Abstract
Cocaine addiction remains a global health crisis with no FDA-approved medications. The complexity of the brain's response to cocaine, primarily through the dopamine transporter (DAT) and its associated proteomic network, presents a significant challenge for drug discovery. This study introduces a proteome-informed machine learning (ML) platform that integrates protein-protein interaction (PPI) networks with deep learning to screen over 60,000 compounds. By evaluating potency, safety (hERG), and ADMET criteria, the researchers identified 141 potential targets and two "nearly optimal" lead compounds.
Introduction
Despite decades of research, the "one drug, one target" approach has failed to produce effective treatments for cocaine addiction. Cocaine's primary mechanism involves blocking the dopamine transporter (DAT), but this action triggers a cascade of effects across a vast network of proteins. Traditional drug discovery is hindered by polypharmacology—where drugs hit unintended targets causing side effects like heart arrhythmias via hERG channel inhibition—and the prohibitive cost of experimental testing across the entire proteome. To address these hurdles, a computational framework was developed to navigate the "proteome" rather than isolated targets.
Methods
The researchers implemented a multi-stage pipeline designed to identify potent and safe inhibitors.
PPI Network and Target Identification
Using the STRING database, the team mapped the neighborhood of the dopamine transporter, identifying a global network of 141 proteins. A core network of 11 critical proteins, including VMAT2, alpha-synuclein, and various dopamine receptors, was prioritized for modeling.
Representation Learning and Predictive Modeling
To represent molecular structures, a bidirectional Long Short-Term Memory (Bi-LSTM) autoencoder was trained on 104 million molecules from Zinc15, PubChem, and ChEMBL27. This model converts SMILES strings into 512-dimensional Latent-Vector Fingerprints (LV-FPs). Subsequently, 32 separate Gradient Boosting Decision Tree (GBDT) models were constructed to predict binding affinities (BA) across the identified targets.
Multi-Objective Screening
A rigorous filtering process involving 38 criteria was applied to over 60,000 candidates. This included high potency for DAT, low affinity for the hERG safety target, cross-target analysis to minimize side effects, and favorable ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) profiles.
Results
The predictive models demonstrated high reliability, with most achieving Pearson correlation coefficients ($R$) exceeding 0.80 in 10-fold cross-validation.
Lead Compound Identification
The screening process identified two "nearly optimal" leads:
- CHEMBL558086: A potent DAT inhibitor with excellent ADMET scores.
- CHEMBL3764304: Originally from a D2R dataset, this compound showed significant repurposing potential as a DAT inhibitor with minimal hERG risk.
Insights into "Atypical" Inhibitors and Safety
The platform highlighted the advantage of "atypical" DAT inhibitors, which stabilize an inward-facing conformation of the transporter, effectively blocking cocaine binding even with moderate affinity. Furthermore, the model provided a mechanistic explanation for the failure of experimental drugs like Ibogaine, predicting high risks for hERG and other proteins associated with severe adverse effects.
Discussion
This proteome-informed approach reveals why many previous candidates failed: they lacked the necessary cross-target safety profiles. While the identified leads are promising, they remain in silico predictions requiring in vivo validation. Future work will involve expanding the network scope beyond the initial 141 proteins to capture even more subtle proteomic interactions and potential side effects.
Summary updated