Multi-Task Learning with Loop Specific Attention for CDR Structure Prediction
This work addresses a domain-specific problem in antibody engineering by improving CDR structure prediction, though it appears incremental as it builds on existing multi-task learning approaches.
The paper tackles the challenge of predicting the Complementarity Determining Region (CDR) structure, particularly the difficult H3 loop in antibody engineering, by proposing a multi-task learning model with loop-specific attention (MLSA) that reduces H3 loop prediction error by at least 19% compared to baselines.
The Complementarity Determining Region (CDR) structure prediction of loops in antibody engineering has gained a lot of attraction by researchers. When designing antibodies, a main challenge is to predict the CDR structure of the H3 loop. Compared with the other CDR loops, that is the H1 and H2 loops, the CDR structure of the H3 loop is more challenging due to its varying length and flexible structure. In this paper, we propose a Multi-task learning model with Loop Specific Attention, namely MLSA. In particular, to the best of our knowledge we are the first to jointly learn the three CDR loops, via a novel multi-task learning strategy. In addition, to account for the structural and functional similarities and differences of the three CDR loops, we propose a loop specific attention mechanism to control the influence of each CDR loop on the training of MLSA. Our experimental evaluation on widely used benchmark data shows that the proposed MLSA method significantly reduces the prediction error of the CDR structure of the H3 loop, by at least 19%, when compared with other baseline strategies. Finally, for reproduction purposes we make the implementation of MLSA publicly available at https://anonymous.4open.science/r/MLSA-2442/.