Pre-trained molecular representations enable antimicrobial discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Antimicrobial resistance (AMR) is a growing global health crisis, diminishing the effectiveness of existing antibiotics.
• Experimental screening of novel compounds for antimicrobial activity is resource-intensive and slow.
• Current deep learning models for drug discovery often require extensive custom training data.

Purpose of the Study:

• To introduce a lightweight computational strategy for accelerating antimicrobial discovery.
• To leverage self-supervised learning for generating molecular representations.
• To develop a predictive model for assessing antimicrobial potential using unlabeled chemical structures.

Main Methods:

• Utilized MolE (Molecular representation through redundancy reduced Embedding), a self-supervised deep learning framework.
• Combined MolE's learned molecular representations with existing compound-bacteria activity data.
• Developed a general predictive model for antimicrobial potential assessment.

Main Results:

• The model successfully identified novel growth-inhibitory compounds distinct from current antibiotics.
• Discovered and experimentally validated three existing human-targeted drugs as inhibitors of Staphylococcus aureus.
• Demonstrated the model's ability to assess compounds for antimicrobial potential.

Conclusions:

• The proposed computational framework offers a cost-effective and viable approach to accelerate antibiotic discovery.
• This strategy can aid in identifying new antimicrobial agents by leveraging unlabeled chemical data.
• The findings highlight the potential of self-supervised learning in addressing the AMR crisis.