Cancer drug discovery as a low rank tensor completion problem

Vasanth S. Murali, Didem Ağaç Çobanoğlu, Michael Hsieh, Meyer Zinn, Venkat S. Malladi, Jonathan Gesell, Noelle S. Williams, Erik S. Welf, Ganesh V. Raj, Murat Can ÇobanoğlubioRxiv 2021 · DOI

TL;DR

This paper proposes a novel approach to cancer drug discovery by framing it as a low-rank tensor completion problem. By integrating diverse public datasets into a drug-target-disease association tensor, they can efficiently identify promising drug candidates with significantly fewer experiments than traditional methods. This computational approach, implemented in their web server "Cancer Vulnerability Explorer" (CaVu), has been experimentally validated for melanoma and breast cancer, demonstrating its potential to accelerate the discovery of effective cancer therapies.

What Problem Does This Paper Address?

The discovery of new cancer drugs is a notoriously challenging and expensive process. Cancer's inherent heterogeneity means that a single drug rarely works for all patients, necessitating the development of targeted therapies for specific cancer subtypes. However, traditional drug discovery methods, such as high-throughput screening, are often inefficient and costly, leading to a decline in productivity. This paper addresses the critical need for more efficient and effective methods to identify promising drug candidates that can be rapidly translated into clinical treatments. They aim to overcome the limitations of brute-force screening by computationally prioritizing a small number of exceptionally well-suited drug candidates for experimental validation.

Methods

The core of this paper's methodology lies in low-rank tensor completion. They construct a drug-target-disease association tensor by integrating various public datasets. This tensor represents the complex relationships between drugs, their protein targets, and specific cancer diseases (defined by mutations and tissue lineage).

The key steps include:

  1. Defining Cancer Subtypes: Utilizing data from the Cancer Cell Line Encyclopedia (CCLE) to define subtypes based on mutations and tissue lineage.
  2. Identifying Vulnerabilities: Employing statistical tests (like hypergeometric enrichment) on perturbation datasets (e.g., DRIVE, Achilles) to find genes whose knockdown is specifically lethal to cancer cells within a defined subtype.
  3. Linking Vulnerabilities to Drugs: Using databases like STITCH to connect identified vulnerabilities to small molecules (chemicals) that interact with their protein targets.
  4. Drug-like Filtering: Filtering these chemicals for drug-likeness using the DrugBank database.
  5. Tensor Construction and Completion: Assembling the identified drug-target-disease associations into a third-order tensor. This tensor is then completed using CANDECOMP/PARAFAC (CP) decomposition, a low-rank tensor completion technique. This allows them to infer missing entries in the sparse tensor, predicting novel drug-target-disease associations.
  6. Web Server Development: Creating the Cancer Vulnerability Explorer (CaVu) web server to make this methodology accessible to researchers.

Datasets

The study leverages several publicly available datasets:

  • Cancer Cell Line Encyclopedia (CCLE): Used for defining cancer subtypes based on mutations and tissue lineage.
  • Project DRIVE and Achilles: Perturbation datasets used for identifying gene vulnerabilities specific to cancer subtypes.
  • STITCH: A database for linking proteins to chemicals (drug-target interactions).
  • DrugBank: Used to filter chemicals for drug-likeness by identifying those similar to approved drugs.
  • Ensembl: Used to map protein targets to genes.

The paper does not introduce a new benchmark dataset but rather integrates existing ones into a novel tensor framework.

Key Insights and Ideas

  • Cancer Drug Discovery as a Tensor Problem: The novel framing of cancer drug discovery as a low-rank tensor completion problem is a significant conceptual shift, allowing for the integration of diverse data types into a unified model.
  • Efficient Prioritization: The tensor completion approach dramatically reduces the number of experiments needed by computationally prioritizing the most promising drug candidates, moving away from brute-force screening.
  • Data Integration for Synergistic Discovery: By combining data from different sources (genomics, proteomics, chemical interactions), the model can uncover complex relationships that might be missed by analyzing each dataset in isolation.
  • Experimental Validation of Computational Predictions: The successful experimental validation of novel drug candidates (SU11652 for melanoma and TC-E 5008 for breast cancer) provides strong evidence for the predictive power of their approach.
  • Mechanism Prediction: The tensor framework can also be used to predict the mechanism of action of drugs by identifying their likely targets.

Key Results

  • Identification of Novel Drug Candidates: The CaVu platform successfully identified SU11652 as a promising drug candidate for melanoma, outperforming existing treatments in some in vitro and in vivo models. Similarly, TC-E 5008 was identified as a potential therapeutic for ER+ breast cancer.
  • High Predictive Power: Computational simulations demonstrated that their low-rank tensor completion method achieved a significantly higher "hit rate" (over 20%) compared to random guessing, with a learning factor of 3,000-4,000 fold improvement.
  • Robustness of Predictions: The predictions were shown to be robust, with consistent results achieved when using 16 or more independent initializations for the CP decomposition.
  • Identification of Known Vulnerabilities: The model accurately identified known lineage-specific vulnerabilities for various cancer types, validating its ability to recover established biological insights.
  • Mechanism of Action Prediction: The study demonstrated that the tensor completion approach could predict the likely target of SU11652 as PAK1, which was then experimentally validated.

Limitations and Future Work

  • Data Limitations: The accuracy of the CaVu predictions is inherently limited by the quality and completeness of the underlying datasets (STITCH, DRIVE).
  • Computational Cost: While more efficient than brute-force screening, tensor completion can still be computationally intensive, especially for very large tensors.
  • Extending to Predicted Interactions: Future work aims to incorporate predicted drug-target interactions, not just experimentally validated ones, to further increase the actionable insights.
  • Mechanistic Queries: The authors envision building more sophisticated computational tools within CaVu to answer complex mechanistic questions about drug action.
  • Incorporating Gene Expression Data: Expanding the framework to include gene expression data could enable vulnerability identification for cancer subtypes where knockdown data is scarce.
  • Broader Applicability: The goal is to make CaVu a comprehensive platform for bridging bioinformatics and cancer research, facilitating more efficient drug discovery.

Summary updated

Abstract

Abstract The heterogeneity of cancer necessitates developing a multitude of targeted therapies. We propose the view that cancer drug discovery is a low rank tensor completion problem. We implement this vision by using heterogeneous public data to construct a tensor of drug-target-disease associations. We show the validity of this approach computationally by simulations, and experimentally by testing drug candidates. Specifically, we show that a novel drug candidate, SU11652, controls melanoma tumor growth, including BRAF WT melanoma. Independently, we show that another molecule, TC-E 5008, controls tumor proliferation on ex vivo ER+ human breast cancer. Most importantly, we identify these chemicals with only a few computationally selected experiments as opposed to brute-force screens. The efficiency of our approach enables use of ex vivo human tumor assays as a primary screening tool. We provide a web server, the Cancer Vulnerability Explorer (accessible at https://cavu.biohpc.swmed.edu ), to facilitate the use of our methodology.
BibTeX
@misc{murali2021cancer,
  author = {Vasanth S. Murali and Didem Ağaç Çobanoğlu and Michael Hsieh and Meyer Zinn and Venkat S. Malladi and Jonathan Gesell and Noelle S. Williams and Erik S. Welf and Ganesh V. Raj and Murat Can Çobanoğlu},
  title = {{Cancer drug discovery as a low rank tensor completion problem}},
  year = {2021},
  doi = {10.1101/2021.03.08.434311},
  howpublished = {bioRxiv},
}
Figure