Structure-Independent Genetic Algorithm-Based Molecular Optimization Using Pocket-Aware Molecular Representations
Abstract
De novo molecular design typically depends on target structures or ligand poses, limiting its applicability when structural information is unavailable. Here, we introduce the pocket-aware molecular representation-guided genetic algorithm (PAMR-GA), a target-structure-independent molecular optimization framework that learns pocket-aware representations from structure–activity relationship (SAR) data using only molecular structures represented as SMILES strings and experimental pIC50 values. By encoding latent spatial constraints inferred from SAR without requiring target structures, the learned representations guide genetic algorithm-based molecular optimization through a Gaussian process surrogate model. Benchmarking on BACE-1 demonstrated that PAMR-GA achieved comparable predictive performance, docking scores, and drug-likeness profiles to existing optimization approaches while generating molecules with substantially greater structural novelty and higher Murcko scaffold diversity. Evaluation on four additional targets, including a kinase, serine proteases, and an NADbinding enzyme, further showed consistent improvements in molecular novelty. These results highlight PAMR-GA as a complementary approach for incorporating SAR-derived spatial constraints into de novo molecular design without requiring target structures. 1 Scientific Contribution PAMR-GA enables structure-independent molecular design by learning binding constraints implicitly from 2D structure–activity relationship data, without requiring protein structures or ligand poses. By introducing a pocket-aware molecular representation that integrates solvent-accessible surface area and partial-charge features, PAMR-GA bridges molecular topology with 3D binding compatibility. Our results demonstrate that constrained generation, rather than improved activity prediction alone, drives structural novelty. Evaluation across five diverse targets from four protein classes highlights the generalizability of PAMR-GA as a broadly applicable framework for target-aware molecular design.
Related
- Task-adaptive multimodal molecular representations for structure-sensitive property prediction
- GLOSS: A Multi-Strategy Sampling Algorithm for Optimization in Vast Chemical Spaces
- A Latent Ability Model for Benchmarking Molecular Representations
- C-MAGE: Automated Extraction and Organization Engine for Molecular Representations in Chemical Documents
- Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
BibTeX
@misc{kumar2026structureindependent,
title = {{Structure-Independent Genetic Algorithm-Based Molecular Optimization Using Pocket-Aware Molecular Representations}},
author = {Abhishek Kumar and Hyun Woo Kim},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010348/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010348/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Structure-Independent Genetic Algorithm-Based Molecular Optimization Using Pocket-Aware Molecular Representations
AU - Abhishek Kumar
AU - Hyun Woo Kim
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010348/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010348/v1
AB - De novo molecular design typically depends on target structures or ligand poses, limiting its applicability when structural information is unavailable. Here, we introduce the pocket-aware molecular representation-guided genetic algorithm (PAMR-GA), a target-structure-independent molecular optimization framework that learns pocket-aware representations from structure–activity relationship (SAR) data using only molecular structures represented as SMILES strings and experimental pIC50 values. By encoding latent spatial constraints inferred from SAR without requiring target structures, the learned representations guide genetic algorithm-based molecular optimization through a Gaussian process surrogate model. Benchmarking on BACE-1 demonstrated that PAMR-GA achieved comparable predictive performance, docking scores, and drug-likeness profiles to existing optimization approaches while generating molecules with substantially greater structural novelty and higher Murcko scaffold diversity. Evaluation on four additional targets, including a kinase, serine proteases, and an NADbinding enzyme, further showed consistent improvements in molecular novelty. These results highlight PAMR-GA as a complementary approach for incorporating SAR-derived spatial constraints into de novo molecular design without requiring target structures. 1 Scientific Contribution PAMR-GA enables structure-independent molecular design by learning binding constraints implicitly from 2D structure–activity relationship data, without requiring protein structures or ligand poses. By introducing a pocket-aware molecular representation that integrates solvent-accessible surface area and partial-charge features, PAMR-GA bridges molecular topology with 3D binding compatibility. Our results demonstrate that constrained generation, rather than improved activity prediction alone, drives structural novelty. Evaluation across five diverse targets from four protein classes highlights the generalizability of PAMR-GA as a broadly applicable framework for target-aware molecular design.
ER -