| 日時 | 2026年07月09日木曜日・16時00分～17時30分 |
|---|---|
| 場所 | 創薬科学研究館2F講義室 |
| 講師 | Prof. Hsiang-Yu (Angie) Wang（National Tsing Hua University, Taiwan/Professor） |
| 連絡先 | 加藤竜司（ kato.ryuji.j7@f.mail.nagoya-u.ac.jp） |
| ファイル | 178228203020260709_ProfAngieWang.pdf |
Abstract：
Ionic liquids (ILs) are highly tunable “designer solvents” with negligible vapor pressure, wide electrochemical windows, and excellent thermal stability, making them valuable for metal recovery, biopolymer dissolution, energy storage, catalysis, and pharmaceutical processing. However, the vast IL chemical space, estimated to exceed 10¹⁸ possible cation–anion combinations, makes conventional trial-and-error screening impractical. Early machine learning (ML) approaches were limited by scarce curated datasets, inadequate molecular featurization, poor interpretability, and weak transferability across chemically distinct IL families. The absence of physically grounded feature selection also restricted model extrapolation and prospective discovery. To address these challenges, we developed an integrated ML and thermodynamic modeling framework combining rigorous feature engineering, interpretable regression, and transfer learning. The framework uses a Mutual Information–Decision Tree (MI-DT) hybrid strategy for feature selection, coupled with Elastic Net regularized Quantitative Structure–Property Relationship (QSPR) modeling, to identify physically meaningful molecular descriptors from heterogeneous datasets. COSMO-RS thermodynamic simulations are incorporated to augment limited experimental data and anchor predictions in thermochemical principles, improving accuracy while revealing mechanistic insights into IL solvation and coordination behavior. For metal extraction, the framework was validated using 497 literature samples involving imidazolium-based ILs for extracting Ni²⁺, Au³⁺, and Cu²⁺. The MI-DT/QSPR models achieved extrapolation R² values of 0.91, 0.95, and 0.89, respectively, demonstrating strong predictive performance and providing molecular guidance for designing selective ILs for hydrometallurgy and environmental remediation. The framework was further extended to fucoidan dissolution, where limited experimental data required transfer learning. A QSPR model trained on cellulose dissolution was fine-tuned for fucoidan using COSMO-RS simulations, enabling virtual screening of over 2,500 ILs. This process identified promising candidates and key structural features governing effective dissolution while preserving bioactivity. Overall, the platform offers a robust, data-efficient, and transferable strategy for accelerating IL discovery across diverse applications.
Ionic liquids (ILs) are highly tunable “designer solvents” with negligible vapor pressure, wide electrochemical windows, and excellent thermal stability, making them valuable for metal recovery, biopolymer dissolution, energy storage, catalysis, and pharmaceutical processing. However, the vast IL chemical space, estimated to exceed 10¹⁸ possible cation–anion combinations, makes conventional trial-and-error screening impractical. Early machine learning (ML) approaches were limited by scarce curated datasets, inadequate molecular featurization, poor interpretability, and weak transferability across chemically distinct IL families. The absence of physically grounded feature selection also restricted model extrapolation and prospective discovery. To address these challenges, we developed an integrated ML and thermodynamic modeling framework combining rigorous feature engineering, interpretable regression, and transfer learning. The framework uses a Mutual Information–Decision Tree (MI-DT) hybrid strategy for feature selection, coupled with Elastic Net regularized Quantitative Structure–Property Relationship (QSPR) modeling, to identify physically meaningful molecular descriptors from heterogeneous datasets. COSMO-RS thermodynamic simulations are incorporated to augment limited experimental data and anchor predictions in thermochemical principles, improving accuracy while revealing mechanistic insights into IL solvation and coordination behavior. For metal extraction, the framework was validated using 497 literature samples involving imidazolium-based ILs for extracting Ni²⁺, Au³⁺, and Cu²⁺. The MI-DT/QSPR models achieved extrapolation R² values of 0.91, 0.95, and 0.89, respectively, demonstrating strong predictive performance and providing molecular guidance for designing selective ILs for hydrometallurgy and environmental remediation. The framework was further extended to fucoidan dissolution, where limited experimental data required transfer learning. A QSPR model trained on cellulose dissolution was fine-tuned for fucoidan using COSMO-RS simulations, enabling virtual screening of over 2,500 ILs. This process identified promising candidates and key structural features governing effective dissolution while preserving bioactivity. Overall, the platform offers a robust, data-efficient, and transferable strategy for accelerating IL discovery across diverse applications.
