Projects

← Back to Home

FiDES: Adverse Event Analysis via Interpretable ML (KAN Approach)

• Overview: Focusing on AI-driven scientific discovery, I am researching ways to enhance the interpretability of chemical structures contributing to adverse events. As one approach, we utilize Kolmogorov-Arnold Networks (KAN).
• Addressing Technical Trade-offs: While KAN offers high function approximation capabilities, it faces challenges like computational cost spikes and overfitting in high-dimensional settings (e.g., ECFP). This project investigates overcoming the scalability-interpretability trade-off through specialized regularization and architectural optimizations.
• Status: Presented at the PSJ 146th Annual Meeting, the 9th JSPHCS Freshers Conference, and the 53rd Annual Meeting of the Japanese Society of Toxicology (JSOT).
• Publication: Preprint available: Preprints.org (FMES)
• Resources: ecfp_cli ｜ Technical Review

EUOS25 challenge: Optical Property Prediction

• Overview: Winner of the Fluorescence track in a large-scale competition involving ~100k compounds.
• Technical Approach:
  • Massive Informatics: Expanded PathCounts (up to 50th order) beyond 1,800+ Mordred descriptors. Integrated Conjugation Features to evaluate π-conjugated systems, constructing a specialized feature space for fluorescence prediction.
  • GNN Feature Generation: Integrated graph-based embeddings using a custom GINE-Net.
  • Sequential Stacking: Chained architecture modeling biophysical dependencies from transmittance to fluorescence.
• Result: Winner of the Fluorescence track (February 2026)
• Publication: Preprint available: ChemRxiv
• Tools & Links: mordred_descriptor_calculator ｜ Official News ｜ Challenge Info ｜ Technical Review