BIO-Distiller: Boosting Supervised Baselines by Distilling Biological Foundation Models

The success of foundation models has driven their application to biological data, including RNA, DNA, and protein sequences. These are expected to adapt to new tasks, while also providing novel insights into molecular function and biological mechanisms underlying health and disease. Despite recent advances, benchmarks show that biological foundation models fail to consistently outperform simpler supervised approaches. Moreover, key challenges remain, such as the development of faithful interpretation methods and the integration of multiple modalities. Here, we propose BIO-Distiller, a framework to distill the rich information from biological foundation models into smaller models. In our benchmark on six RNA downstream tasks, we first show that well-tuned supervised baselines can still outperform foundation models. Furthermore, knowledge distillation consistently boosts the baselines’ performance by up to 10% across all tasks. Additionally, our framework is capable of integrating multiple foundation models, whether from the same modality by exploiting design and pre-training differences, or across different modalities. The results not only confirm the potential of BIO-Distiller but also provide guidelines for its application to new tasks and modalities, paving the way toward high-performing, efficient, and easily interpretable supervised models for biology.