How are deep learning architectures adapted for genomic sequence analysis?

Deep learning architectures for genomics: 1) Convolutional Neural Networks (CNNs) - detect motifs and patterns in DNA/protein sequences; 1D convolutions scan sequences like text; DeepBind, Basenji for regulatory element prediction. 2) Recurrent Networks (LSTMs, GRUs) - capture long-range dependencies in sequences; used in protein language models. 3) Attention/Transformers - self-attention models capture long-range interactions without position constraints; protein language models (ESM, ProtTrans); Enformer for expression prediction spanning 200kb. 4) Graph Neural Networks - represent molecules and protein structures as graphs; capture spatial relationships. 5) Variational Autoencoders - generate novel sequences, protein design. Considerations: one-hot encoding vs embeddings; handling variable-length sequences; interpretability (attention weights, gradient-based attribution, in-silico mutagenesis); data augmentation (reverse complement); transfer learning from pre-trained models.