Molecular representation learning and generative chemistry have emerged as transformative approaches in computational drug discovery, enabling the design, optimization and synthesis planning of molecules through data-driven methods. Traditional enumeration approaches cannot cope with the combinatorial complexity of chemistry, necessitating more sophisticated methods for navigating vast chemical spaces. Recent advances in deep generative models - particularly diffusion models and flow matching - have demonstrated remarkable capabilities in learning complex molecular distributions and generating novel chemical structures with desired properties.

These generative frameworks become particularly powerful when combined with conditional sampling strategies, allowing precise control over molecular generation through various constraints such as pharmacophoric features, binding pockets, or structural fragments. By conditioning on specific molecular properties or structural motifs, these models can systematically explore chemical space while satisfying design requirements, supporting applications from fragment elaboration to scaffold hopping.

Perhaps most intriguingly, conditional generative models provide elegant solutions to challenging inverse problems in drug discovery- scenarios where we must infer molecular structures from indirect observations. We present two complementary applications: First, MSFlow addresses mass spectrometry-based structure elucidation, employing flow matching to translate spectral data into molecular structures. Second, Flowr.root tackles pocket-aware ligand design with joint affinity prediction, using SE(3)-equivariant flow matching for 3D molecular generation conditioned on protein binding sites.

Through parameter-efficient domain adaptation and inference-time importance sampling, these frameworks enable steering molecular design toward higher-affinity compounds while maintaining computational efficiency. Validation studies across kinase inhibitors and fragment-based design demonstrate that unified generative approaches can address the full spectrum of structure-based drug discovery, from molecular identification through lead optimization, establishing a comprehensive foundation for AI-driven chemistry.

Djork-Arné Clevert

Pfizer Inc

Appearances