Mechanism Meets Big Data: Different Strategies for Machine Learning in Cancer Research

Emanuele Ratti Ratti (University of Notre Dame), Ezequiel López Rubio (University of Málaga)
We argue that mechanistic models elaborated by machine learning cannot be explanatory by discussing the relation between mechanistic models, explanation and the notion of intelligibility of models. We show that the ability of biologists to understand the model that they work with (i.e., intelligibility) severely constrains their capacity of turning the model into an explanatory model. The more a mechanistic model is complex (i.e., it includes an increasing number of components), the less explanatory it will be. Since machine learning increase its performances when more components are added, then it generates models which are not intelligible, and hence not explanatory.

Trey Ideker (University of California, San Diego)
Although artificial neural networks capture a variety of human functions, their internal structures are hard to interpret. In the life sciences, extensive knowledge of cell biology provides an opportunity to design 'visible' neural networks (VNNs) which couple the model's inner workings to those of real systems. Here we develop D-Cell, a VNN embedded in the hierarchical structure of 2526 subsystems comprising a eukaryotic cell. Trained on several million genotypes, D-Cell simulates cellular growth nearly as accurately as laboratory observations. During simulation, genotypes induce patterns of subsystem activities, enabling in-silico investigations of the molecular mechanisms underlying each genotype-phenotype association.

William Bechtel (University of California, San Diego)
Machine-learning tools, including tools for analyzing network representations of big data in cancer research, do not seem to characterize mechanisms. By annotating the results with information from Gene Ontology (GO), researchers attempt to interpret networks in terms of knowledge about cell components and operations. Recently researchers have used machine-learning techniques to create an alternative ontology, NeXO, directly from data. Although on their surface, neither GO nor NeXO describe mechanisms, the hierarchical structured resulting from employing directed acyclic graphs provides the needed bridge. I will discuss how mechanistic knowledge embodied in these ontologies is informing network analyses of cancer data.

Sara Green (University of Copenhagen)
This paper explores whether mechanistic research can improve predictions based on correlation-based cancer biomarkers. I examine a case where a model based on the best-characterized genetic marker for neuroblastoma was improved through connections to a mechanistic model of a cell death-promoting signaling pathway. Inclusion of mechanistic information on an experimentally resolved feedback loop was shown to improve the ability to stratify patients according to treatment outcomes. The case illustrates how a dynamic approach to biomarkers can be facilitated through research on signaling pathways and network dynamics. Moreover, it illustrates important complementary benefits and limitations of mechanistic heuristics and machine-learning strategies.