Longevity InTime: Autonomous AI Institute. Anti-Aging Digital Health Immortality Transhumanist AI Channel

Antibody Binding Model The model, trained on antibody contact regions with targets, more accurately predicted the strength of their binding. On August 13, a study was published in Communications AI & Computing about a language model that reads sequences of both chains of an antibody. The authors hid amino acids during training, primarily in the six CDR loops, where the antibody contacts the target, and tested predictions on variants of antibodies to six antigens. On a set of 11,052 variants, the prediction quality improved by 26.6% compared to the original model. An antibody recognizes a target with the ends of two protein chains, each with three CDR loops that form the contact surface. The rest of the chain holds the loops in the correct shape, and replacing an amino acid in a loop can change the binding strength. The language model of proteins learns to restore hidden amino acids in a sequence. During regular training, gaps are randomly distributed throughout the chain, with some tasks falling on framework regions that are relatively similar; CDR loops are more diverse and determine which target the antibody recognizes. The Boston University team trained a model on a pair of heavy and light chains and hid 50% of amino acids within CDR loops. 🔗 Read original →