Valence Labs

To identify the most promising drug candidates, we need to understand not only whether a molecule is likely to bind to a target, but how strongly.

Today, we’re excited to introduce Nesso-1, an open-source, coarse-grained co-folding model that significantly accelerates binding-affinity predictions.

For machine learning models to be lifelong learners, they must adapt to new tasks without forgetting previous capabilities.

For continuous-time generative models, like flow matching and diffusion models, the challenge isn’t only absorbing new information, but also remembering what came before.

In a new work, we introduce trajectory replay, a continual learning technique that is designed to remind the model what it already knows.

Traditional techniques, like generative replay, use generated samples of the historical data distribution to retrain the model on what it’s previously seen. But these methods can be expensive, and errors in the generated samples accumulate over time.

By contrast, trajectory replay involves retraining the model on the velocity of the previous model’s generative path, without ever generating a full sample. This approach is cheaper and more robust to errors, which is crucial for lifelong learning.

For more information about trajectory replay and how it can be used to support the continual learning of perturbation models in new biological contexts, check out the links below.

Paper: arxiv.org/abs/2610.26749
Code: github.com/valence-labs/t…

To predict binding affinity, would you defer to lessons from physics or AI?

No matter which option you choose, you’re unlikely to get where you’re looking to go.

Check out Daniel Ramirez’s op-ed where he offers a third option, one that has the potential of going further than any one field alone.

Link to op-ed: valencelabs.substack.com/p/chasing-the-…

Just as your two hands are mirror images of one another, so are many of the molecules in the human body.

But mirror images are not identical, and their differences are biologically relevant.

A fact that our models can silently fail to capture, even with chiral tags.

Enter CIP labels.

To learn how chemicals in the body are labeled so that AI tools can tell the difference between mirror images, check out Cristian's video below.

Full Video: youtu.be/0tT-UqhHbpI

A generative model can look like it's working and still fail to output the disease-relevant cell groups or small molecules you care most about.

Fortunately, we have techniques to diagnose and correct these failures.

In our most recent Valence at the Whiteboard, Liam explains how to evaluate generative models using Precision, Recall, Density and Coverage (PRDC) to spot subtle failure modes and know what to do about them.

Full video: youtu.be/teIoP92Aeow