Signal Dissection and Modern Analysis for Molecular Medicine

Michael Grau

Gene expression data from cancer patients capture the net superposition of many simultaneously active biological programs, activatory and inhibitory pathways alike. Their individual contributions are entangled in the measured signal. Inferring the specific generating interactions from such overlapping sums is a fundamental unsolved problem in molecular medicine and beyond. Signal Dissection by Correlation Maximization (SDCM) is an unsupervised learning algorithm designed to address exactly this challenge. Rather than imposing orthogonality or linearity as PCA does, SDCM is driven by a nonlinear bimonotonic consistency requirement — the mathematical key that enables the otherwise highly ambiguous decomposition of measured signals into high-dimensional matrix summands. Each discovered signature defines clear gene and sample rankings, based on extracted significant correlations between top genes and samples. [1] Validation against versatile test signals has demonstrated substantially higher specificity to true source signals than principal components of PCA. Applied to ~500 DLBCL patient samples spanning ~63,000 transcripts, the dimensionality-reduced signature space captured heterogeneity in the form of ~115 signatures — including cell-of-origin, immune cell infiltration, gender, and unknown effects. A bivariate model based on two signatures in particular revealed previously unknown and strong differences in patient survival, constituting an independent biological validation since survival data are not part of the input. The talk will further discuss an outlook on our ongoing work dissecting pancancer transcriptomics, and perspectives integrating transcriptomic and genomic data. [1] Grau, M., Lenz, G. & Lenz, P. Dissection of gene expression datasets into clinically relevant interaction signatures via high-dimensional correlation maximization. Nature Communications 10, 5417 (2019). https://www.nature.com/articles/s41467-019-12713-5