DF-S4: Disentangled FiLM-Conditioned Molecular Generation Using the S4 Architecture

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence in Medicine

Background:

• De novo molecular design faces challenges in balancing biological affinity with drug-likeness.
• Existing generative models often lack controllability and structural diversity.
• Controlling molecular generation for specific targets and properties remains difficult.

Purpose of the Study:

• To introduce DF-S4, a novel conditional molecular generation framework.
• To enable fine-grained and stable multi-objective control in de novo molecular design.
• To address limitations of existing generative models in controllability and diversity.

Main Methods:

• Utilizing Structured State Space Models (S4) for molecular generation.
• Implementing a disentangled latent representation and hierarchical feature-wise linear modulation (FiLM).
• Decoupling structural and property variables with multi-level conditional signal injection.

Main Results:

• DF-S4 achieved robust target-steering performance across EGFR, BRAF, and FGFR1 kinase targets (74.2-83.9% active ratios).
• Generated molecules maintained high novelty (>95%) and competitive internal diversity (∼0.85).
• DF-S4 improved the Pareto Frontier for simultaneous high affinity and drug-likeness, surpassing prior methods.

Conclusions:

• DF-S4 offers a powerful framework for de novo molecular design with enhanced control.
• Latent disentanglement and hierarchical FiLM modulation are crucial for generative stability and property alignment.
• The framework generates physically plausible candidates with superior binding affinities compared to reference inhibitors.