[2S00m] 01. ASBA - Part 4

10:45 AM - 11:00 AM JST(1:45 AM - 2:00 AM UTC)

[2ASBA-04] Deep Learning Algorithms Using Transformer Attention for Biopart Analysis in Synthetic Biology

○Donghyuk Kim¹ (1. Korea Advanced Institute of Science and Technology (Korea))

Keywords:

Machine Learning,Synthetic Biology,Enzyme,Transcription Regulation

This research introduces an explainable AI framework based on the transformer architecture to analyze and design critical biological parts in synthetic biology. The study first presents DISCODE, a deep learning-based method that achieves a 97.4% accuracy in classifying NAD and NADP-dependent binding proteins by leveraging ESM-2 language model-based embeddings and multi-head self-attention mechanisms. Beyond simple classification, DISCODE utilizes attention maps to identify specific protein residues responsible for cofactor preference, which enables the design of enzyme mutants for cofactor switching through an iterative design pipeline. Additionally, the framework is applied to predict condition-specific transcription factor binding for the global regulator CRP in Escherichia coli. By employing a custom tokenizer and positional encoding, the model accurately predicts binding intensities across different environmental conditions and various E. coli strains, demonstrating robust generalization and high correlation with experimental ChIP-exo data. These results highlight the potential of transformer-based attention mechanisms to reveal complex sequence correlations and provide actionable evidence for the future of genomic analysis and synthetic biopart design.