FLOWR.ROOT - A flow matching-based foundation model for joint multi-purpose structure-aware 3D ligand generation and...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• De novo drug design requires accurate prediction of ligand-target interactions.
• Existing methods often lack the ability to unify ligand generation and binding affinity prediction.
• The development of foundation models offers a promising avenue for integrated molecular design.

Purpose of the Study:

• To introduce FLOWR.ROOT, a unified SE(3)-equivariant flow-matching foundation model.
• To enable pocket-aware 3D ligand generation and multi-endpoint binding affinity prediction.
• To provide pLDDT-based confidence estimation within a single framework.

Main Methods:

• Utilizing SE(3)-equivariant flow matching for molecular generation.
• Implementing a mixed isotropic-anisotropic prior placement strategy for diverse sampling.
• Employing a three-stage training process: large-scale pre-training, refinement on co-crystal data, and LoRA finetuning.
• Integrating joint structure-affinity modeling for inference-time guidance.

Main Results:

• FLOWR.ROOT supports various generative tasks: pocket-conditional generation, interaction- and pharmacophore-conditional sampling, scaffold hopping, fragment growing/replacement.
• The model achieves accurate binding affinity prediction (pIC50, pKi, pKd, pEC50) and confidence estimation (pLDDT).
• Joint structure-affinity modeling allows for inference-time importance-sampling guidance without external scoring functions.

Conclusions:

• FLOWR.ROOT successfully unifies 3D ligand generation and binding affinity prediction in a single foundation model.
• The model demonstrates versatility across multiple drug design tasks, from hit identification to lead optimization.
• Case studies on kinase selectivity and scaffold elaboration highlight the practical utility of FLOWR.ROOT in accelerating drug discovery pipelines.