(A-8) TransCompR: Translatable Components Regression Reveals Human Sepsis Mechanisms through Mouse-Derived Gene Expression Patterns

Introduction:: Sepsis, a life-threatening condition triggered by a dysregulated host response to infection, continues to pose significant challenges to global health due to its complex pathophysiology and high mortality rates. Despite extensive research, our understanding of sepsis remains incomplete, which underscores the need for innovative approaches to unravel its underlying molecular mechanisms and find a curative treatment. A key limitation of sepsis research lies in the difficulty of translating findings from pre-clinical animal models, predominantly mice, to humans, due to species-specific differences in gene expression and immune response. This study leverages a novel analytical approach, translatable components regression (TransComp-R), to bridge this interspecies gap. TransComp-R is a method that applies Principal Component Analysis (PCA) on mouse transcriptomic data (e.g. microarray, RNA-seq, scRNA-seq) to derive Principal Components (PCs), projects the homologously aligned human transcriptomic data into such a transformed space, and utilizes various metrics such as Davies-Bouldin score, individual component logistic regression accuracy, and general logistic regression model coefficients of PCs to quantitatively evaluate the ability of these mouse-derived patterns to distinguish between healthy and septic humans. This approach allows us to identify and uncover mouse model-derived molecular patterns of value for prediction and potential therapeutic intervention, thereby advancing our understanding of this complex disease.

Materials and Methods:: The total RNA microarray data from the blood of patients (septic, healthy) and mouse models (infected, control) was obtained from the Gene Expression Omnibus dataset GSE33341. Data has been processed as described in the original manuscript. Probe IDs were matched to the gene symbol referred to in the mouse4302SYMBOL and hgu133plus2 database in R. Human-mouse homologous genes were paired and aligned using biomaRt in R. TransCompR computed the top 50 principal components that best captured the variance in mouse models and projected the human gene expression onto the mouse-derived feature space by multiplying each human gene’s expression with the loadings of its mouse homolog’s on PCs. 3 metrics–Davies-Bouldin score, individual component logistic regression accuracy, and general logistic regression model coefficients of PCs–were used to quantify the ability of each mouse-derived PC to partition different disease statuses in humans. Specifically, for the later 2 logistic regression-based metrics, the Python scikit-learn LogisticRegression model with l1 penalty and Bayesian optimized hyperparameters was used to fit the data and reported a 5-fold cross-validation-averaged accuracy. The approach was also packaged as the Python toolkit TransCompR (https://github.com/BrubakerLab/TransCompR), in which the main function performs the PCA, and several auxiliary functions visualize and interpret the results.

Results, Conclusions, and Discussions::

Following the PCA, we discovered that the difference between the score of healthy and septic patients varied most significantly on the mouse-derived PC 3, 4, 7, 30, and 49, especially PC 3 and PC 7. Specifically, healthy patients had a lower score on PC 7 and a higher score on PC 3 compared to the septic patients. GSEA on the top 100 positive- and negative-loading genes on these 2 PCs revealed that the upregulation of fatty acid beta-oxidation favors the prediction of healthy patients while the upregulation of oxygen species metabolic process, copper ion transport, phosphatidylcholine synthesis and the downregulation of mitochondrion organization skewed towards the side of septic patients. These findings not only validate previous research but also underscore the utility of TransComp-R as a powerful tool for uncovering predictive molecular patterns.

One of the main advantages of TransComp-R is its ability to bridge the gap between human and mouse data, providing a method to project human data into a "mouse-like" transformed space. This can reveal relationships and structures that might be obscured in the human data alone, offering new avenues for research. Moreover, our improvement in TransComp-R, which incorporates various metrics to evaluate the effectiveness of these mouse-derived patterns, contributes to a more comprehensive and robust analysis of their predictive power.

Acknowledgements (Optional): :

References (Optional): : Douglas K. Brubaker, et al., An interspecies translation model implicates integrin signaling in infliximab-resistant inflammatory bowel disease.Sci. Signal.13,eaay3258(2020).DOI:10.1126/scisignal.aay3258