The Complementarity-Determining Region 3 (CDR3) is the most variable region of an antibody and the primary determinant of antigen specificity. Located at the center of the antigen-binding site, CDR3 directly contacts the target antigen and determines binding affinity.
CDR3 sequences are highly diverse due to V(D)J recombination and somatic hypermutation, making them ideal targets for rational antibody design. However, this diversity also makes it challenging to predict which CDR3 sequences will bind to a specific antigen.
Traditional antibody discovery relies on immunization, hybridoma technology, or phage display—processes that are time-consuming, expensive, and often yield suboptimal candidates. Computational approaches can accelerate discovery, but predicting antigen-specific CDR3 sequences remains difficult due to:
- The vast sequence space (20n possible sequences for length n)
- Complex structure-function relationships
- Limited training data for machine learning models
- The need for both high affinity and developability
ABYSS uses a machine learning-guided optimization pipeline to generate antigen-specific CDR3 sequences de novo. The system combines:
ESM-2 Protein Language Model
Leverages evolutionary information from millions of protein sequences to generate biologically plausible embeddings of antigen-CDR3 pairs.
Iterative Beam Search
Explores sequence space efficiently by maintaining top candidates at each step and applying targeted mutations guided by a learned scoring function.
Scoring Model
A neural network trained on antigen-CDR3 binding pairs that predicts binding likelihood from ESM-2 embeddings, enabling rapid candidate evaluation.
Diversity Clustering
Groups similar candidates and selects diverse representatives to maximize coverage of promising sequence space regions.
Advanced Options
Generate → Filter → Cluster
Protein language model
Downloadable results