Icon Legend

This session is not in your schedule.

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

Back

Bioinformatics, Computational and Systems Biology

Characterizing Altered Cell States in Blast-Related Traumatic Brain Injuries Using Graph Attention Networks

(A-34) Characterizing Altered Cell States in Blast-Related Traumatic Brain Injuries Using Graph Attention Networks

Thursday, October 12, 2023

9:30 AM – 10:30 AM PDT

Location: Exhibit Hall - Row A - Poster # 34

Presenting Author(s)

GG

Gabriel A. Gonzalez, n/a

Student
Columbia University
Ocala, Florida, United States

Co-Author(s)

Eva Soler

Undergraduate Researcher
Columbia University, United States

Primary Investigator(s)

EA

Elham Azizi

Professor
Columbia University, United States

Introduction::

Sub-Track-Single-Cell Measurements and Models: Blast-induced traumatic brain injury (bTBI) is a major medical concern for individuals exposed to war zones and urban terrorist attacks. In response to brain injury, neural stem cells (NSCs) in the subventricular zone (SVZ) exhibit heightened cell proliferation, differentiation, and migration towards the affected regions. However, the mechanisms and cellular responses that instigate and regulate these activities in the pathophysiological progression following bTBI are superficially understood and hard to diagnose. The advent of single-cell RNA sequencing (scRNA-seq) has brought about a revolution in biological research by providing unprecedented resolution in understanding the complex heterogeneity within biological tissues. Despite being widely employed to gain insights into both healthy and pathological states, the utilization of scRNA-seq for disease diagnosis or prognostication has been rather limited. In this study, we employed Graph Attention Networks (GATs), to analyze single-cell RNA sequencing (scRNAseq) data derived from the SVZ of mice following bTBI. Our objective was to unravel the intricate cellular and molecular alterations associated with bTBI that traditional statistical methods may fail to capture. GATs serve as a promising framework to perform cell fate analysis as they have the ability to learn from original features and graph structures using snRNAseq data. This project may pave the way for the development of novel therapeutic strategies and diagnostic tools. By integrating cutting-edge machine learning techniques with biological insights, our study opens new avenues for research in the field of TBI and may ultimately lead to improved clinical outcomes and patient care.

Materials and Methods:: Our GAT model was trained on published scRNAseq data obtained from mice SVZ (n = 6), divided equally into a bTBI group and a sham group, providing transcriptional profiles of 15,272 cells. The training was done with Python 3.9 and PyTorch Geometric using Google Colab. We use this dataset to perform inductive and transductive inference with hyperparameters tuned via grid search and Bayesian optimization. For the transduction task, we randomly assigned 10% of the nodes each for validation and testing while keeping the ratio of Sham:bTBI cells the same as in the full dataset. We randomly mask the labels of 50% of the data to produce a graph where half of the nodes are unlabeled, for the model to predict the labels based on the known label of the nodes in the graph and the features of the node. We feed in the features of all nodes during training with the goal to predict the masked labels. We first trained baseline models using common GAT parameters within our grid search. All models used the Adagrad optimizer, and gradient clipping was incorporated after the baseline search because of improved performance in the inductive task. We use the attention weights matrix to visualize the new graphs learned by the model, and extract meaningful features and investigate them in the context of existing biological knowledge on bTBI. From this, outside of normal trajectories produced from existing pipelines, we extract relevant biological information from the attention heads.

Results, Conclusions, and Discussions:: The GAT models exhibited impressive accuracy in predicting disease state. The best-performing GAT model achieved a test accuracy of 95% in the inductive task, outperforming other common machine learning models such as XGBClassifier and SVC. This highlights the effectiveness of GATs in analyzing scRNA-seq data and their potential as a diagnostic tool for bTBI. Through cell fate trajectory analysis, we explored the differentiation of neural stem cells (NSCs) into astrocytes, microglia, oligodendrocytes, and neurons after bTBI. The trajectories revealed the high differentiation potential of cells closer to NSCs and the distinct terminal states of the cells of interest. We identified top expressing genes for each cell type, including Cntnap2, Plp1, and Slc1a3, which are known to influence neural functions such as axon regeneration, myelination, and glucose transport. Biological insights obtained from the GAT model and gene analysis provided valuable information regarding the secondary biochemical response to bTBI. These findings align with existing studies on the relationship between traumatic brain injury and neuroplasticity changes. Overall, our research demonstrates the efficacy of GAT models in analyzing scRNA-seq data for bTBI research as they identified expressive genes shed light on the underlying mechanisms of bTBI-induced pathogenesis. Expanding the dataset would enhance the robustness and generalizability of the model, allowing for a more comprehensive understanding of brain cells from different individuals. Future work would focus on extracting biological information from the attention heads of the GAT model to gain deeper insights into the unique mechanisms underlying bTBI. With further improvements and expansion of the dataset, we envision that the development of this preliminary pipeline can serve as the start of a diagnosis and clinical tool to accurately identify bTBIs and facilitate treatment.

Acknowledgements (Optional): : Our research is supported by Professor Elham Azizi and the TAs Cameron Park, Linyue (Joy) Fan, and Yinuo Jin in the Department of Biomedical Engineering at Columbia University.

References (Optional): :

[1] M. Li et al., “Single-nucleus profiling of adult mice sub-ventricular zone after blast-related traumatic brain injury,” Scientific Data, vol. 10, no. 1, p. 13, Jan. 2023, doi: https://doi.org/10.1038/s41597-022-01925-y.

[2] S. Yamamoto, D. DeWitt, and D. Prough, “Impact & Blast Traumatic Brain Injury: Implications for Therapy,” Molecules, vol. 23, no. 2, p. 245, Jan. 2018, doi: https://doi.org/10.3390/molecules23020245.

[3] R. R. Hicks, S. J. Fertig, R. E. Desrocher, W. J. Koroshetz, and J. J. Pancrazio, “Neurological Effects of Blast Injury,” The Journal of trauma, vol. 68, no. 5, pp. 1257–1263, May 2010, doi: https://doi.org/10.1097/TA.0b013e3181d8956d.

[4] J. Cao et al., “The single-cell transcriptional landscape of mammalian organogenesis.” Nature 566, 496–502, 2019. doi: https://doi.org/10.1038/s41586-019-0969-x

[5] Finak, Greg, et al. “MAST: A Flexible Statistical Framework for Assessing Transcriptional Changes and Characterizing Heterogeneity in Single-Cell RNA Sequencing Data.” Genome Biology, vol. 16, no. 1, Dec. 2015, p. 278. BioMed Central, doi:

[6] F. A. Wolf, P. Angerer, and F. J. Theis, “SCANPY: large-scale single-cell gene expression data analysis,” Genome Biol, vol. 19, no. 1, p. 15, Dec. 2018, doi: 10.1186/s13059-017-1382-0.

[7] Haghverdi, Laleh, et al. “Diffusion Pseudotime Robustly Reconstructs Lineage Branching.” Nature Methods, vol. 13, no. 10, Oct. 2016, pp. 845–48. DOI.org (Crossref), https://doi.org/10.1038/nmeth.3971.

[8] Paul, Franziska, et al. “Transcriptional Heterogeneity and Lineage Commitment in Myeloid Progenitors.” Cell, vol. 163, no. 7, Dec. 2015, pp. 1663–77. DOI.org (Crossref), https://doi.org/10.1016/j.cell.2015.11.013.

[9] Velicˇkovic, Petar, et al. GRAPH ATTENTION NETWORKS. 2018.

[10] Oikari, Lotta E., et al. “Data Defining Markers of Human Neural Stem Cell Lineage Potential.” Data in Brief, vol. 7, Feb. 2016, pp. 206–15. PubMed Central, https://doi.org/10.1016/j.dib.2016.02.030.

[11] Xie, Xuanhua P., et al. “High-Resolution Mouse Subventricular Zone Stem-Cell Niche Transcriptome Reveals Features of Lineage, Anatomy, and Aging.” Proceedings of the National Academy of Sciences, vol. 117, no. 49, Dec. 2020, pp. 31448–58. pnas.org (Atypon), https://doi.org/10.1073/pnas.2014389117.

[12] Ravindra, Neal, et al. “Disease State Prediction from Single-Cell Data Using Graph Attention Networks.” Proceedings of the ACM Conference on Health, Inference, and Learning, ACM, 2020, pp. 121–30. DOI.org (Crossref), https://doi.org/10.1145/3368555.3384449.

[13] French, Louis, et al. “The Military Acute Concussion Evaluation (MACE).” Journal of Special Operations Medicine, vol. 8, Jan. 2008, pp. 68–77.

[14] Setty, Manu, et al. “Characterization of Cell Fate Probabilities in Single-Cell Data with Palantir.” Nature Biotechnology, vol. 37, no. 4, 4, Apr. 2019, pp. 451–60. www.nature.com, https://doi.org/10.1038/s41587-019-0068-4.

[15] Mizrak, Dogukan, et al. “Single-Cell Analysis of Regional Differences in Adult V-SVZ Neural Stem Cell Lineages.” Cell Reports, vol. 26, no. 2, Jan. 2019, pp. 394-406.e5. PubMed Central, https://doi.org/10.1016/j.celrep.2018.12.044.

[16] Glorot, Xavier, and Yoshua Bengio. “Understanding the Difficulty of Training Deep Feedforward Neural Networks.” Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, JMLR Workshop and Conference Proceedings, 2010, pp. 249–56. proceedings.mlr.press, https://proceedings.mlr.press/

v9/glorot10a.html.