Kinome profiling allows examination and prediction of kinase inhibitor cardiotoxicity

This study demonstrates that integrating proteomic kinome profiling with machine learning enables the prediction of kinase inhibitor cardiotoxicity by identifying specific off-target kinase engagements, such as PDGFRB and EGFR, that drive adverse cardiac events.

The Gist

Imagine you are a chef trying to create a new, super-powerful spice blend to cure a specific type of cancer. This spice blend is made of "Kinase Inhibitors" (KIs). These are drugs designed to turn off specific "bad switches" (kinases) inside cancer cells that make them grow out of control.

For years, scientists thought the main problem with these spice blends was that they were too "messy." They believed that if a drug hit too many switches at once (a "promiscuous" drug), it would cause chaos in the heart, leading to heart failure or other cardiac issues. The logic was: More hits = More heart trouble.

However, this new study by Jimmy Tabet and his team at UNC Chapel Hill says: "Not so fast!"

Here is the story of what they found, explained simply:

1. The "Messy Chef" Myth

The researchers gathered data on 44 different cancer drugs. They wanted to see if the "messier" drugs (those hitting many switches) caused more heart problems than the "cleaner" drugs (those hitting only one or two).

The Surprise: They found no connection.

• Analogy: Imagine two cars. Car A has a very messy driver who hits 50 potholes. Car B has a very precise driver who hits only 2 potholes. You might think Car A is more likely to break down. But in this study, the "messy" cars didn't break down any more often than the "precise" ones.
• The Real Culprit: It's not how many switches the drug hits; it's which specific switches it hits. Even a "clean" drug can wreck your heart if it accidentally hits the wrong switch, even if it's just one.

2. The "Ghost Targets"

The FDA (the government agency that approves drugs) usually lists the main target a drug is supposed to hit. But the researchers used a high-tech microscope called Proteomics (think of it as a super-sensitive metal detector) to see what the drugs were actually hitting in the lab.

The Discovery: The drugs were hitting dozens of "ghost targets" (off-targets) that the FDA didn't even know about.

• Analogy: You hire a security guard (the drug) to stop a specific thief (the cancer kinase). The FDA says, "Great, he stopped the thief!" But the high-tech metal detector reveals the guard also accidentally tripped the fire alarm, knocked over a vase, and locked the janitor in the closet.
• The Heart Connection: The study found that specific "ghost targets" (like PDGFRB, EGFR, and MEK1/2) were the ones actually causing the heart trouble. These are switches that, when turned off, confuse the heart's electrical system or weaken its muscle.

3. The "Heart Network"

The researchers realized that these "ghost targets" aren't just random; they are part of a connected web, like a city's power grid.

• Analogy: If you cut one wire in a house, the lights might flicker. But if you cut a wire that connects the kitchen, the bedroom, and the hallway all at once, the whole house goes dark.
• The study showed that cardiotoxic drugs disrupt a specific network of signals that the heart needs to beat and pump blood. It's not just one broken part; it's a coordinated failure of the whole system.

4. The Crystal Ball (Machine Learning)

Since they couldn't predict heart trouble just by looking at how "messy" a drug was, the team built a Crystal Ball using Artificial Intelligence (Machine Learning).

• How it works: They fed the AI the "fingerprint" of every drug (a list of every single switch it hits) and told it which drugs caused heart problems.
• The Result: The AI learned to predict heart trouble with 66% to 84% accuracy.
• Why it matters: Before, we had to wait until patients got sick to know a drug was dangerous. Now, we can test a new drug in the lab, scan its "fingerprint," and the AI can say, "Hey, this drug hits the 'Heart-Breaker' switch. Don't use it, or fix it."

The Big Takeaway

This study changes the rules of the game for making cancer drugs:

1. Don't just aim for "clean" drugs. A drug that hits only one target can still be deadly to the heart if that one target is the wrong one.
2. Watch out for the "Ghosts." We need to look deeper than the FDA's official list to see what else a drug is touching.
3. Use the Crystal Ball. By combining lab data with AI, we can design safer drugs before they ever reach a patient, saving lives and preventing heart failure in cancer survivors.

In short: It's not about how many targets a drug hits; it's about which targets it hits. And thanks to this new "fingerprint" method, we can finally spot the dangerous ones before they hurt our hearts.

Technical Summary

1. Problem Statement

Kinase inhibitors (KIs) have revolutionized cancer treatment, yet they carry a significant risk of cardiotoxicity, manifesting as heart failure, hypertension, QT interval prolongation, arrhythmias, and myocardial infarction.

• Current Limitations: Existing methods to predict KI-induced cardiac adverse events (CAEs) rely on chemical structure, transcriptomics, or FDA-annotated primary targets. These approaches fail to capture off-target kinase engagement, which is a primary driver of toxicity due to the high conservation of ATP-binding pockets across the kinome.
• The Gap: There is a lack of systematic integration between empirical, proteome-wide binding data and clinical CAE incidence rates. Furthermore, the hypothesis that "promiscuous" (broadly acting) inhibitors are inherently more toxic than selective ones has not been rigorously tested using quantitative kinome data.

2. Methodology

The authors developed a data-driven framework integrating proteomics and machine learning (ML) to predict cardiotoxicity.

A. Data Curation

• Kinome Profiles: Utilized unbiased chemoproteomic data (Kinobeads/MIB-MS) from Klaeger et al. (39) covering 229 drugs. This provided quantitative binding profiles (percent inhibition) across nearly the entire kinome at 1 µM concentration.
• Clinical CAE Data: Curated incidence rates for six specific CAEs (QT prolongation, heart failure, ischemia/MI, arrhythmia, LV dysfunction, hypertension) for FDA-approved KIs, based on a review by Jin et al. (23) and supplemented with negative controls.
• Final Dataset: Merged 44 FDA-approved KIs with both proteomic binding profiles and documented CAE incidence rates.
  • Labeling: Incidence rates were binarized: "No reported event" (negative class) vs. "Reported event" or ">10% incidence" (positive class).

B. Analytical Approach

1. Promiscuity Analysis: Calculated a "promiscuity score" (number of kinases inhibited >80%) for each drug to test if broad inhibition correlates with CAE incidence.
2. Frequency & Network Analysis:
   • Identified kinases recurrently inhibited by CAE-linked drugs.
   • Mapped these kinases to protein-protein interaction networks (HIPPIE) and KEGG pathways to understand biological context.
3. Machine Learning Modeling:
   • Trained six separate models (one per CAE) to predict the binary CAE outcome based on the full kinome inhibition profile.
   • Algorithms Tested: Logistic Regression, Random Forest, XGBoost, LightGBM, Multi-Layer Perceptron (MLP), and TabPFN (transformer-based).
   • Validation: 5-fold cross-validation with hyperparameter tuning (randomized search).
   • Interpretability: Used SHAP (SHapley Additive exPlanations) values to identify which specific kinase features drove the model's predictions.

3. Key Results

A. Promiscuity is Not a Predictor

• Contrary to the intuitive assumption that "dirtier" (more promiscuous) drugs are more toxic, kinase promiscuity scores showed no significant statistical difference between cardiotoxic and non-cardiotoxic drugs across five of the six CAE endpoints.
• This suggests that cardiotoxicity is driven by specific kinase engagement rather than the sheer number of off-targets.

B. Identification of Critical Off-Targets

• Frequency analysis revealed that specific kinases are recurrently inhibited in CAE-linked compounds, even though they are rarely the FDA-annotated primary targets.
• Key Off-Targets: RET, PDGFRB, and DDR1 were frequently inhibited across multiple CAE categories.
• Network Insights: Pathway enrichment analysis showed that these frequently inhibited kinases form dense subnetworks involved in cardiac-relevant processes, such as axon guidance (linked to hypertension) and actin cytoskeleton organization (linked to LV dysfunction).

C. Machine Learning Performance

• Predictive Accuracy: ML models achieved 66–84% cross-validated accuracy and ROC-AUC scores of 0.75–0.80 across the six CAE endpoints.
  • Best Performance: Hypertension (84% accuracy).
  • Lowest Performance: QT interval prolongation (66% accuracy).
• SHAP Feature Importance: The models identified distinct, biologically coherent kinase signatures for each CAE:
  • Highly Predictive Kinases: PDGFRB, EGFR, and MEK1/2 (known to be involved in myocardial stress and hypertrophic remodeling).
  • Novel Predictors: Kinases less traditionally associated with cardiac toxicity, such as RET, LCK, LIMK1, TEC, TNK1, and FLT3, emerged as highly influential features, suggesting underappreciated roles in specific cardiotoxic phenotypes.

4. Key Contributions

1. First Systematic Integration: This is the first study to directly link empirical, proteome-wide kinase binding data with clinical CAE incidence rates across the FDA-approved KI landscape.
2. Refutation of Promiscuity Hypothesis: Demonstrated that broad inhibition is not the primary driver of cardiotoxicity; rather, the identity of specific off-target kinases is the critical factor.
3. Mechanistic Framework: Provided a systems-level model showing that cardiotoxicity arises from the coordinated disruption of interconnected signaling networks (e.g., cytoskeletal and vascular tone pathways) rather than single-node inhibition.
4. Predictive Tool: Established a machine learning framework capable of predicting CAE risk early in drug development using only kinome binding profiles, achieving meaningful accuracy without prior clinical data.

5. Significance and Clinical Implications

• Drug Design: The findings suggest that drug developers should focus on avoiding specific off-target kinases (like RET or PDGFRB) rather than simply striving for general selectivity. This enables "off-target-aware" rational drug design.
• Preclinical Screening: The proteomic profiling + ML approach offers a scalable, reproducible tool to flag high-risk compounds early in the pipeline, potentially reducing clinical trial attrition (currently estimated at 14–45% due to cardiotoxicity).
• Personalized Medicine: As the cancer population ages and accumulates cardiovascular comorbidities, these tools can help stratify patients and select therapies that balance antitumor efficacy with minimized cardiac risk.

Conclusion: The study establishes that proteomic kinome profiling combined with machine learning provides a mechanistically grounded, predictive framework for assessing kinase inhibitor cardiotoxicity, shifting the paradigm from "selectivity vs. promiscuity" to "specific target engagement."