Mimic: A Unified Neural Network Architecture for Modelling Genomic Sequences and Module-level Comparison

Mimic, a unified neural network architecture for genomic sequences, significantly outperformed Bichrom and DeepSTARR in modelling protein-DNA interactions and enhancer activity, respectively.

• Why it matters: Existing genomics NNs differ mainly in sequence-embedding components, but prior comparisons are limited or use drastically different architectures, and many models require manual changes for different sequence lengths.
• What they did: The study analyzed established genomics NN architectures and built Mimic with a unified k-mer embedding, sequence embedding, and output mapping, testing recurrent and convolutional sequence-embedding strategies on protein-DNA interactions and enhancer activity datasets.
• The result: Specific sequence-embedding components produced significantly better models, and Mimic significantly outperformed Bichrom for protein-DNA interactions and DeepSTARR for enhancer activity, enabling more principled component-level architecture comparisons.