the binding affinity of the same set of compounds against on and related off-targets.

• Beyond public benchmarks: evaluating AI models over real-world medicinal chemistry. When developing Nesso-1, we observed that public benchmarks commonly adopted in existing works (FEP+ and CASP16) reward overfitting. Models trained for longer can perform better on these datasets, while significantly struggling out-of-distribution. While we were able to achieve “SOTA” performance on FEP+, surpassing the values reported by closed-source models, these checkpoints resulted in worse accuracy on our internal assays. Similarly, removing any structure-based filtering — i.e. keeping complexes where the trunk output is noisier and less reliable — turned out to benefit the accuracy on these benchmarks. Overall, this is a function of the leakage constraints adopted, which only consider sequence similarity. Nesso-1 adheres to these constraints for enabling fair comparisons with existing approaches, but we will move away from them in future releases to enforce more realistic splits. To overcome these limitations, our evaluations focus on real-world medicinal chemistry, including the OpenBind affinity benchmark and 25 internal biochemical assays, with an average of several hundreds of compounds per assay. These benchmarks are far more out-of-distributions for models like Nesso-1 and Boltz-2 trained on BindingDB and ChEMBL, as shown below. Nesso-1 manages to outperform Boltz-2 on these more realistic cases, but there is still significant room for improvement. Unsurprisingly, zero-shot evaluation on proprietary chemistry that can differ wildly from the simpler training data, is a formidable challenge. In fact, on some of our most difficult assays we did not find a strong correlation between the model’s structural confidence and the downstream activity, indicating that the signal carried by the learned structural representations may not be sufficiently expressive out-of-distribution. Nonetheless, knowledge of the targets and chemistry can be leveraged to further improve internal versions of the model—we will share more on this soon.

Nesso-1 is a faster, more efficient foundation to keep iterating on. It represents a step forward towards deploying fast, reliable, and open-source AI in the early stages of drug discovery. Beyond that, Nesso-1 is deployed in production and is already supporting drug discovery programs at Recursion.The work continues: stay tuned for new features and updates on Nesso-1 and beyond.

If you want to know more about Nesso-1 and our evaluations, there is a technical report now on bioRxiv. More importantly, Nesso-1 is fully open-source and you can try it yourself, check it on github and our Model Card on Hugging Face.