Speaker
Description
Innovation:
Utilizes Fourier Neural Operators (FNOs) to model continuous 3D physical landscapes (hydrophilic/hydrophobic surfaces).
Embeds physical properties directly into the molecular generation process, enhancing physical realism and relevance.
Methodology:
Evolution from LSTM-based generators [2] to GPT-based transformers to current dual-FNO-enhanced PINO-GPT architectures.
FNO layers process spatial physical fields, enabling generation of molecules optimized for orthosteric binding pockets.
Fine-tuning enables adaptation to experimental constraints in GPCRs (e.g., agonist activity).
Results:
Outperforms traditional sequence-based and diffusion-based models [3] in structure-informed drug generation tasks.
Experimental validation (binding and functional assays) confirms predictive performance and real-world applicability.
Impact:
Bridges computational design and wet-lab validation.
Enables more interpretable, physically grounded, and experimentally aligned GenAI workflows in drug discovery.
