Artificial intelligence is expediting the development of therapeutic immunotherapies

CD8+ T cells recognise antigens through their T-cell receptors (TCRs), which bind peptides presented by major histocompatibility complex (pMHC) class I. However, the identification of TCRs from patient material for therapeutic use is laborious, technically challenging and potentially restricted by the TCR repertoire in the donors used for selection. The introduction of artificial intelligence (AI)-driven computational structure tools has provided the opportunity to design high-specificity pMHC-binding artificial TCRs for precise targeting of tumour antigens.

At the Technical University of Denmark, we used generative AI models, including the RFdiffusion and AlphaFold2 structure-generation and -prediction tools, to generate TCR mimics called ‘minibinders’ that are specific to the NY-ESO-1 peptide bound to HLA-A*02 (Science. 2025;389:380–385). Screening in a cellular expression system demonstrated that some of these binders recognised the peptide MHC sequence of interest but did not, to any great extent, cross-recognise other peptide MHCs. Our proof-of-concept findings were published simultaneously with those of two other groups, one of which looked more closely at the broad capacity of this approach with respect to the extent of possible peptide MHC targets (Science. 2025;389:375–379). The other provided greater insights into cross-binding aspects and the potential use of minibinders as T-cell engagers (Science. 2025;389:386–391).

One of the advantages of the AI-driven approach, compared with generation of TCRs from an established repertoire, is the chance to build libraries of minibinders that recognise a particular peptide and then to screen each agent for its binding properties. In-depth characterisation of the interactions had between TCRs that recognise the same epitope and a pMHC complex can be achieved using DNA barcode-labelled MHC multimer ‘fingerprinting’ (Nat Biotechnol. 2018:10.1038/nbt.4303). This tool can be used to help design a minibinder that possesses the properties appropriate for a particular therapeutic approach.

The potential application of minibinders is far reaching. For example, our work has shown that they can be adapted for use in a chimeric antigen receptor construct – with the minibinder, rather than the antibody, as the recognition element – where the specificity led to effective target-cell death (Science. 2025;389:380–385). The relatively small size of protein minibinders compared with TCRs means that they are more compatible to being produced in soluble form. In addition, they may be useful as payload carriers or T-cell engagers targeting pMHC. They also have potential application in the blocking of pMHC recognition or degradation in the context of autoimmune disease.

The pace of AI-associated change means that minibinders may find their way into the clinic sooner than might have been thought. However, before then, we need comprehensive, robust data on specificity and peptide cross-recognition aspects. While exclusive recognition of a particular peptide is unrealistic, it is important to set limits regarding the degree of flexibility that is safe and to ensure that there is no cross-binding to critical proteins. Investigation of better therapeutic target candidates should also be pursued. The enormous amounts we can expect to learn from these AI models in a relatively short time frame will really help to expedite the rational design of protein structures for improved therapeutic approaches.