Ai-based computational screening platform for small-molecule binder discovery and optimization

Overview

An AI-based computational screening platform designed to accelerate the discovery and optimization of small-molecule binders. The platform currently supports protein targets and is being extended to glycosaminoglycans, enabling efficient identification of ligands with improved biochemical properties. By mapping chemical space into a meaningful latent space, the platform supports multiobjective optimization along key axes including binding affinity, blood-brain barrier penetration, toxicity, solubility, and synthesizability. This approach allows researchers to adapt existing binder solutions and explore novel chemical variants faster than traditional screening methods.

Technical specifications

Core capabilities:

• AI-based virtual screening for predicting protein-ligand binding affinity, validated against public binding datasets with state-of-the-art results on in-vitro binding affinity prediction
• Variational autoencoder approach using SELFIES molecular representations to embed molecules in a navigable latent space
• Generation of valid, similar molecules through latent space exploration
• Navigation of the embedding space along any axis of predictable values, including toxicity, solubility, and binding affinity
• Extension to glycosaminoglycan binder discovery, enabling exploration of this underexplored chemical space
• Flexible adaptation to new optimization axes based on collaborator needs

Key benefits:

• Faster discovery of ligands with improved biochemical properties
• Multiobjective optimization addressing binding affinity, blood-brain barrier penetration, toxicity, solubility, and synthesizability simultaneously
• Reduced experimental screening burden through computational pre-filtering

Technology readiness level

The virtual screening platform has been validated on public binding datasets and achieved state-of-the-art results in predicting in-vitro binding affinity values. The latent space navigation approach using variational autoencoders has been validated for embedding molecular descriptions and generating valid similar molecules, as well as for navigating along predictable biochemical axes. The protein-based platform is mature and ready for collaborative application. Extension of the platform to glycosaminoglycans is estimated at three months, leveraging existing infrastructure.