Session Introduction

Kelly Regan, Zachary B. Abrams, Michael F. Sharpnack, Arunima Srivastava, Kun Huang, Nigam Shah, Philip R. O. PaynePSB 2016

The 2016 Pacific Symposium on Biocomputing (PSB) session on the "Discovery of Molecularly Targeted Therapies" marks a pivotal transition in biomedical research. Moving away from the traditional reductionist "one-gene, one-drug" paradigm, this session highlights the emergence of integrative, multi-scalar systems models. These models are designed to bridge the persistent gap between laboratory discovery and clinical application, transforming heterogeneous biological data into actionable medical hypotheses.

The Translational Gap and the Need for Integration

A significant "translational gap" exists where basic science research and clinical care remain poorly integrated. Despite the explosion of data regarding genes, phenotypes, and pharmacological agents, the field has lacked systematic tools to "forward engineer" treatment strategies for real-world patients. Traditional reductionism often isolates biological components, failing to account for the synergistic effects and complex interactions that characterize human disease. To address this, researchers must develop computationally tractable ways to reason across different scales—linking molecular mutations to clinical symptoms and, ultimately, to targeted therapies.

The Framework of Vertical Reasoning

The core methodology advocated in this session is "vertical reasoning," a framework based on the work of Blois and Tsafnat. This approach integrates biological data across multiple scales, from molecules to entire populations, through two primary components:

  1. Component Models: Decision-making models focused on specific scales, such as protein-protein interaction networks.
  2. Interchange Layers: "Vertical linkages" that define the connections between these scales, such as how a specific protein interaction manifests as a clinical phenotype.

Computational Methodologies for Multi-Scalar Modeling

Session contributors employed several advanced computational techniques to realize this vertical reasoning:

  • Factorization-based Models: Utilizing Non-negative Matrix Factorization (NMF) and collective pairwise classification (COPACAR), researchers traversed large-scale databases to predict novel drug-disease relationships.
  • Network-Theoretic Analysis: By employing "scale-free" network paradigms, researchers modeled biological systems as tripartite networks, linking environmental factors, biological pathways, and diseases.
  • Advanced Clustering Algorithms: Innovations such as inverse regularized Markov Clustering (irMCL) and attribute-aware Markov Clustering (airMCL) were introduced to integrate network structures with biological attributes like gene expression or patient demographics, making models more robust against data noise.
  • Differential Dependency (EDDY): The Evaluation of Differential DependencY (EDDY) method was highlighted for its ability to identify how gene dependencies shift across different disease conditions, such as cancer subtypes.

Data Repositories and Evidence

The research presented synthesized data from major biological and clinical repositories, including the Comparative Toxicogenomics Database (CTD) for chemical-gene-disease interactions, The Cancer Genome Atlas (TCGA) for glioblastoma multiforme genomic data, and the AIDS Clinical Trials Group (ACTG) alongside the Roadmap Epigenome Project for SNP-phenotype associations.

Key Insights: Category-Jumping and Knowledge Injection

A powerful concept emerging from this session is "category-jumping"—the ability of multi-relational models to generate hypotheses across heterogeneous entities. For instance, shared clinical symptoms can be shown to indicate shared underlying molecular interactions. Furthermore, the session emphasized "knowledge injection," where prior biological knowledge is integrated into regression-based studies rather than relying solely on unsupervised data mining. This approach allows for "environmental triangulation," mapping how specific environmental toxins (e.g., arsenic) disrupt specific signal transduction pathways.

Clinical Impact and Results

The application of these methods yielded high-impact findings:

  • Association Discovery: Verma et al. identified over 50,000 statistically significant associations across 26 phenotypes by filtering genomic variance.
  • Cancer Subtyping: Using the EDDY method, researchers discovered that mesenchymal glioblastoma subsets are primarily defined by metabolic changes, whereas proneural subsets are driven by AKT-ERK signaling changes.
  • Drug Repurposing: Factorization models demonstrated high precision in predicting drug-target interactions, facilitating the "repositioning" of existing drugs—a strategy that is significantly faster and more cost-effective than de novo drug development.

Challenges and Future Directions

While the shift toward vertical reasoning is promising, several hurdles remain. Current models, particularly those based on NMF, often struggle to distinguish between correlation and causality. Additionally, there is an inherent tension between the desire for holistic "multi-models" and the risk of creating overly complex systems that are difficult to interpret or validate.

The ultimate goal is the creation of a seamless "end-to-end" pipeline where in-silico hypotheses are automatically funneled into laboratory and clinical validation paradigms. The 2016 PSB session concludes that while vertical reasoning is an emerging field, it represents the most viable path toward data-driven, adaptive, and truly personalized healthcare.

Summary updated

BibTeX
@inproceedings{regan2016session,
  author = {Kelly Regan and Zachary B. Abrams and Michael F. Sharpnack and Arunima Srivastava and Kun Huang and Nigam Shah and Philip R. O. Payne},
  title = {{Session Introduction}},
  booktitle = {PSB},
  year = {2016},
}
Figure