Table of Contents

SpectralMol: Multi-Objective Molecular Generation with Frequency-Controlled Evolutionary Dynamics

Training-free molecular design with Fourier-coefficient genotypes, SELFIES decoding, and NSGA-II

KVL Staff on Project

Didier Barradas-Bautista
didier.barradasbuatista@kaust.edu.sa
Building 1, Level 0, Office 0125

Collaborators

Elia Colleoni
elia.colleoni@kaust.edu.sa

Paolo Guida (KAUST RS)
paolo.guida@kaust.edu.sa

William L. Roberts (KAUST PI)
william.roberts@kaust.edu.sa

Research Breakthrough

SpectralMol is a training-free evolutionary framework for multi-objective molecular generation. It represents molecules as a matrix of Fourier coefficients, projects them through a fixed basis, and decodes the resulting latent sequence with SELFIES-constrained token selection.

  • Uses NSGA-II to optimize multiple chemical objectives without collapsing them into a single scalar reward.
  • Separates global scaffold changes from local substructure edits through frequency-controlled latent dynamics.
  • Avoids neural network pre-training while remaining interpretable and compatible with non-differentiable oracles.

🔬 Research Visualization The project is built around a Fourier-parameterized latent space, where low-frequency coefficients induce broad scaffold-level changes and high-frequency coefficients refine local substructures. The underlying manuscript is available as Preprint.

🔬 Research Summary The method is designed for compact, structured exploration of chemical space. Low-frequency Fourier modes drive coherent global changes across the sequence, while high-frequency modes produce more localized variation. This makes the search trajectory more interpretable than an unstructured latent vector and better suited to scaffold hopping and lead optimization.

🎯 Key Contributions The SpectralMol pipeline combines four ingredients:

  • A fixed truncated Fourier basis for position-wise latent construction.
  • A structured token embedding space for distance-based decoding.
  • Constraint-aware SELFIES masking to reduce invalid generations.
  • Multi-objective optimization with NSGA-II over Fourier coefficients.

📌 Highlights

  • Training-free molecular optimization
  • Explicit and interpretable latent structure
  • Pareto-based multi-objective search
  • Compatibility with docking and other non-differentiable scoring functions

🧪 Why It Matters SpectralMol provides a practical route to explore molecular design spaces without a learned generator. That makes the approach attractive when data are limited, when objectives are expensive, or when the search needs to remain interpretable for downstream medicinal chemistry decisions.

,