CCMI v2.0 Project 3: From Networks and Structures to Hierarchical Whole-Cell Models of Cancer Project Leads: Trey Ideker and Andrej Sali; Co-Investigators: Emma Lundberg, Jennifer Grandis, J. Silvio Gutkind, and Laura van ’t Veer SUMMARY One of the striking discoveries of the cancer genome projects is that each tumor presents a unique set of genetic mutations and molecular alterations. To understand how these alterations give rise to cancer and treatment outcomes, the Cancer Cell Map Initiative (CCMI) has launched systematic efforts to map the physical and functional architecture of tumor cells, capturing the molecular components and pathways on which cancer mutations converge. While parts of this effort are experimental, this Project 3 presents the central computational framework. A first computational theme concerns methods to assemble the structure of the multiscale tumor cell map. Aim 1 focuses on creating 3D models of cancer-associated protein complexes. It will apply established methods of integrative structural biology to data from other projects, including cryo-electron microscopy (cryo-EM), affinity purification mass spectrometry (AP-MS), cross-linking mass spectrometry (XL-MS), and genetic interaction datasets. Initial efforts will focus on PIK3CA-HER3 and mTOR complexes, identified in previous work by the CCMI, then move to new protein complexes identified by our ongoing mapping activities. Aim 2 focuses on mapping tumor cellular components at scales at and above the protein complex, extending to larger cellular components, compartments, and organelles. It will expand on a compelling proof-of-concept for creating an unbiased hierarchical map of human cell components by integration of AP-MS data with protein distribution data from immunofluorescence confocal images. These whole-cell maps will be analyzed to reveal specific cellular components under mutational selection in breast, head-and-neck, and lung cancers. A second computational theme concerns methods to integrate tumor cell maps with functional analysis and predictive medicine. Aim 3 uses the maps to build interpretable deep learning systems for prediction of drug responses. This aim draws from our previous work to establish “visible” learning models (DCell and DrugCell), which are not black boxes but have internal organization determined by prior knowledge of biological structure. We will construct such models from CCMI tumor cell maps, incorporating key improvements over our first- generation pilots. Finally, Aim 4 will use visible deep learning systems alongside other machine learning models to design and evaluate combinatorial biomarkers for breast, head-and-neck, and lung tumors in the patient- derived xenograft (PDX) and clinical settings. Clinical samples and data will be drawn from molecular tumor boards and the I-SPY breast cancer trial. PDX and clinical data will be used for further optimization of our predictive models using nascent techniques from transfer learning. T...