Predicting hERG Cardiotoxicity with a Quantum Kernel

Student Researcher / Intern · Yale Chemistry · Batista Lab · New Haven, Connecticut

Jun 2026 to presentIn progress

Interning at the Batista Lab in Yale Chemistry on cardiotoxicity, one of the biggest safety problems in drug development: a drug that blocks the heart’s hERG potassium channel can cause dangerous heart rhythms, and that risk has pulled otherwise effective medicines off the market.

A glimpse of the work: a two-stage hybrid quantum-classical pipeline, my own take on CardioGenAI, that screens molecules for hERG blockade straight from their structure and then sorts the flagged ones into structural families. It is done with Nvidia, Quantinuum, and QuantumCT, and is still in progress.

Lab or program: Batista Lab, Yale Chemistry
Topic: Predicting hERG cardiotoxicity from molecular structure
Methods: ChemBERTa embeddings, supervised UMAP, a 10-qubit projected quantum kernel, quantum support vector classifier (QSVC), spectral clustering
Data: A balanced 2,000-molecule set from the hERGCentral screen (Therapeutics Data Commons), and patch-clamp (PC) and fluorescence polarization (FP) percent inhibition fitted to pIC50 values
Collaborators: Nvidia, Quantinuum, QuantumCT
Status: In progress

A look at the work

Embedding each molecule with a pretrained ChemBERTa chemical language model, then compressing it with supervised UMAP to ten features that drive the rotations of a ten-qubit quantum circuit.
Encoding those features in a parameterized quantum feature map and screening molecules with a quantum support vector classifier trained on the projected quantum kernel.
Grouping the flagged molecules into structural sub-groups with spectral clustering over the same quantum kernel.
Keeping every comparison fair: hyperparameters chosen on a leakage-free validation split, and the quantum kernel benchmarked against classical kernels and the Tiny Recursion Model on identical inputs.
Fitting patch-clamp (PC) and fluorescence polarization (FP) percent inhibition to pIC50 values.

Results

A validation ROC-AUC of 0.786, and 0.745 ± 0.015 on the test set over ten repeated splits.
On identical ten-dimensional features, the quantum kernel matches or slightly exceeds classical kernels, and its validation score stays competitive with the Tiny Recursion Model.
On a panel of marketed drugs, the pipeline flagged all four true blockers, including sotalol, a well-known hard case for structure-based hERG predictors.
The 1,010 flagged molecules split into three structural sub-groups, separated mainly by polarity and hydrogen bonding.

Quantum machine learning · Quantum kernels · hERG cardiotoxicity · Drug discovery · CardioGenAI

Batista Lab

Questions about this work? shashish.v.vasireddi@gmail.com