A grand challenge in the field of mass spectrometry proteomics is the problem of peptide sequencing. The dominant approach to this problem uses a database search; however, with the use of machine learning, peptide sequencing can be solved without using a database. Casanovo is a recently-developed, state-of-the-art machine learning model that solves this problem. However, Casanovo is trained on biased data, because most mass spectrometry proteomics experiments digest proteins using trypsin, which preferentially cleaves after lysine and arginine. This can result in incorrect predictions for data that was not generated using trypsin. Hence, we hypothesized that using a non-tryptic dataset to refine an existing Casanovo model would produce more accurate predictions on non-tryptic data. I constructed an unbiased dataset by downsampling from preexisting data, yielding a set of peptides with a uniform distribution of n-terminal amino acids. After splitting the data and fine-tuning Casanovo on an unbiased training dataset, I used the model to predict on an unbiased validation set. I also applied Casanovo to two non-tryptic datasets: antibody sequencing data and immunopeptidomics sequencing data. For the latter dataset, an antibody binding affinity tool, NetMHCpan4.0, was used to determine the binding probability of the predictions to test plausibility. We demonstrate that the fine-tuned model, Casanovo_ne, increases performance when predicting on non-tryptic data. Applied to the uniformly distributed validation set, Casanovo_ne predicts more accurately than the original Casanovo model, and predicts more uniform terminal amino acid distributions. Additionally, Casanovo_ne predicts more peptides that are likely to be MHC binders than a database search strategy. Finally, Casanovo_ne accurately represents the digestion rules for most non-tryptic enzymes. As future work, we will modify Casanovo to take as input the identity of the digestion enzyme, alongside each spectrum. We hypothesize this approach will further improve Casanovo’s performance for samples prepared with alternative enzymes.

Adoptive T cell therapy is a promising therapeutic strategy for the treatment of many hematologic malignancies, however, its efficacy in solid tumors poses several challenges. Some of these challenges include the limited infiltration and activation of cytotoxic T cells due to the effects of a diverse immunosuppressive environment within the solid tumor. One of the main suppressive immune cells present in several solid tumors are regulatory T cells (Tregs) and high numbers of Tregs within the solid tumor have been correlated with poor prognosis. Therefore, there is a clinical need to develop strategies targeting the suppressive immune cells that limit the efficacy of adoptive T cell therapy. Tumors have highly dysregulated metabolism, which results in the secretion of multiple metabolites into the extracellular space. This allows a buildup within the tumor microenvironment, which may have an effect on the infiltrating immune cells. Our group has identified one metabolite, succinate, that enhances Treg numbers within the tumor microenvironment. Here we further explored the effect of succinate on the function of Tregs. To examine this, we identified tumor cell lines that produce succinate (lung and melanoma) and further altered these to modify the levels of succinate secreted by these cells. We then co-cultured high succinate secreting tumor lines with healthy donor CD4+ T cells that were isolated from PBMCs. We screened these cells and found that the higher levels of succinate resulted in higher numbers of Tregs and increased anti-inflammatory function, as evidenced by TGFb levels in Tregs. Future experiments will validate these findings in in vivo mouse models with the aim of developing synergistic approaches to enhance adoptive T cell therapy.