Pre-training with pseudo-labeling compares favorably with large language models for regulatory sequence prediction

Predicting molecular processes using deep learning is a promising approach to provide biological insights for non-coding SNPs identified in genome-wide association studies. However, most deep learning methods rely on supervised learning, which requires DNA sequences associated with functional data, and whose amount is severely limited by the finite size of the human genome. Conversely, the amount of mammalian DNA sequences is growing exponentially due to ongoing large-scale sequencing projects, but in most cases without functional data. To alleviate the limitations of supervised learning, we propose a novel semi-supervised learning based on pseudo-labeling, which allows to exploit unlabeled DNA sequences from numerous genomes during model pre-training. The approach is very flexible and can be used to train any neural architecture including state-of-the-art models, and shows in certain situations strong predictive performance improvements compared to standard supervised learning in most cases. Moreover, small models trained by SSL showed similar or better performance than large language model DNABERT2.