A groundbreaking study shows AI could help predict potential antibiotics
The recent paper in Cell details the creation of AMPSphere, a catalog of 863,498 antimicrobial peptide sequences predicted using machine learning methods applied to genomic data. AMPSphere represents a significant advancement in leveraging AI to combat the growing antibiotic resistance crisis.
A recently published paper in the journal Cell details the creation and validation process for AMPSphere a catalog that comprises 863,498 sequences for antimicrobial peptides, which are small molecules with antimicrobial properties and have clinical, agricultural, and food preservation applications. Antimicrobial peptides cannot be reliably detected by standard methods, meaning they are often dismissed as candidates for new antibiotic drugs. Even if they are reliably detected, translation to a clinical setting is difficult, as many of these molecules are toxic, unstable, or costly to synthesize.
Solutions to the challenges surrounding new antimicrobial materials synthetization have become particularly pressing as antibiotic resistance takes millions of lives yearly. According to the WHO, antibiotic resistance could cause as many as 10 million deaths by 2050. To address these challenges, Prof. César de la Fuente and the Machine Biology Group have worked on applying AI and ML methods to antibiotic discovery, biology mining, and molecular de-extinction for several years. Contributing to the creation of the AMPSphere is the most recent fruit of the lab's labor, which has also yielded multiple publications in respected journals including Science, Cell, Cell Host Microbe, Nature Biomedical Engineering, and Nature Communications.
The AMPSphere resulted from applying machine learning methods to uncover AMPs in 63,410 publicly available metagenomes and 87,920 high-quality microbial genomes. After going through the list to remove redundant molecules, the researchers were left with 863,498 individual non-redundant sequences, clustered into 10,715 AMP families with at least 7 sequences each. From these, the research team synthesized a representative sample comprising 100 molecules in vitro and found that 79 were active, and 63 could kill at least one clinically significant pathogen. The researchers behind AMPSphere have opened access to the dataset to empower anyone looking to accelerate discoveries in biology and medicine.
