Pre-trained molecular representations enable antimicrobial discovery

Roberto Olayo-Alarcón, Martin K. Amstalden, Annamaria Zannoni, Medina Bajramovic, Cynthia M. Sharma, Ana Rita Brochado, Mina Rezaei, Christian L. Müller

The "silent pandemic" of antimicrobial resistance (AMR) is one of the greatest threats to modern medicine. While deep learning has shown promise in accelerating the discovery of new antibiotics, most models are data-hungry, requiring massive sets of experimentally labeled data that are expensive and slow to produce.

A new paper published in Nature Communications introduces MolE (Molecular representation through redundancy reduced Embedding), a framework that learns the "grammar" of chemistry from unlabeled data to predict antibiotic activity with surprising efficiency.

TL;DR

Researchers developed a lightweight, two-stage deep learning framework called MolE that learns molecular features from 100,000 unlabeled structures. By fine-tuning this model on small datasets of drug-bacteria interactions, they successfully identified three human-targeted drugs that inhibit the growth of Staphylococcus aureus, demonstrating a cost-effective path for antibiotic repurposing.

What Problem Does This Paper Address?

The traditional "search and find" method for antibiotics is failing. Screening large chemical libraries has a dismal hit rate (1–3%), and experimental validation is a bottleneck.

Current AI approaches often rely on supervised learning, where a model is trained on molecules already known to be "active" or "inactive." The problem? We don't have enough labeled data for many pathogens. While "foundation models" trained on millions of molecules exist, they are computationally heavy. The authors sought a "lightweight" strategy that could learn meaningful chemical representations from unlabeled data and generalize well to new, unseen bacteria.

Methods

The authors employed a two-stage strategy:

  1. Self-Supervised Pre-training (MolE): Using a Graph Isomorphism Network (GIN), the model was trained on 100,000 unlabeled molecules from PubChem. Instead of the common "contrastive learning" (which compares positive and negative examples), they used the Barlow Twins objective. This technique forces the model to produce similar embeddings for two slightly different versions (augmentations) of the same molecule while minimizing redundancy between the components of those embeddings.
  2. Supervised Downstream Prediction: The pre-trained MolE representations were fed into an XGBoost classifier. This model was trained on a publicly available dataset of ~1,200 drugs tested against 40 different gut bacterial strains to predict "Antimicrobial Potential" (AP) scores.

Datasets

  • Pre-training: 100,000 unique chemical structures sampled from PubChem.
  • Benchmarking: The MoleculeNet collection (including Tox21, ClinTox, and HIV datasets) was used to compare MolE against state-of-the-art models.
  • Activity Training: A dataset from Maier et al. (2018) featuring 1,197 marketed drugs and their effects on 40 representative human gut microbes.
  • Discovery Library: An orthogonal library from MedChemExpress (MCE) containing 2,320 compounds (FDA-approved drugs and metabolites) not seen during training.

Key Insights and Ideas

  • Small Data, Big Impact: Unlike other molecular models that require millions of structures, MolE achieved competitive results with just 100,000. This suggests that the quality of the self-supervised objective (Barlow Twins) is more important than the sheer quantity of data.
  • Emergent Chemical Logic: Without being told what a functional group is, the model’s latent space naturally clustered molecules with similar features, such as naphthalenes or amino acid chains.
  • Cross-Species Generalization: The model was trained primarily on gut commensals but successfully predicted activity against S. aureus, a major human pathogen. This indicates the model is learning fundamental "compound-microbe" interaction rules rather than just memorizing specific species.

Key Results

  • Superior Benchmarking: MolE outperformed standard chemical fingerprints (ECFP4) and several complex graph neural networks on 4 out of 6 MoleculeNet classification tasks. On the ClinTox benchmark, it achieved a high ROC-AUC of 92.85%.
  • Discovery of Non-Antibiotic Inhibitors: The model identified 158 non-antibiotic compounds with high antimicrobial potential.
  • Experimental Validation: The team tested six predicted compounds in the lab. Three human-targeted drugs—Elvitegravir (an HIV integrase inhibitor), Opicapone (used for Parkinson’s), and Ebastine (an antihistamine)—were confirmed to inhibit the growth of S. aureus. Elvitegravir was particularly potent, with a Minimum Inhibitory Concentration (MIC) of 8 μg/mL.

Limitations and Future Work

While the model is excellent at predicting if a compound will inhibit growth, it does not yet explain how (the Mechanism of Action). The authors also note that the current framework focuses on growth inhibition; future iterations could be refined to distinguish between bacteriostatic (growth-stopping) and bactericidal (killing) effects. Additionally, incorporating more specific biological features of the target bacteria (like membrane composition or metabolic pathways) could further improve the specificity of the predictions.

Reference: Olayo-Alarcon, R., et al. (2025). Pre-trained molecular representations enable antimicrobial discovery. Nature Communications, 16:3420. https://doi.org/10.1038/s41467-025-58804-4

Summary updated

Abstract

The rise in antimicrobial resistance poses a worldwide threat, reducing the efficacy of common antibiotics. Determining the antimicrobial activity of new chemical compounds through experimental methods remains time-consuming and costly. While compound-centric deep learning models promise to accelerate this search and prioritization process, current strategies require large amounts of custom training data. Here, we introduce a lightweight computational strategy for antimicrobial discovery that builds on MolE (Molecular representation through redundancy reduced Embedding), a self-supervised deep learning framework that leverages unlabeled chemical structures to learn task-independent molecular representations. By combining MolE representation learning with available, experimentally validated compound-bacteria activity data, we design a general predictive model that enables assessing compounds with respect to their antimicrobial potential. Our model correctly identifies recent growth-inhibitory compounds that are structurally distinct from current antibiotics. Using this approach, we discover de novo, and experimentally confirm, three human-targeted drugs as growth inhibitors of Staphylococcus aureus. This framework offers a viable, cost-effective strategy to accelerate antibiotic discovery.
BibTeX
@inproceedings{olayoalarcon2025pre,
  author = {Roberto Olayo-Alarcón and Martin K. Amstalden and Annamaria Zannoni and Medina Bajramovic and Cynthia M. Sharma and Ana Rita Brochado and Mina Rezaei and Christian L. Müller},
  title = {{Pre-trained molecular representations enable antimicrobial discovery}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-58804-4},
}
Figure