SALVE: prediction of interorgan communication with transcriptome latent space representation

Massive transcriptomics data allow gene relationships to be discovered from their correlated expression. We describe SALVE, a method to infer the associations between secretome-encoding transcripts and gene modules in a distal organ from RNA sequencing data. This method builds upon similar bioinformatics approaches by introducing transcriptome latent space representations and transfer learning to simultaneously increase discovery power and predict downstream functional associations. Applied to GTEx v8 data, we show the method readily recapitulates canonical endocrine relationships, including insulin and adiponectin signaling, while inferring new candidate organokines and their signaling modality. We also explore its utility for generating new hypotheses on cardiokine candidates and finding distal factors that may affect cardiac protein synthesis and metabolism. The predictions suggest a potential role of circulating galectin-3 (LGALS3) in regulating cardiac protein synthesis and homeostasis, which can be recapitulated in part in human induced pluripotent stem cell (hiPSC)-derived cardiomyocytes. This method may aid in ongoing efforts to delineate interorgan communications and endocrine networks in various areas of study.Supplemental Data S1. All filtered predictions.Suppl. Figure S1. Tissue distributions of GTEx v8 transcriptome data<b>A.</b><b> </b>Plot of t-distributed stochastic neighbor embedding (t-SNE) of GTEx v8 transcriptome data grouped by gross tissue type (SMTS). Tissue subtypes are not further distinguished by color.<b>B.</b><b> </b>Plot of t-SNE of data following normalization and batch correction steps.Suppl. Figure S2. Significant gene to latent variable associations clustered by tissues.Heatmap showing significant correlation among prioritized organokines (adj.P ≤ 0.005, S* ≥ 2) across all source-target tissue pairs.Suppl. Figure S3. Association of adipose <i>LBP</i> with liver metabolismScatter plots showing the correlated expression of adipose (visceral – omentum) and liver latent variables related to glucose and lipid metabolism. Each data point is an individual donor sample with both adipose and heart transcriptomes in GTEx v8. Red dashed line: best fit linear fit. Ribbon: standard error.Suppl. Figure S4. Candidate organokines associated with cardiac collagen formation.<b>A.</b><b> </b>Bar chart showing the number of significant organokines from several cardiometabolically relevant tissues that are significantly associated with heart latent variable 2 with REACTOME Collagen Formation annotation.<b>B.</b><b> </b>Dot plot showing the correlation coefficient and significance of organokines from the adipose tissue (subcutaneous), liver, and skeletal muscle with heart latent variable 2 (REACTOME Collagen Formation).<br>