(B-75) Diffusion Probabilistic Modeling for Rapid Assessment of T-cell Receptor Specificity

Introduction:: Predicting the binding affinities of TCR-antigen plays a crucial role in recognizing and eliminating foreign and tumor-associated antigens. The greatest challenge in predicting TCR specificity arises from the diversity of possible TCR and antigen combinations relative to those that can be studied or even realized in a single individual. One of the current state-of-the-art models for synthetic TCR-antigen generation is generative adversarial networks, notorious for their difficulty in training and mode collapse. Denoising Diffusion Probabilistic Models (DDPMs), a class of generative models based on statistical thermodynamics, produce high-quality synthetic samples in certain areas. We propose and extend a hybrid model which combines a diffusion model with inferential statistical techniques, Rapid Coarse-grained Epitope TCR (RACER) ^[1]. This novel method is trained on robust binding TCR-antigen systems with available crystal structures which are improved by synthetic samples to efficiently generate reliable predictions. Combining the strengths of RACER and diffusion models improved accuracy and runtime compared to traditional search-based and generative models.


Materials and Methods:: Introduction: The T cell repertoire, composed of approximately ~10⁸ unique T cell receptors (TCRs) plays a vital role in detecting and neutralizing a wide range of antigens. These TCRs contribute significantly to the adaptive immune response, assisting the immune system to recognize and eliminate various pathogens and cancerous cells while maintaining tolerance to self-antigens. Given their potential impact on immunology research and cancer immunotherapy advancements, there is a strong demand for accurate, high-throughput prediction methods for TCR-antigen pairs. Such prediction methods could also contribute to a deeper understanding of immune system function, T cell-mediated autoimmunity and the development of novel vaccines or targeted therapies for infectious diseases.



Materials and Methods: We present a hybrid approach integrating a diffusion model which learns a distribution over diverse and larger protein complex structures with a coarse-grained protein energy model to calculate the TCR-peptide binding energy based on a pairwise amino acid energy matrix. By developing this model, we can learn a distribution over diverse and longer protein complexes' backbone structures to sample protein complexes from this distribution, conditioned on a given motif within the interacting proteins. This approach enables the generation of protein complex structures that can accommodate specific motifs, facilitating the study and design of novel protein interactions. Simultaneously, we optimize the energy matrix for maximal resolution of robust and randomized TCR-antigen pairs. Training samples undergo clustering by structural and functional similarity, generating optimized models per cluster. We evaluate cross-cluster transferability through prediction sensitivity and specificity.


Results, Conclusions, and Discussions:: Results and Discussion: Using the RACER approach on untrained data achieves a sensitivity of 92%. Incorporation of the diffusion model into the training process on the MART1 system and unique TCR-antigen pairs leverages features from the augmented protein database, utilizing the enriched and diverse data to demonstrate improved performance. Our results, when combined with hybrid sequence and structural-based predictive models, suggest that accurate prediction of TCR-peptide pairs within a given HLA restriction is improved by learning generic features from protein-protein interactions.



Conclusions: Our results demonstrate that our procedure is capable of generating diverse complex scaffolds with computation time reliably, and the RACER model utilizes these scaffolds, crystal structures, and a pairwise energy model to accurately predict TCR-peptide binding affinities. We presented a diffusion generative model, tailored to the task of TCR-peptide binding affinities, representing a paradigm shift from previous machine learning approaches. Our work demonstrates the potential of machine learning methods to be applied in realistic TCR-peptide binding affinities design settings aiming to predict therapeutic TCR-antigen pairs in the medical setting through largescale clinical implementation.


Acknowledgements (Optional): : Acknowledgments: This work is supported by CPRIT grant (RR210080).


References (Optional): : References: [1] Lin X, George JT, Schafer NP, Ng Chau K, Birnbaum ME, Clementi C, Onuchic JN, Levine H. Nat Comput Sci 2021;1:362-373.