1130 Rapid feedback cycles of generative AI and cell-based screening facilitate the high throughput immunoengineering of T cell receptors (TCRs) with specific binding profiles

<h3>Background</h3> T cell receptors (TCRs) are a unique class of targeting molecule. Through highly specific interactions with peptide:HLA complexes present on the cell-surface, they offer a means of therapeutically targeting any protein, with this specificity highlighted by their ability to selectively recognise cancer-associated single amino acid mutations.¹ However, identifying TCRs that have the desired specificity toward a target peptide has traditionally been challenging; often screening large numbers of peripheral T cells to find only a handful of potential molecules.² Even when a target-peptide specific TCR has been identified, its selectivity becomes crucial in developing a successful therapeutic, with reactivity toward off-target peptides having potentially serious safety implications.³ To increase the speed and success of TCR discovery, we have developed a <i>in silico</i> TCR design process leveraging recent advances in machine learning for protein sequences with rapid cell-based screening, where both specificity and selectivity are considered at the point of design. <h3>Methods</h3> Targets for TCR generation are computationally selected based on transcriptomic and immunopeptidomic data. Potential off-targets are identified through a combination of immunopeptidomic data, sequence similarity, and machine learning predicted binding similarity, in order to generate a list of sequences which can be used to guide generative AI models based on selective binding. Large protein language models are fine-tuned on experimental data which captures binding specificities between known TCRs and their peptide ligands, learning meaningful latent representations of both peptide and TCR sequences. These trained models are then used to generate TCRs <i>in silico</i> that are optimised for selective target binding, leveraging an energy-based diffusion process. Computationally-generated TCR sequences, synthesised alongside naive permutations, are expressed in a cell-based high throughput screen, providing binding data on up to a million TCR sequences per cycle, enabling further model refinement based on desired binding properties. <h3>Results</h3> Our pipeline has been applied to design novel TCRs against both viral targets and a public neoantigen derived from the common KRAS mutation G12D. We show how a generative model produces TCRs optimised for multiple-antigen specificity, binding to the target peptide but not self-peptides including the unmutated wild type KRAS protein. <h3>Conclusions</h3> Combining state of the art generative AI with high throughput screening approaches empowers a TCR design process that considers both selectivity and specificity, opening up the next generation of TCR-based therapeutics. <h3>References</h3> Simnica D, Kobold S. Neoantigen T-cell receptor gene therapy in pancreatic cancer. <i>N Engl J Med</i> 2022;<b>386</b>:2112–2119. Fonteneau JF, Larsson M, Somersan S, Sanders C, Münz C, Kwok WW, Bhardwaj N, Jotereau F. Generation of high quantities of viral and tumor-specific human CD4+ and CD8+ T-cell clones using peptide pulsed mature dendritic cells. <i>J Immunol Methods</i>. 2001;<b>258</b>:111–26. Cameron BJ, Gerry AB, Dukes J, Harper JV, Kannan V, Bianchi FC, Grand F, Brewer JE, Gupta M, Plesa G, Bossi G, Vuidepot A, Powlesland AS, Legg A, Adams KJ, Bennett AD, Pumphrey NJ, Williams DD, Binder-Scholl G, Kulikovskaya I, Levine BL, Riley JL, Varela-Rohena A, Stadtmauer EA, Rapoport AP, Linette GP, June CH, Hassan NJ, Kalos M, Jakobsen BK. Identification of a Titin-derived HLA-A1-presented peptide as a cross-reactive target for engineered MAGE A3-directed T cells. <i>Sci Transl Med</i>. 2013;<b>5</b>(197):197-ra103.