Therapeutic Antibody Design Using High-Capacity Machine Learning
MIT EECS, Biological Engineering, and CSAIL
We introduce a family of machine learning methods for designing the complementarity-determining regions of human antibodies that allows for the simultaneous optimization of antibody affinity and target specificity. We performed phage-panning experiments to generate large training data sets that characterize the binding of ~10^8 antibody sequences to five target molecules. We used these data to train machine learning methods to predict target binding for previously unseen antibody sequences. We generalized these models to synthesize novel antibody sequences that are optimized for affinity and specificity objectives. A total of 77,596 machine learning proposed sequences were combined with controls and fabricated as 104,525 distinct oligonucleotides, cloned into antibody frameworks, and tested in the context of competitive sequences. We find that our synthetic methods create antibodies with a tenfold increase in affinity when compared with the best sequences previously observed. This project is a collaboration between MIT and the Novartis Institutes for BioMedical Research.