Prediction of chemical compounds’ toxicity enables efficient and rapid screening at the cost of utilizing experimental data as a foundation for artificial intelligence (AI) models. Given the constraints of limited data availability, few-shot learning techniques, like those based on meta-learning models, may be beneficial. In this study, the ionic liquid (IL) toxicity dataset was employed. Initially, models underwent pretraining on data from related domains before being adapted in a few-shot manner, adhering to the Model-Agnostic Meta-Learning (MAML) algorithm. The MAML algorithm utilized a neural network to predict toxicity based on the descriptors of the molecular structure. MAML enhanced the accuracy of IL toxicity prediction even with limited data, particularly when similar tasks were provided during meta-training, such as between Escherichia coli and Vibrio fischeri. Its effectiveness, however, is dependent on the diversity of the chemical space of the task for which adaptation is performed. As the final part of the study, a new version of MAML (i.e., attentive MAML) was proposed to address the limitations of the basic method. In our Att-MAML modified version of the algorithm, a part of the neural network was dedicated to adapting the importance of latent features with respect to the given test task. This approach allowed a significant reduction in standard deviation and overall improvement of the performance metrics in the low-data regime. This work advances the toxicity prediction in extremely low-data regimes beyond the current state of the art. Incorporation of attentive latent feature scaling (with Att-MAML) handles the limitations inherent to extremely low-data regimes. This novel method allows reductions in the standard deviation of metric values even while adapting on sets as small as four records.
Baran et al. (Thu,) studied this question.