Content area
Full Text
Deep learning has been powerful in learning complex functions from data and has been applied in computer vision, natural language processing and biology. If we view the human genome as a book with three billion letters of nucleotides represented by A, C, G and T, genes and gene-controlling sequences are encoded in the book and variations in the genome can link to disease conditions. Neural network models that extract patterns from the sequences can help predict functional genomic elements and interpret genetic variations. However, the current deep learning models for genomics usually involve expert-designed neural network structures and require extensive tuning, making such models unapproachable for most other scientists. In a recent publication in Nature Machine Intelligence, Zhang and colleagues1 present a framework called Automated Modelling for Biological Evidence-based Research (AMBER), which incorporates a deep learning architecture searching algorithm and demonstrates efficient and automatic model selection for genomic problems.
Recent work that applies deep learning to biological sequences has improved our understanding of the human genome. Examples include clinical impact inference for protein-coding mutations2, pattern recognition among sequences bound by transcription factors (TF)3, and epigenetic effect prediction for genetic variations4. Specifically, as epigenetic profiles of a genomic region reflect biological functions, deep learning models that predict epigenetic profiles from genomic sequences have been powerful in extracting patterns in functional biological sequences (Fig. 1a). Convolutional neural networks (CNNs) are well suited for this task due to their advantage in extracting spatial patterns in sequences. For instance, Zhou et al.4 constructed a CNN-based model called DeepSEA to map genomic sequences to epigenetic profiles. The trained model can thus predict genetic variants with significant epigenetic effects. Built on the previous success with the expert-designed CNN in DeepSEA, Zhang et al. expanded the...