Antibody Steering

Overview

I am building an antibody language-modeling stack for targeted, antigen-conditioned HCDR3 editing. Rather than generate entire antibodies from scratch, the system first learns antibody sequence grammar from Observed Antibody Space data, refines paired heavy/light-chain representations, and then fuses them with antigen context for local infilling.

The current implementation includes OAS and antibody-antigen preprocessing, a dual-stream cross-attention model, real-label compatibility training, and an HCDR3 infiller that can optionally use a separate compatibility model to guide each residue choice.

Steering with an honest ceiling

Inference-time guidance can only reweight residues that the base model already considers plausible. The project therefore includes a reachability probe that measures whether the compatibility signal can meaningfully change a decision before spending compute on a guidance sweep. If the base policy assigns a useful residue effectively no probability, increasing the guidance strength cannot recover it.

That distinction keeps three ideas separate: guidance changes a sampling run, preference post-training would change the model’s weights, and sparse-feature methods may eventually help interpret what changed. The latter two remain future directions until the antigen-conditioned policy passes independent evaluation gates.

Current focus

• leakage-aware preprocessing for OAS and antigen-specific antibody data,
• paired heavy/light and antibody-antigen representation learning,
• fixed- and variable-length HCDR3 infilling,
• guided generation and steering-reachability diagnostics, and
• reproducible experiment, checkpoint, and evaluation contracts.