From Latent Manifolds to Functional Probes: An Interpretable, Kinome-Scale Generative Machine Learning Framework for...

Area of Science:

• Medicinal Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Designing selective kinase inhibitors is challenging due to conserved ATP-binding sites.
• Generative AI models offer rapid chemical space exploration but often lack interpretability.

Purpose of the Study:

• To develop a modular, interpretable generative framework for *de novo* SRC kinase inhibitor design.
• To address bottlenecks in AI-driven drug design, focusing on interpretability and chemical rationale.

Main Methods:

• Integration of ChemVAE latent space modeling, a Kinase Inhibition Likelihood (KIL) scoring function, and Bayesian optimization.
• Utilized cluster-guided local neighborhood sampling for scaffold transformation and exploration of chemical space.
• Employed a modular framework combining interpretable machine learning with generative AI.

Main Results:

• Kinase inhibitors form a coherent manifold in latent space, with SRC as a 'hub' for scaffold transformation.
• Successfully converted LCK inhibitors into novel SRC-like chemotypes, with LCK-derived molecules comprising ~40% of outputs.
• Identified a representation gap: SMILES-based generation struggles with multi-ring pharmacophores crucial for clinical kinase inhibitors.

Conclusions:

• A hybrid approach combining interpretable ML and generative AI is crucial for transparent and effective drug design.
• Topology-aware representations are needed to overcome limitations in current generative models for complex pharmacophores.
• Unbiased, cluster-guided sampling is more effective than active-biased optimization for navigating chemical space in inhibitor design.