Najat Khan, Chief Commercial Officer at The Janssen Pharmaceutical Companies and board member at Recursion, shared on LinkedIn:
“Predicting how tightly a molecule binds to its target is one of the fundamental challenges in drug discovery. Until now, researchers have largely faced a tradeoff: fast methods sacrifice accuracy, while more accurate approaches are often too computationally expensive to use at the scale modern drug discovery demands.
Last year, our team at Recursion worked with the Massachusetts Institute of Technology to advance the field with Boltz-2, showing that AI could make binding affinity prediction both more accurate and more accessible.
Today, we are excited to share the next release from our Valence Labs: Nesso-1, which we are also open sourcing.
For years, there’s been an assumption that achieving better binding affinity prediction requires increasingly detailed, atomistic models. But Nesso-1 builds on thoughtfully-designed course-grained representations architecture to achieve comparable predictive performance at a fraction of the computational cost.
The result is a model that predicts binding affinity in approximately 1 second on a single GPU – up to 20× faster than Boltz-2 – while matching or surpassing its performance across public benchmarks and, importantly, a diverse set of internal biochemical assays that better reflect real-world medicinal chemistry.
By leveraging NVIDIA cuEquivariance, we’ve also accelerated training and inference by an additional 2-3×.
Why does this matter?
Because in AI-native drug discovery, speed isn’t just an engineering metric – it determines how much science you can explore. Faster affinity prediction means screening much larger chemical libraries, running many more design iterations, and bringing autonomous drug discovery workflows closer to reality. Instead of forcing researchers to choose between computational efficiency and predictive performance, it demonstrates that thoughtful model design can advance both simultaneously.
We’re already using Nesso-1 internally alongside our proprietary datasets to accelerate active drug discovery programs, and we’re excited to continue improving it with the community.
Open science accelerates innovation. We hope Nesso-1 becomes a foundation the community can build on as we collectively push toward faster, more effective drug discovery.
Weights and code are fully open source:
Title: Nesso-1: Accelerating Open-Source Binding Affinity Predictions
Authors: Valence Labs, Recursion, Nikhil Shenoy, David Errington, Emmanuel Bengio, Kacper Kapuśniak, Kerstin Klaeser, Yui Tik Pang, Vladimir Radenkovic, Prudencio Tossou, Therence Bois, Andrew Wedlake, Francesco Di Giovanni
Read The Full Article
Other articles about AI on. OncoDaily.