Adverse Event Analysis via Kolmogorov-Arnold Networks (KAN)
Summary: Utilizing KAN to enhance the interpretability of chemical structures contributing to adverse events.
Addressing Technical Trade-offs:
While KAN provides superior function approximation, it faces well-documented challenges such as computational overhead and susceptibility to overfitting when applied to high-dimensional datasets (e.g., ECFP). This project investigates overcoming the scalability-interpretability trade-off through specialized regularization techniques and architectural optimizations.
Status: Presented at the PSJ Annual Meeting; [Manuscript in preparation].
EUOS25 challenge: Optical Property Prediction
Summary: A machine learning competition for predicting optical properties of ~100k compounds. The challenge consisted of two tracks: Transmittance and Fluorescence prediction. Our team won the Fluorescence track.
Technical Approach:
Multimodal Strategy: Integrating high-precision quantum chemistry (MOPAC/MACE-xTB) and 1,800+ descriptors. Specifically, integrated specialized Conjugation Features to evaluate the quality of $\pi$-conjugation systems, constructing a feature space specialized for fluorescence prediction.
Sequential Stacking: Developed a chaining architecture that models the biophysical dependency from transmittance to fluorescence.
Result: Winner of the Fluorescence track (February 2026)
Outlook: Technical details and code to be submitted to “SLAS Technology”. [Status: In preparation]