Protein Language Model Based Structure-guided Antibody Screening for Disordered Protein Targets

A crucial step in the pathogenesis of Parkinson’s disease involves cell-to-cell transmission of α -Synuclein proto-fibrils via endocytosis, driven primarily by the interaction of its disordered C-terminal peptide with domain 1 of Lymphocyte Activation Gene 3 (LAG3) neuronal receptors. High-affinity antibodies have been proposed as therapeutic modalities to delay this progression and subsequent amyloid formation. In our work, we develop an end-to-end computational pipeline to enable rapid screening of antibody sequences that have a high-affinity for the disordered C-terminal peptide of α -Synuclein using no information of known binders. This de novo screening was enabled by a structural bioinformatics based in silico data generation pipeline combined with a deep learning framework. Our simple feed forward network model built upon sequence embeddings from a protein language model ranked the binding affinities (ΔG) of antibodies to α -Synuclein with a high accuracy (Spearman ρ = 0.86) when the training and the evaluation datasets contained sequences having some overlap in the complementarity determining regions (CDRs). However, for vastly different CDR sequences, a transformer encoder model trained using the antibody sequence embeddings showed a low Spearman rank correlation of ρ = 0.18. The models have a mean Precision@100 of 38 and 12 respectively, significantly outperforming a random process. Overall, our work demonstrates a computational protocol for generating a high quality dataset of antibody-antigen complexes spanning a very large diversity in antibody sequences followed by training of a deep learning model for prediction of high-affinity antibody sequences for a specific protein target with no known binders.