Using the structural kinome to systematize kinase drug discovery

Zheng Zhao, Philip E. BournearXiv 2021 · DOI · arXiv

This paper, authored by Zheng Zhao and Philip E. Bourne from the University of Virginia, provides a comprehensive review of how structural biology and data science are transforming kinase drug discovery from a "trial-and-error" process into a systematic engineering discipline.

TL;DR

Designing drugs that target specific kinases is notoriously difficult because most kinases share a nearly identical "pocket" for binding ATP. This paper reviews how researchers are using Function-site Interaction Fingerprints (FsIFP) and structural bioinformatics to map the "structural kinome," enabling the design of highly selective or intentionally multi-targeted drugs while minimizing toxic side effects.

What Problem Does This Paper Address?

Kinases are the "switchboard" of the cell, regulating everything from growth to death. When they malfunction, they drive diseases like cancer and rheumatoid arthritis. While they are excellent drug targets, they present a massive engineering challenge: the human "kinome" contains 538 different kinases, and almost all of them have a highly conserved catalytic domain designed to bind ATP.

If a drug binds to the "wrong" kinase (off-target binding), it can lead to severe side effects like heart failure or cytopenia. Conversely, some diseases require hitting multiple specific kinases simultaneously (polypharmacology). The core problem is selectivity: how do we find the subtle structural "nuances" that distinguish one kinase from 537 others?

Methods

The authors highlight a systematic in silico (computer-based) approach to navigate this complexity:

  1. Structural Mapping: Utilizing the massive growth in the Protein Data Bank (PDB), which now contains over 5,200 structures for ~300 kinases.
  2. Binding Site Alignment: Using software like SMAP to align binding pockets independently of the overall protein sequence. This allows researchers to compare "apples to apples" at the atomic level.
  3. Function-site Interaction Fingerprints (FsIFP): This is the paper's centerpiece. It converts complex 3D protein-ligand interactions into a 7-bit 1D fingerprint. Each bit represents a specific type of interaction:
  4. Van der Waals forces
    • Aromatic interactions (face-to-face and edge-to-face)
    • Hydrogen bonds (donor and acceptor)
    • Electrostatic interactions (positive and negative)
  5. Inhibitor Classification: Systematizing drugs into Type-I (bind the active "DFG-in" state), Type-II (bind the inactive "DFG-out" state and reach into allosteric pockets), and Type-III (purely allosteric).

Datasets

The paper synthesizes data from several major repositories:

  • PDB (Protein Data Bank): 5,208 kinase structures.
  • KLIFS: A kinase-specific structural database.
  • ChEMBL Kinase SARfari: Contains ~530,000 data points across ~980 kinase targets.
  • Commercial Profiling Panels: Data from platforms like KINOMEscan (DiscoverRx) and HotSpot (Reaction Biology), which provide experimental validation for virtual screenings.
  • FDA Reference Set: A curated list of 63 FDA-approved small-molecule kinase inhibitors (as of Feb 2021).

Key Insights and Ideas

  • The Cysteine Map: One of the most actionable insights is the identification of 63 distinct spatial positions where accessible cysteines reside across the kinome. This provides a "roadmap" for designing covalent (irreversible) inhibitors, which are often more potent and selective.
  • From One-Target to Polypharmacology: The field is moving away from the "one-drug-one-target" dogma. By understanding the structural kinome, researchers can intentionally design drugs like Midostaurin that hit a specific spectrum of targets to treat complex diseases like AML.
  • The "Structural Biomedical Data Science" Shift: The authors argue that drug discovery is becoming a data science problem. By combining structural fingerprints with machine learning, we can predict how a new molecule will interact with the entire kinome before it ever enters a wet lab.

Key Results

  • Selectivity via Nuance: The paper demonstrates that even though kinase pockets look similar, "subtle differences" in the vicinity of the ATP-binding site (like the C-lobe or the G-rich loop) are sufficient to achieve high selectivity if mapped correctly.
  • Approved Landscape: As of early 2021, 63 kinase inhibitors were FDA-approved. The paper provides a vital lookup table (Table 2) classifying these by their binding mode and PDB availability, serving as a benchmark for future drug design.
  • Resistance Mechanisms: The authors highlight how structural data explains drug resistance. For example, the T790M "gatekeeper" mutation in EGFR increases the kinase's affinity for ATP, effectively "crowding out" drugs like Erlotinib.

Limitations and Future Work

  • The "Dark" Kinome: While we have structures for over 300 kinases, a significant portion of the 538 human kinases remains structurally uncharacterized ("undruggable" or "dark").
  • Acquired Resistance: Mutations like L1196M in ALK continue to outpace drug development. The authors suggest that "on-the-fly" FsIFP approaches will be necessary to predict and counteract these mutations rapidly.
  • Integration of ML: While the potential for Deep Learning is high, the paper notes that we need better integration between virtual structure-based screening and high-throughput phenotypic assays to truly "systematize" discovery.

Summary updated

Abstract

Kinase-targeted drug design is challenging. It requires designing inhibitors that can bind to specific kinases when all kinase catalytic domains share a common folding scaffold that binds ATP. Thus, obtaining the desired selectivity, given the whole human kinome, is a fundamental task during early-stage drug discovery. This begins with deciphering the kinase-ligand characteristics, analyzing the structure-activity relationships, and prioritizing the desired drug molecules across the whole kinome. Currently, there are more than 300 kinases with released PDB structures, which provides a substantial structural basis to gain these necessary insights. Here, we review in silico structure-based methods - notably, a function-site interaction fingerprint approach used in exploring the complete human kinome. In silico methods can be explored synergistically with multiple cell-based or protein-based assay platforms such as KINOMEscan. We conclude with new drug discovery opportunities associated with kinase signaling networks and using machine/deep learning techniques broadly referred to as structural biomedical data science.
BibTeX
@misc{zhao2021using,
  author = {Zheng Zhao and Philip E. Bourne},
  title = {{Using the structural kinome to systematize kinase drug discovery}},
  year = {2021},
  eprint = {2104.13146},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.13146},
}
Figure