Abstract A036: Developing the next generation of T cell engager (TCE): Combining high throughput immunoassays with generative AI to develop TCEs toward multiple RAS mutations

Abstract Introduction Through highly specific interactions with peptide: HLA complexes (pHLA) present on the cell-surface, T cell receptors (TCRs) offer a means of targeting any protein, with this specificity highlighted by their ability to selectively recognise RAS-associated neoantigens (mRAS). However, to date TCR-based therapies targeting mRAS have been restricted to a limited patient population that expresses both a distinct RAS mutation and a specific HLA allele. To increase the success of mRAS-targeting TCEs, we have developed an in silico TCR design process that optimises for multiple mRAS binding properties, leveraging high throughput immunoassays and recent advances in machine learning for protein sequences. Methods TCR training datasets were generated through extensive site-directed perturbation of previously identified single complex mRAS binders. Both site and type of perturbation were steered using structural modelling and in silico prediction. Training sets of up to one million sequences were screened for TCR functionality toward multiple mRAS and off-target peptides, using a T cell reporter-based screen optimised to provide a proxy for apparent affinity. Output from training datasets were used to fine tune large protein language models and generate TCRs in silico that are optimised for selective mRAS binding. This enables further model refinement based on desired binding properties and target affinities. Multispecific mRAS candidates were profiled for selectivity using cell-based pHLA library screens that simultaneously present an mRAS peptide alongside all single amino acid peptide variants, generating a unique peptide specificity footprint for each molecule. Results Training libraries were built around multiple TCRs with selectivity toward different mRAS mutations. As expected, initial libraries predominantly respond to cognate ligand, with discrete populations that show levels of cross reactivity. This data is used by the model to understand local binding rules for the target, as well as sites involved in sequence selectivity. TCR binding data is complemented by peptide specificity footprints using a pHLA screen, revealing the key areas of peptide sequence associated with selectivity. Monitoring individual TCR sequences through learning cycles demonstrates a significant improvement of binding affinity toward additional mRAS targets, with no apparent loss of general selectivity. This is demonstrated in a TCE format where molecules show near equivalent binding toward two key mRAS targets, and are able to potently activate bystander T cells in the presence of multiple mRAS targets. Conclusions Through innovative library design and high throughput mammalian cell screening we have generated a unique mRAS training dataset for use in generative AI. This data has enabled in silico mRAS TCR generation that demonstrate unique selectivity profiles, opening a new class of mRAS TCE therapeutics. Citation Format: Dubravka Pezic, Victoria Koullourou, Somayya Manzoor, Anton Shmelev, Joshua Meyers, Valerie Coppard, Lilly Wollman, Alex S. Powlesland. Developing the next generation of T cell engager (TCE): Combining high throughput immunoassays with generative AI to develop TCEs toward multiple RAS mutations abstract. In: Proceedings of the AACR Special Conference in Cancer Research: RAS Oncogenesis and Therapeutics; 2026 Mar 5-8; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Res 2026;86 (5Suppl₁): Abstract nr A036.