Seminarium Zakładu Biofizyki

This project presents a general computational platform for optimizing drug-like molecules. The system takes a target protein structure, an initial ligand library, and a defined binding site as input, then generates new molecular candidates using a SELFIES-based genetic algorithm. The generated molecules are filtered according to physicochemical and cheminformatics criteria, including molecular weight, lipophilicity, polar surface area, hydrogen bonding capacity, rotatable bonds, ring size, charge, and synthetic accessibility.The filtered candidates are evaluated in an active-learning loop. In the current version, AutoDock Vina is used as the scoring function, while a surrogate machine learning model based on Morgan fingerprints and an Extra Trees regressor predicts the performance of newly generated molecules. The system selects candidates by combining predicted docking quality, prediction uncertainty, and chemical diversity.The goal is to create a flexible "drug optimizer" that supports early stage hit-to-lead optimization by rapidly exploring chemical space and prioritizing promising candidates.