Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session C

Poster A2 - Generative Machine Learning for Single-Cell Clustering Identifies Determinants of CAR T Cell Longevity

Friday, October 25, 2024

10:00 AM – 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Leo Shen

Student
Yale University
Auburn, Alabama, United States

Co-Author(s)

Zhiliang Bai, PhD (he/him/his)

Postdoc Associate
Yale University
New Haven, Connecticut, United States

Primary Investigator(s)

RF

Rong Fan

Harold Hodgkinson Professor
Yale University
New Haven, Connecticut, United States

Introduction: Generative foundation machine learning models, which are pretrained on vast datasets and utilize a self-attention-based transformer architecture, have gained prominence in fields such as human language, computer vision, and healthcare due to their ability to recognize complex patterns within high-volume data. Analogous to large language models capturing the contextual associations within words of a sentence, the development of generative foundation models with biomedical applications has introduced the ability to mathematically represent the intricate relationships between cells and individual genes through the generation of vector embeddings. These models, which require extensive data for pretraining, demonstrate clear potential for single-cell biological analysis, enabling the semi-autonomous analysis of single-cell omics datasets in an accurate and time-effective manner. Concurrently, CD19-targeted Chimeric Antigen Receptor (CAR) T cell therapy has demonstrated exceptional response rates against acute lymphocytic leukemia (ALL) in early-phase trials, but frequently exhibited patient relapse represents a significant challenge in the next stage of cellular immunotherapy. Our translational studies investigating the relationship between CAR T cell functional properties and clinical response, including five individuals with eight-year ALL remission durations, has produced a comprehensive single-cell RNA sequencing (scRNA-seq) data atlas capable of providing insights into mechanisms underlying ultra-long-term relapse-free remission. Generative foundation models emerge as an optimal method for understanding complex patterns and relationships within this extensive dataset, ultimately providing a robust understanding of the molecular determinants contributing to long-term response.

Materials and

Methods: The whole-human scGPT generative foundation model, pretrained on 33 million cells from the CELLxGENE database, has demonstrated significantly accurate performance on single-cell downstream tasks versus both traditional analysis methods and other foundation models. Utilizing CAR T scRNA-seq data from a labeled reference partition of five healthy donors within the referenced ultra-long study, scGPT was further fine-tuned in a supervised manner over fifteen epochs to learn unique patterns and features within the provided data. Model layer weights were adjusted accordingly to generate final embeddings accurately capturing task- and context-specific information. Using these embeddings, in conjunction with B-cell aplasia (BCA) duration as a surrogate marker describing functional persistence of CAR T cells in clinical practice, robust reference mapping was performed to obtain a novel clustering and cell type annotation of the remaining query partition of 33 individuals across three BCA duration categories. Reference mapping was conducted using a similarity search between the cell embeddings of the reference and query partitions, incorporating a majority voting and k-nearest neighbors mechanism. To delineate the characteristics of CAR T cells, associate them with clinical outcomes, and identify mechanisms mediating long-term response, we considered two conditions: basal unstimulated CAR T and CD19-specific activated CAR T. After the complete annotation of all cells, the sequencing data was divided into the aforementioned treatment groups. Cell type proportions within treatment groups were calculated for each BCA group, providing insight into relationships between CAR T longevity and clinical response. Specific cell type trends were highlighted for further investigation.

Results, Conclusions, and Discussions: By analyzing the association between cell type proportions within the annotated clusters generated by reference mapping and patient BCA duration, notable molecular determinants contributing to durable remission are identified. Several characteristics of basal CAR T cell profiles distinguish between BCA groups. Within proliferative clusters and the cluster of FOS and JUN expressing cells, a substantially lower proportion of BCA-L cells was present compared to other groups. For the cluster enriched in memory gene markers such as CCR7, TCF7, IL7R, and CD27, and the naive cluster expressing CD62L, BCA-L cells were instead observed at a much higher comparative proportion. Together, these four observations imply conserved memory and naive states, low proliferating activity, and prevention of unintended activation, all of which suggest that preserving the inactive state is associated with CAR T longevity. Additionally, the analysis of CD19-specific activated group also presents unique trends differentiating BCA groups. The observation of lower BCA-L cell proportions within highly proliferative clusters suggests that CAR T cells undergoing rapid expansion may sustain less effective immune functions. Likewise, within the cluster of cells displaying high exhaustion signature, BCA-L proportions were also notably lower in comparison with less persistent BCA2 and BCA1 CAR T cells. Conversely, in the cluster characterized by expression of the C chemokine encoding XCL1 and XCL2 genes, BCA-L was observed in high comparative proportion. These trends point towards reduced proliferation, enhanced potential of immune memory, and active immune recruitment, all contributing towards an optimal state of anti-tumor defense for CAR T cells sustaining long-term response. Identifying and understanding biomarkers with robust potential to optimize CAR T persistence could effectively prevent relapse within ALL patients treated with CAR T therapy. Additionally, generative foundational models are able to rapidly and efficiently analyze patient cell behavior through gene expression data, opening the possibility of providing on-demand access to personalized biological insights for CAR T patients. Overall, this study highlights the clear potential of utilizing generative foundation models for expediting translational research within immuno-oncological personalized medicine.

Acknowledgements (Optional):