Elucidating Compound Mechanism of Action and Polypharmacology with a Large-scale Perturbational Profile Compendium

Lucas ZhongMing Hu, Eugene F. Douglass, Mikko Turunen, Sergey Pampou, Adina Grunn, Ron Realubit, Albert A. Antolín, Alexander L.e. Wang, Hai Li, Prem S. Subramaniam, Prabhjot S. Mundi, Charles Karan, Mariano J. Alvarez, Andrea Califano, Barry Honig, Brent R. StockwellbioRxiv 2023 · DOI

TL;DR

Researchers have unveiled PANACEA, a massive compendium of over 18,000 perturbational profiles mapping how 700+ oncology drugs affect the activity of ~6,500 proteins across 23 high-fidelity cancer models. This framework moves beyond simple "lock-and-key" drug targets to reveal complex, tissue-specific polypharmacology and identifies existing drugs that can inhibit "undruggable" oncoproteins like MYC.

What Problem Does This Paper Address?

The traditional view of a drug’s Mechanism of Action (MoA) is often too simplistic, focusing on a few high-affinity "on-target" binding proteins. In reality, a drug’s efficacy and toxicity are driven by polypharmacology: a complex web of low-affinity off-targets and tissue-specific secondary effectors.

Existing resources like the Connectivity Map (CMap) or LINCS have paved the way but suffer from three major flaws:

  1. Low Resolution: They often measure only a small subset of "landmark" genes (~1,000), missing the broader regulatory context.
  2. Poor Models: The cell lines used don't always faithfully represent the biology of actual human tumors.
  3. Confounding "Death Signatures": Testing drugs at high, toxic concentrations often triggers generic cell-stress and death pathways that mask the drug's specific mechanism.

Methods

The authors developed an integrative experimental and computational framework to build the PANACEA (PAN-cancer Activity-based Compound Efficacy Analysis) dataset:

  • PLATE-Seq: A low-cost, high-throughput RNA sequencing technology used to capture whole-genome transcriptional responses.
  • VIPER Algorithm: Instead of looking at mRNA levels alone, the authors used VIPER to infer protein activity. It does this by analyzing the expression of a protein’s downstream targets (its "regulon"), acting like a massively multiplexed reporter assay.
  • OncoMatch: A tool used to select 23 cell lines that are "high-fidelity" matches to specific human cancer subtypes, ensuring the results are clinically relevant.
  • Sublethal Dosing: To avoid the "death signature" problem, all perturbations were performed at the $IC_{20}$ concentration (the dose that inhibits only 20% of growth), ensuring the observed changes reflect the drug's primary mechanism rather than a cell's dying gasps.

Datasets

  • PANACEA: The core dataset containing 18,046 RNASeq profiles. It covers 700+ drugs (FDA-approved and late-stage experimental) across 23 cell lines representing aggressive tumor subtypes (e.g., glioblastoma, TNBC, neuroendocrine tumors).
  • Benchmarking: The study used LINCS (L1000) and CCLE for comparison and reproducibility testing.
  • Interactomes: Tissue-specific regulatory networks were reverse-engineered from large patient cohorts (like TCGA) using the ARACNe algorithm.

Key Insights and Ideas

  • Protein Activity is More Reliable: The study found that drug-induced changes in protein activity (inferred via VIPER) are significantly more reproducible across replicates than raw gene expression changes.
  • MoA is Context-Dependent: A drug’s mechanism can be "orthogonal" (completely different) across different tissues. For example, some PI3K inhibitors showed highly conserved MoA in one breast cancer line but not in another, despite both being the same subtype.
  • Canalization: The researchers observed that different drugs often "converge" on the same downstream regulatory bottlenecks. This explains how chemically unrelated drugs can produce similar therapeutic effects.
  • Inhibiting the "Undruggable": Many master regulators of cancer (like MYC or CTNNB1) lack traditional binding pockets for drugs. PANACEA identifies drugs that inhibit these targets indirectly by modulating the cellular network—either by degrading the protein, sequestering it away from the nucleus, or downregulating its gene.

Key Results

  • Drug Modules: The framework identified 30 distinct drug modules based on shared MoA. This revealed surprising "moonlighting" functions, such as the discovery that the leukemia drug Cladribine acts similarly to HDAC inhibitors.
  • Validation of "Undruggable" Targets: The authors experimentally validated several novel drug-target pairs using Western blots and confocal microscopy. Highlights include:
  • Amsacrine and Bortezomib inducing nuclear delocalization of CENPF.
  • Pimasertib sequestering MYC to the nuclear lamina.
  • Clinical Fidelity: MoA predictions made in PANACEA cell lines were shown to be faithfully recapitulated in Patient-Derived Xenograft (PDX) models in vivo.

Limitations and Future Work

  • Scope: While the framework is generalizable, this study focused primarily on oncology. The authors suggest extending this to neurological and viral diseases (it was already used to find 18 drugs that might counteract SARS-CoV-2).
  • Interactome Dependency: The accuracy of VIPER depends on high-quality "interactomes" (maps of which proteins regulate which genes). For rare diseases with little available patient data (like meningioma), the predictions were less robust.
  • Future Scaling: The authors envision using this as a "universal phenotypic assay" for drug discovery, allowing researchers to screen novel chemical libraries against thousands of protein activity signatures simultaneously.

Summary updated

Abstract

Drug Mechanism of Action (MoA) is generally represented as a small repertoire of tissue-independent, high-affinity binding targets. Yet, drug activity and polypharmacology are associated with a broad range of off-target and tissue-specific effector proteins. To address this challenge, we have generated a compendium of drug perturbation profiles for >700 oncology drugs in cell lines representing high-fidelity models of 23 aggressive tumor subtypes and developed an integrative computational framework for the proteome-wide assessment of drug-mediated, tissue-specific differential protein activity. This allows systematic, mechanism-based elucidation of tissue context-specific compound MoA and polypharmacology, including discovery and experimental validation of post-translationally-mediated inhibition of undruggable oncoproteins-such as MYC, CTNNB1, CENPF and UHRF1-as well as previously unknown inhibitors of MAPK, PI3K, and folate pathways, among others, as further suggested by protein structure analysis. This framework can help characterize the MoA of discovery compound libraries, thus supporting more systematic and quantitative approaches to precision oncology.
BibTeX
@misc{hu2023elucidating,
  author = {Lucas ZhongMing Hu and Eugene F. Douglass and Mikko Turunen and Sergey Pampou and Adina Grunn and Ron Realubit and Albert A. Antolín and Alexander L.E. Wang and Hai Li and Prem S. Subramaniam and Prabhjot S. Mundi and Charles Karan and Mariano J. Alvarez and Andrea Califano and Charles Karan and Mariano Alvarez and Barry Honig and Brent R. Stockwell and Andrea Califano},
  title = {{Elucidating Compound Mechanism of Action and Polypharmacology with a Large-scale Perturbational Profile Compendium}},
  year = {2023},
  doi = {10.1101/2023.10.08.561457},
  howpublished = {bioRxiv},
}
Figure