SpectralMol: Training-free multi-objective molecular design using Fourier evolution
Evolutionary algorithm uses Fourier coefficients to generate diverse drug-like molecules without any pre-training.
A new paper from arXiv introduces SpectralMol, an evolutionary algorithm that generates novel molecules by representing chemical structures as compact matrices of Fourier coefficients. Unlike deep learning models that require extensive pre-training and suffer from statistical biases, SpectralMol uses frequency-controlled evolutionary dynamics. It projects these coefficients onto a fixed basis to create position-wise latent vectors, which are then decoded into valid molecular strings via SELFIES. The NSGA-II algorithm enforces diversity and handles multiple objective functions separately—rather than collapsing them into a single scalar reward—enabling true multi-objective optimization.
In standardized benchmarks, SpectralMol achieved comparable aggregate performance to state-of-the-art methods, with a task-dependent profile that excelled on several multi-parameter optimization tasks. An ablation study confirmed the advantages of the structured latent matrix. On a realistic ClpP-targeted drug discovery benchmark, SpectralMol was compared to a reinforcement-learning-based model under a fixed oracle-call budget. It generated more docking hits and more diverse scaffolds while maintaining competitive physicochemical properties. The Fourier representation naturally separates scaffold-level changes (low-frequency perturbations) from localized substructure variations (high-frequency perturbations), providing interpretability and efficiency in a training-free framework.
- SpectralMol encodes molecules as Fourier coefficient matrices, decoded via SELFIES, eliminating the need for pre-training.
- Uses NSGA-II for true multi-objective optimization; outperformed RL models on ClpP-targeted drug discovery with more docking hits and diverse scaffolds.
- Frequency control cleanly separates scaffold modifications from local substructure changes, enhancing interpretability.
Why It Matters
A training-free, interpretable approach to multi-objective drug discovery that could accelerate hit finding without expensive deep learning.