A hybrid metaheuristic algorithm for efficient de novo drug design using SELFIES and adaptive search strategies
- 1Institute of Statistical Science, Academia Sinica
Abstract
The discovery of drug-like molecules is a fundamental challenge in computational chemistry because the chemical space contains an enormous number of possible molecular structures. Although deep learning has achieved remarkable success in de novo molecular design, its reliance on large training datasets and high computational costs limits its efficiency and practical applicability. This study aims to develop an efficient training-free framework for de novo drug design by integrating swarm intelligence, evolutionary optimization, and the robust SELF-referencIng Embedded Strings (SELFIES) molecular representation. In specific, we propose SIB-SOMO-SELFIES, a hybrid metaheuristic algorithm that combines the “guided-by-the-best” search strategy of Particle Swarm Optimization with the crossover and mutation mechanisms of Genetic Algorithms. The proposed framework enhances molecular optimization through SELFIES-aware mutation operators, an improved MIX crossover strategy, adaptive probability updating, similarity penalties, and complementary VARY and Random Jump operations to balance exploration and exploitation. Performance was evaluated through computational molecular optimization simulations using the Quantitative Estimation of Drug-likeness (QED) as the optimization objective, and benchmarked against representative SELFIES-based methods including SIB-SOMO, GA-D, SELFIES-VAE, SELFIES-LSTM-HC, SELFIES-REINVENT, and Pasithea. Simulation results demonstrate that the proposed framework efficiently explores the chemical space while maintaining rapid convergence and high optimization stability. The proposed algorithm was evaluated through computational molecular optimization experiments using QED as the objective function. The results show that the hybrid PSO-GA framework efficiently explores the molecular search space and consistently identifies molecular candidates with high predicted drug-likeness as assessed by QED. The proposed method generates molecules with QED scores exceeding 0.948 while maintaining computational efficiency. These findings support the computational potential of the proposed hybrid metaheuristic framework for de novo molecular design without requiring large-scale model training; however, the generated candidates have not yet been validated through experimental synthesis or biological assays.
Related
- Cross-cultural modeling of body satisfaction using social media selfies and self-descriptions
- Higher-order adaptive behaviors outperform pairwise strategies in mitigating contagion dynamics
- Evaluating Multiple Treatments and Strategies for Sepsis in Adaptive Platform (EURASSIA) trial in Japan: a core protocol.
- Algorithm‐Defined Muscle Dysmorphia Estimates Across Weighting and Case‐Definition Strategies in a Gender‐Balanced Adult Online Sample
- Dynamic monitoring and adaptive algorithm optimization of fatigue recovery in track and field athletes based on inertial measurement unit sensor data.
BibTeX
@article{feng2026hybrid,
title = {{A hybrid metaheuristic algorithm for efficient de novo drug design using SELFIES and adaptive search strategies}},
author = {Shen-Ching Feng and Frederick Kin Hing Phoa and Hsin-Ping Liu},
journal = {Scientific Reports},
year = {2026},
month = sep,
doi = {10.1038/s41598-026-69048-7},
url = {https://doi.org/10.1038/s41598-026-69048-7},
note = {CC BY},
}
RIS
TY - JOUR
TI - A hybrid metaheuristic algorithm for efficient de novo drug design using SELFIES and adaptive search strategies
AU - Shen-Ching Feng
AU - Frederick Kin Hing Phoa
AU - Hsin-Ping Liu
PY - 2026
DA - 2026/09/01
JO - Scientific Reports
DO - 10.1038/s41598-026-69048-7
UR - https://doi.org/10.1038/s41598-026-69048-7
AB - The discovery of drug-like molecules is a fundamental challenge in computational chemistry because the chemical space contains an enormous number of possible molecular structures. Although deep learning has achieved remarkable success in de novo molecular design, its reliance on large training datasets and high computational costs limits its efficiency and practical applicability. This study aims to develop an efficient training-free framework for de novo drug design by integrating swarm intelligence, evolutionary optimization, and the robust SELF-referencIng Embedded Strings (SELFIES) molecular representation. In specific, we propose SIB-SOMO-SELFIES, a hybrid metaheuristic algorithm that combines the “guided-by-the-best” search strategy of Particle Swarm Optimization with the crossover and mutation mechanisms of Genetic Algorithms. The proposed framework enhances molecular optimization through SELFIES-aware mutation operators, an improved MIX crossover strategy, adaptive probability updating, similarity penalties, and complementary VARY and Random Jump operations to balance exploration and exploitation. Performance was evaluated through computational molecular optimization simulations using the Quantitative Estimation of Drug-likeness (QED) as the optimization objective, and benchmarked against representative SELFIES-based methods including SIB-SOMO, GA-D, SELFIES-VAE, SELFIES-LSTM-HC, SELFIES-REINVENT, and Pasithea. Simulation results demonstrate that the proposed framework efficiently explores the chemical space while maintaining rapid convergence and high optimization stability. The proposed algorithm was evaluated through computational molecular optimization experiments using QED as the objective function. The results show that the hybrid PSO-GA framework efficiently explores the molecular search space and consistently identifies molecular candidates with high predicted drug-likeness as assessed by QED. The proposed method generates molecules with QED scores exceeding 0.948 while maintaining computational efficiency. These findings support the computational potential of the proposed hybrid metaheuristic framework for de novo molecular design without requiring large-scale model training; however, the generated candidates have not yet been validated through experimental synthesis or biological assays.
SN - 2045-2322
ER -