# A Machine Learning-Guided Approach for Identifying Potential HCAR1 Antagonists in Lactate-Driven Cancers

## TL;DR

This study uses machine learning to identify potential HCAR1 antagonists, which could help treat lactate-driven cancers by targeting a receptor involved in tumor progression and drug resistance.

## Contribution

A novel SVM model and framework integrating conformational docking and SHAP analysis to identify and interpret HCAR1 antagonists.

## Key findings

- The SVM model achieved 79.3% accuracy and an AUC of 0.94 in classifying HCAR1 ligands as agonists or antagonists.
- Polar, rigid, and aromatic substructures were identified as key features for HCAR1 antagonist selectivity.
- Three compounds, including two FDA-approved drugs, showed promising antagonistic potential based on model predictions and docking scores.

## Abstract

GPR81 (HCAR1) is a lactate-sensing G protein-coupled
receptor (GPCR)
involved in tumor progression, immune evasion, and therapeutic resistance
across various cancers. Despite their clinical relevance and druggable
nature, selective HCAR1 antagonists have yet to be identified. This
study aimed to construct a statistically significant Support Vector
Machine (SVM) model for binary classification (agonists versus antagonists)
of HCAR1’s potential ligands and the prioritization of molecular
substructures driving antagonism and receptor selectivity. An SVM
model was trained on 144 ligands (66 agonists, 78 antagonists), listed
in the IUPHAR/BPS Guide to Pharmacology, from 12 structurally related
Class A GPCRs (HCAR1, HCAR2, HCAR3, OXER1, GPR35, SUCNR1, P2Y2, MCHR1,
OPRD1, AGTR1, ADORA2A, and ADRA1A). Their ligands were encoded using
physicochemical descriptors, 2048-bit ECFP4 fingerprints, and ΔAffinity
scores from molecular docking to active and inactive receptor conformations.
The data set was split 80/20 for training and testing, respectively,
with hyperparameters (C, γ) being optimized
via 5-fold cross-validation. SHAP analysis was performed for feature
interpretation. The final SVM model achieved a test set accuracy of
79.3%, with a sensitivity of 69.2% and specificity of 87.5%. The ROC
analysis yielded an AUC of 0.94, while bootstrapping confirmed robust
performance with a mean AUC of 0.874 and a 95% confidence interval
[0.711, 1.000]. SHAP analysis highlighted polar, rigid, and aromatic
substructures as selectivity-driving features. We applied the model
to screen 3,377 compounds from natural products, synthetic libraries,
and FDA-approved drugs, prioritizing potential HCAR1 ligands with
antagonist-like features. Based on ΔAffinity, off-target scores,
and prediction confidence, Ketanserin, Cryptopyranmoscatone A1 diacetate,
and Cefuroxime emerged as reference ligands with promising antagonistic
potential, two of which are FDA-approved drugs. Rather than representing
final hits, these molecules illustrate how structural and electronic
features can favor the stabilization of inactive states in HCAR1.
Overall, this work presents a proof-of-concept framework that integrates
conformational docking, machine learning, and substructure interpretation
to elucidate the chemical and structural determinants of HCAR1 antagonism.
The findings provide fragment-level insights that may guide future
bioisosteric and fragment-based design of selective antagonists for
lactate-driven tumors.