Researchers are using AI to compress the hunt for new antimicrobial molecules from years to hours, scanning genomes and protein databases that would be impossible to sift through by hand. The approach treats DNA and amino acid sequences as a kind of alphabet, then trains models to spot patterns linked to antimicrobial activity. Alongside in-house models, the lab leans on Codex and ChatGPT for hypothesis brainstorming, code writing and refinement, dataset wrangling, and cross-disciplinary translation.<\/p>\n
Bacterial antimicrobial resistance was associated with about five million deaths in 2021, and that annual toll is projected to roughly double by 2050. No new class of antibiotics has been approved in roughly 50 years, which is why researchers are looking beyond familiar chemistry and into the code of life itself.<\/p>\n
For decades, antimicrobial development has meant tweaking existing drugs or mining well-known chemical classes. Those strategies now return less and less with each round. The newer approach starts with the fact that only a small fraction of any genome has a clearly understood function, and an even smaller fraction encodes molecules that can fight infection.<\/p>\n
The lab of C\u00e9sar de la Fuente at the University of Pennsylvania takes that observation as a starting point. His deep-learning models are trained to recognize patterns in biological sequences, then search genome and protein databases for candidates that look like they could act as antimicrobials. The team can then hand a short, prioritized list to experimentalists for testing.<\/p>\n
The gain for researchers is straightforward: an initial screen that once took years can now be narrowed down in hours, freeing bench scientists to focus on confirmation, optimization, and safety work.<\/p>\n
Flagging a candidate is only the first step. A promising molecule still has to clear a sequence of real-world tests before it can be prescribed:<\/p>\n
That is why de la Fuente insists that AI predictions must be paired with ground-truth lab experiments. Models can point to needles in a haystack, but only wet-lab work confirms they really work as drugs.<\/p>\n
His group is intentionally mixed, with biologists, chemists, computer scientists, and engineers working side by side. The challenge is that someone who can write code may know little about peptide chemistry, and a microbiologist may not want to wrestle with a new programming language. Codex and ChatGPT help close those gaps by letting biologists build small programs, programmers tackle biology questions, and everyone review unfamiliar literature, clarify terminology, and compare methods across fields.<\/p>\n
Other practical uses have shown up in daily workflows:<\/p>\n
De la Fuente also uses ChatGPT as a brainstorming partner. The shared workspace collects input from lab members with different training, becoming a kind of sounding board for both strong and weak ideas. His caveat is the standard one for research AI: always double-check for accuracy, because the model can produce plausible-sounding mistakes.<\/p>\n
The team’s larger bet is that useful antimicrobials are hiding in places traditional screening never looked: in extinct organisms, in obscure branches of the tree of life, in stretches of sequence no one has annotated yet. AI makes it feasible to scan those unread spaces at all.<\/p>\n
De la Fuente frames the work as part of a long tradition of instruments that extend what scientists can see. Telescopes opened up the cosmos and microscopes revealed the invisible world inside cells; machine learning models are now doing the same for the information written into biology itself. His group is searching those edges between disciplines, where, in his view, the next generation of antimicrobial drugs is most likely to be found.<\/p>\n
Researchers train deep-learning models on biological sequences such as DNA and proteins to spot patterns linked to antimicrobial activity. The models scan huge genome and protein databases, then prioritize a short list of candidate molecules for laboratory testing, reducing initial discovery from years to hours.<\/p>\n
No new class of antibiotics has reached patients in roughly 50 years, and modifying existing medicines is producing diminishing returns. Bacterial antimicrobial resistance was associated with about five million deaths in 2021, with that toll projected to roughly double by 2050.<\/p>\n
In de la Fuente’s lab, Codex and ChatGPT are used to brainstorm hypotheses, write and refine code, process datasets, analyze results, and help biologists and programmers collaborate across disciplines. Researchers caution that all AI outputs need to be verified against ground-truth experiments.<\/p>\n