CRAFT: an antigen-specific TCR generative model for conditional de novo design with physicochemical constraints.

Authors

Xiaochuan Yang, Zaiduo Li, Zhijuan Li, Wei Zhang, Xianyu Zhang, Tiansong Yang, Mingliang Liu, Xiaokun Li

Published in

Briefings in bioinformatics. Volume 27. Issue 6. Nov 02, 2026.

Abstract

T-cell receptors (TCRs) are essential to adaptive immunity and mediate the recognition and elimination of malignant or infected cells through antigen-specific interactions. Despite the clinical progress of TCR-based immunotherapies, computational de novo design of antigen-specific TCRs remains difficult. The main difficulties include the complexity of peptide-major histocompatibility complex (pMHC)-dependent recognition, strict physicochemical requirements at the binding interface, and the small amount of paired training data available for model development. Current generative approaches depend largely on 1D sequence statistics and do not capture pMHC presentation context, interfacial physicochemical properties, and native receptor priors within a unified framework. We present conditional reconstruction of antigen fragments for TCRs (CRAFT), a decoupled generative model for pMHC-conditioned de novo design of antigen-specific TCR CDR3$\beta $ sequences under physicochemical constraints. With large-scale pretraining, CRAFT learns base representations that help reduce the limitations imposed by data scarcity. We further incorporate antigen-derived contextual information to guide antigen-specific TCR CDR3$\beta $ sequence design. The architecture comprises an antigen encoder (PrimeBERT) that captures the pMHC context and an autoregressive decoder (PrimeTCR) that preserves native receptor priors. The encoder and decoder are coupled through cross-attention and augmented with Atchley factor embeddings, allowing CRAFT to incorporate physicochemical constraints relevant to antigen-conditioned receptor design during sequence decoding. Across multiple evaluations, CRAFT demonstrates competitive sequence naturalness, together with favorable performance in specific binding, and predicted structural plausibility. Overall, CRAFT provides a computational framework for developing personalized tumor immunotherapies. It also holds potential to support clinical translation by enabling the efficient, targeted discovery of candidate TCRs.

PMID:
43016166
Bibliographic data and abstract were imported from PubMed on 12 Nov 2026.

Read full publication at:
Please sign in to see all details.

Stats

1-terrible, 9-excellent. How would you rate this publication? Sign in in to submit your rating.

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.