(T0930-10-56) An End-to-End Method for AI-Driven Peptide Vaccine Formulation and Translation through Deep Learned Exosome Precision Drug Delivery

Purpose: Extracellular vesicles, namely exosomes, are unique biological nanoparticles that mediate cellular communications through trafficking towards different cellular phenotypes, inspiring wide biotechnological development for clinical translation. Therapeutic exosomes are reported to significantly reduce immunogenicity, a common challenge when formulating new biologics. However, mechanistic studies into exosome therapy by tissue-specified targeting are still required to elucidate critical pathways. Moreover, experimental approaches to mechanistic investigations are arduous and low-throughput, instanced through industry’s PBPK approaches to small molecules and monoclonal antibodies. Given the vast biological space that exosomes encompass, current experimental methodologies are unable to fully tap the potential of exosome therapeutics. Fortunately, artificial intelligence (AI) has emerged as a method to streamline mechanistic studies. Thus, we sought to combine AI with exosome pharmaceutical design to doubly illuminate key molecular drivers and establish translatable downstream efficacy of exosome therapy in vitro­. Specifically, we demonstrate our methodology to unlock exosome therapeutic potentiality with an AI-driven pipeline, ExoGAN. ExoGAN is a deep learning model that learns to tether key exosome cell-targeting cues with pharmaceutical development by in silico generation of surface functionalize-able peptide aptamers (PA). As proof-of-concept, we induced immunogenicity by targeting CD8+ T cells within peripheral blood mononuclear cells to simulate exosome vaccine development through mechanistically understanding HLA-presentation. Our methodology tackles the combined challenge of revealing exosome therapeutic processes and producing a feasible precision pharmaceutical faster than existing strategies. Broadly, we expect our AI-driven methodology to serve as a generalized-and-benchmarkable strategy for researchers and clinicians when developing future exosome therapeutics.

Methods: Our deep learning approach begins with extracting all HLA-relevant data from the Immune Epitope Database (IEDB) to incorporate into ExoGAN. ExoGAN functions by understanding the physiochemical modalities of HLA molecules when complexed to MHC-I, developed through molecular geometric optimization by utilizing the ORCA platform. We benchmarked ExoGAN’s physiochemical understanding of HLA against conventional machine learning algorithms. Next, we applied ExoGAN’s training into constructing PAs through generative adversarial networking (GAN). GANs take their trained molecular knowledge and iteratively evolve new molecules distally by passing on parental molecular information until completion. Completion is determined as molecular similarity, akin to parent-and-child. Generated PAs are computationally assessed by molecular docking, top ranked PAs are synthesized and evaluated for true MHC-I affinity by biolayer interferometry. Finally, top PAs are conjugated onto exosomes to induce immunogenicity by CD8+ T cell activation through presentation, measuring cytokine secretion by a multiplex cytokine assay.

Results: We first structured the data acquired from IEDB, where they report individual peptide sequences corresponding to specific MHC-I variants. We selected HLA-A2:01 (*A2) to focus our translation of peptide vaccines for human diseases, such as multiple myeloma. Multiple myeloma is reported to present *A2 on their surface, but evades CD8+ T cells, thus we aimed to present synthetic *A2s to trigger immunogenicity against multiple myeloma. Our deep learning network, ExoGAN, performed amazingly when benchmarked against conventional machine learning algorithms, classifying binders with 100% accuracy on purely binding affinity. Implementing residue position and importance through ordinal coding retained this accuracy, however, downstream translation was deemed inefficient by these coding methods. We opted to switch to a user-friendly alphanumeric encoding strategy, of which brought the previous classification to 60%. We returned this accuracy to 99% by attaching additional physiochemical concepts commonly considered in vaccine design. Following this result, we progressed our investigation by applying ExoGAN’s knowledge to generate new PAs by GAN. GAN functions by learning established *A2 molecular modalities, and attempts evolutionary generation of PAs, iteratively refining evolved PAs to be distal from *A2s. ExoGAN successfully turnovers 200 PAs in under 15 minutes, of which each were computationally docked and contrasted against their *A2 counterpart. PAs outperformed their *A2 on a computational basis, which motivated us to synthesize top PAs (i.e most distal) and their corresponding *A2 for experimental evaluation. Experimentally, PAs outperformed *A2s in binding affinity and immunogenicity, measured by biolayer interferometry and IFN-Y secretion within serotyped HLA*02:01 peripheral blood mononuclear cells, respectively. This strong result allowed us to extend our methodology by attempting exosome presentation with PAs and *A2s, doubly simulating peptide vaccine formulation and translation. Ultimately, we have successfully constructed and demonstrated a peptide vaccine design platform integrating mechanistic investigation with pharmaceutical development by applying artificial intelligent understanding of exosome presentation with great in vitro efficacy.

Conclusion: We developed a novel methodology integrating mechanistic investigations and downstream pharmaceutical development, demonstrating its significance for developing a peptide vaccine with precision delivery by exosomes. Peptides precisely targeting MHC-I and presented by exosomes were successfully created through the deep learning framework, ExoGAN. ExoGAN capably produced latent peptide vaccines that were computationally screened and rigorously experimentally validated by biolayer interferometry and multiplex cytokine screening. Hence, our framework will culminate as the first high-throughput mechanistic peptide vaccine design and precision pharmaceutical development strategy utilizing extracellular vesicles, with future work focusing on establishing in vivo efficacy

Acknowledgements:Funding support through NIHGMS R35GM133794


Figure 1. Graphical abstract detailing our AI-driven methodology to simultaneously investigate mechanistic drivers of therapy and translate these discoveries directly.


Figure 2. ExoGAN capably understands HLA and outputs peptide aptamers. A) ExoGAN's deep learning methodology to produce MHC-I aptamers. B) How feature engineering physiochemical modalities into ExoGAN's learning course is achieved. C). Benchmarking against conventional machine learning algorithms based on binding affinity (BA) and learning with additional physiochemical modalities (F). Colors represent implementations to ordain positional importance into the algorithms, with our method shaded with dashed box (SNN: Shallow neural network, SVM: Support vector machine, NB: Naive Bayes, EN: Elastic Net, RF: Random Forest). D) Principal component analysis depicting a combination of weighted fractions of physiochemical modalities clearly distinguishes HLA molecules. E) A receiver operator curve and its associated AUC, listed in the barplot, demonstrating how Alphanumeric encoding is practical and downstream utilizable. F) T-stochastic neighbor embedding visualization of HLA (True peptides) and its distal peptide aptamers (Generated) post ExoGAN turnover. An example of computational screening is highlighted by the boxed selection of two points that were decoded into their structures, molecular docked using HDOCK, and compared their binding strengths in RMSD. Higher negative values indicate better binder.


Figure 3. Experimental validation and translation of ExoGAN. A) Our experimental methodology to test immunogenicity with both HLAs and their distal peptide aptamers. B) Biolayer interferometry evaluation of binding affinity to MHC-I's variant HLA*A2:01 for selected peptide aptamers and their HLA counterparts. C). IFN-Y secretion from peripheral blood mononuclear cells measured by a sandwich ELISA. D) The tracked peripheral blood mononuclear cell viability when incubated with either peptide aptamers or HLAs. E) Our established methodology to capture-and-functionalize-and-release MHC-I+ exosomes with selected peptide aptamers.