Abstract Deep learning foundation models are becoming increasingly popular for use in bioactivity prediction. Recently, Feng et al. developed ActFound, a bioactive foundation model that jointly uses pairwise learning and meta-learning. By utilizing these techniques, the model is capable of being fine-tuned to a more specific bioactivity task with only a small amount of new data. Here, to investigate the generalizability of the model, we looked to fine-tune the foundation model on an antibacterial natural products (NPs) dataset. Large, labelled NPs datasets, which are needed to train traditional deep learning methods, are scarce. Therefore, the bioactivity prediction of NPs is an ideal task for foundation models. We studied the performance of ActFound on the NPs dataset using a range of few-shot settings. Additionally, we compared ActFound’s performance with those of other state-of-the-art models in the field. We found ActFound was unable to reach the same level of accuracy on the antibacterial NPs dataset as it did on other cross-domain tasks reported in the original publication. However, ActFound displayed comparable or better performance compared to the other models studied, especially at the low-shot settings. Our results establish ActFound as a useful foundation model for the bioactivity prediction of tasks with limited data, particularly for datasets that contain the bioactivities of similar compounds.
Butt et al. (Thu,) studied this question.