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

Abstract

There exist a wide range of Neural Networks (NNs) that have modelled genomic sequences well. The architectures of these NNs consist of three steps: k-mer embedding, sequence embedding, and output mapping. The NNs are primarily differentiated from one another by their sequence embedding components. Although different NNs have been compared before, these comparisons are often either limited or involve drastically different NNs. Another issue is that some NNs are configured for a specific sequence length, and different length sequences require manual modification of the architectures. This study examines the architectures of established NNs for genomics sequences. Based on these architectures, a novel NN called Mimic is developed that provides a unified architecture for the principled comparison of NN components. Mimic was used to model two different types of genomic datasets: protein-DNA interactions and enhancer activity. Along with other architecture components, different sequence embedding strategies were tried such as recurrent layers and convolutional layers. It was found that there exist specific NN components that yield significantly better performing models. Furthermore, Mimic significantly outperformed prevailing standard NNs, Bichrom and DeepSTARR, in modelling protein-DNA interactions and enhancer activity, respectively. This study provides a framework for more extensive studies toward a general prescription of NN architecture for modelling genomics data.