New Tumor-Specific Molecular Interaction Maps Capture Complexity of Cancer Networks

Though cancer researchers have made considerable headway in elucidating signaling and regulatory pathways, current representations fail to capture the full complexity of the mechanisms that mediate the effects of both genetic and pharmacological perturbations. More specifically, prior efforts to identify signaling and regulatory pathways do not capture their cross-tissue differences (context-dependence), are based largely on interactions reported in the literature, and do not effectively reflect their role in mediating oncoprotein-specific signals. Some researchers have recently begun to incorporate cell line-, tumor-, or tissue-specific information in efforts to produce findings that are context-specific, but comprehensive, proteome-wide depictions of human interactomes across different tissue contexts remain elusive.

In a paper recently published in Nature Biotechnology, Andrea Califano, Dr, chair of the Department of Systems Biology at Columbia University Vagelos College of Physicians and Surgeons, and Barry Honig, PhD, professor of systems biology, and their co-authors propose a fundamentally different representation of the signaling and regulatory machinery needed to modulate the function of a specific protein of interest in a specific tissue context, i.e., a protein mechanism of action. This novel representation is called a signaling map, or SigMap.

The paper presents OncoSig, an integrative machine learning (ML) framework for the systematic protein-centric reconstruction of tumor-specific SigMaps. The SigMaps reported are centered on virtually all known oncoproteins and contextualized to more than 20 tumor contexts. OncoSig is trained on integrated information derived from predicted physical protein-protein interactions, inferred from 3D structural data; and transcriptional and post-transcriptional interactions, from gene-expression and mutational profiles in large-scale repositories such as The Cancer Genome Atlas (TCGA).

The researchers first generated a KRAS-specific SigMap for lung cancer (the KRAS oncogene is commonly mutated in lung cancer). The SigMap not only included known KRAS interactors reported in published pathways, but also identified many novel synthetic lethal proteins that were subsequently validated in 3D spheroid models at a validation rate exceeding 80%; it also identified crosstalk with Rab and Rho pathway proteins. OncoSig generates a single integrated score that represents the probability that a protein belongs to a specific SigMap. Using lung adenocarcinoma as an example, of the 40 highest-scoring proteins predicted in the KRAS SigMap, 20 were already known, and 16 of the remaining 20 were experimentally validated.

“This is an extraordinary validation rate,” says Dr. Califano, who co-leads the Precision Oncology and Systems (POSB) research program at the Herbert Irving Comprehensive Cancer Center.

SigMaps might also be used to identify pharmacologically accessible candidate targets for many mutated oncoproteins, including KRAS, thus providing a valuable tool for guiding hypothesis-based studies to validate their disease-related relevance.

“This is fundamentally a new methodology,” says Dr. Califano, “and its usefulness should extend well beyond cancer. Pathways and networks are central to how cells recognize one another and communicate. One reason we are more advanced in dealing with cancer than with, say, Alzheimer’s disease, is that the cells of other diseases don’t proliferate. It’s much easier to culture cancer cells—proliferating is what they like to do. But SigMaps can also be constructed for proteins unrelated to cancer, greatly expanding the value of the algorithm.”

“Advances in ML methodology have enabled us to integrate these very different sources of input,” says Dr. Honig, a member of the HICCC’s POSB program. “The complexity of SigMaps is daunting, but biology is complex, and sometimes it is important to embrace that complexity.”