Exploring Chemoinformatics Aspects of Few-Shot Meta-Learning by Example of an Infinite Dilution Activity Coefficient in Ionic Liquid Prediction

Collection of substantial training sets for structure-property modeling often poses a significant challenge, especially for niche sorption systems. This study provides computational experiments with meta-learning that presents a compelling solution to the data scarcity challenge in molecular property prediction. Our analysis utilized infinite dilution activity coefficients for several ionic liquid-solute systems, with coefficient prediction for systems with a particular solute constituting a task. The systems were modeled predominantly using model-agnostic meta-learning (MAML), supported by a study on Reptile and its modified variants. The obtained results provide promising insights into training MAML models by expanding the adaptation set size. Metrics such as R2, RMSE, and MAE indicate comparable performance to graph neural networks, even when trained on only 64 or 128 data points. The versatility of the fine-tuned models suggests that, in certain cases, performance on a single task may be achieved at the expense of reduced model versatility (catastrophic forgetting). The similarity between the test and training tasks (as approximated by Tanimoto similarity of the solutes' molecules) was identified as a factor affecting performance on the test task. Consequently, Task Similarity-Aware Reptile (TSA-Reptile) was proposed to target those dissimilar tasks. This novel method scales the loss function by similarity to the closest training task. It was shown to outperform MAML on out-of-distribution tasks. Beyond providing the comparative analysis of meta-learning and traditional deep learning, potential strengths of both MAML and TSA-Reptile are discussed.