← Back to Home

AI Designs Recognition Sequences for T Cells: Artificial Receptors Show Activation Capability in Experiments

Finding T-cell receptors that recognize specific targets often requires extensive screening. ERTransformer proposes candidate sequences based on antigen fragments, and some designs have passed cell-based experiments. Whether they can be used safely in treatment still requires step-by-step validation.

By SURL BioNews

Enabling immune cells to recognize disease starts with finding suitable receptors. For researchers developing T-cell receptor therapies, this search is often time-consuming and expensive. A study published in *Genome Biology* used artificial intelligence to propose recognition sequences for subsequent experimental screening. Under specified conditions, some artificially designed receptors triggered T-cell activation responses comparable to, or even stronger than, those of natural receptors.

T-cell receptors (TCRs) recognize protein fragments displayed by other cells through human leukocyte antigen (HLA) molecules. This pairing mechanism is an important foundation for researchers seeking to guide immune cells to recognize tumor targets. The new model, ERTransformer, focuses on the CDR3 region of the receptor’s β chain, a variable sequence involved in recognition. Its outputs still need to be incorporated into complete receptors and tested in the corresponding antigen-presentation environment.

The model combines EpitopeBERT and ReceptorBERT, pretrained on 1.9 million antigen epitope sequences and 33.1 million TCR sequences, respectively. The team selected five epitopes with existing natural receptors available for comparison and generated 1,000 candidate CDR3 sequences for each. These candidates had low similarity to natural β-chain sequences while retaining reasonable lengths and amino acid compositions, providing more starting points for subsequent screening.

The real test came in cell-based experiments. The researchers selected some artificial receptors and measured T-cell activation using flow cytometry, obtaining responses comparable to those of natural receptors. However, not all 5,000 generated sequences underwent functional validation. The results were also limited to the receptor and peptide–MHC combinations used in the experiments. They cannot be used to infer the success rate of the entire candidate set, much less be directly equated with anticancer efficacy in humans.

The findings also illustrate the time needed to move from computational design to biological validation. In a public post, corresponding author Qingpeng Zhang said the project began during Jiannan Yang’s internship at Tencent AI Lab. The computational work was largely completed in 2022, followed by wet-lab validation carried out by Jianhua Yao and a collaborating team at Xijing Hospital. Experiments are still needed to bridge the gap between a model’s ability to propose sequences and those sequences’ ability to elicit a cellular response.

The team’s public ER-BERT repository provides instructions for pretraining, generation, and use, along with links to trained models and demonstration data. The workflow described in the documentation also uses an additional discriminative model to predict binding specificity and help select candidates. Such predictions still require experimental confirmation. The complete data must be requested from the researchers, so although the public code helps others examine the method, full reproduction still depends on obtaining the data.

The clearer use at present is to expand the candidate pool and prioritize experimental screening. Moving toward treatment will require confirming whether the designs work across more epitopes and HLA backgrounds, and whether they might recognize unintended targets and harm normal tissue. The paper was published on September 2, 2026. The publisher labels it as an accepted, peer-reviewed early version that will undergo further editing. It provides a starting point for functional validation but has not yet established evidence of clinical efficacy and safety.

References

  1. BIOENGINEER
  2. Genome Biology
  3. Qingpeng Zhang, study corresponding author (LinkedIn)
  4. TencentAILabHealthcare (GitHub)