Unveiling key descriptors via machine learning: toward rational molecular design of chromophores with excited-state intramolecular proton transfer

Abstract

Precise design of excited-state intramolecular proton transfer (ESIPT) molecules targeting advanced optoelectronic or biological sensing applications presents a fundamental challenge. Controlling the energy difference (ΔE*) between normal (N*) and tautomeric (T*) excited-state forms is crucial, yet the complex interplay of hydrogen bond (H-bond) strength, proton donor acidity, and proton acceptor basicity with ΔE* remains insufficiently explored. Conventional trial-and-error approaches for designing tailored ESIPT compounds suffer from inefficient synthesis. To address this, we constructed a high-quality ESIPT dataset by introducing ten substituents with progressively increasing electron-donating capacity into six representative ESIPT parent scaffolds. Integrating qualitative descriptors with data-driven machine learning (ML) enabled precise ΔE* prediction, significantly accelerating high-throughput screening. An interpretable Shapley additive explanations (SHAP)-based ML approach was applied to evaluate the relative importance of key H-bond descriptors while achieving accurate ΔE* prediction. Novel ESIPT candidates were generated using a variational autoencoder (VAE) model and filtered using predicted ΔE*, synthetic accessibility (SA) scores, and pharmacokinetic properties. Critically, we synthesized two AI-designed ESIPT molecules exhibiting distinct N*/T* dual emission, which provides a closed-loop experimental validation of this data-driven molecular design strategy. This work establishes a predictive framework for accurate ΔE* determination and accelerated exploitation of novel promising ESIPT compounds.