Can AI Tools Discover Next-Gen Peptide Therapeutics? One Medical Student's Research Says Yes

Drug discovery is one of the most expensive, time-consuming, and failure-prone endeavors in all of human science. Bringing a single new medication from laboratory discovery to patient approval takes an average of 10 to 15 years and costs over $2 billion — with a failure rate that exceeds 90% in clinical trials. For diseases like type 2 diabetes and obesity, which affect hundreds of millions of people globally, that pace of innovation is simply not fast enough.

That's exactly the problem that Anthony Wong, a fourth-year medical student at Carle Illinois College of Medicine in Urbana, Illinois, set out to address — not with a new laboratory technique, but with artificial intelligence. His research, "Machine Learning-Guided Design of Next-Generation Triple Agonist Peptide Therapeutics for Metabolic Disease," was honored in the 2025 AMA Research Challenge with both the Innovative Research Award and People's Choice recognition. It's a window into how AI is beginning to reshape one of medicine's hardest problems.

The Problem: Why We Need Better Metabolic Disease Medications

To understand why Wong's research matters, you first need to understand the scale of the metabolic disease crisis.

According to the Centers for Disease Control and Prevention (CDC), more than 36 million U.S. adults have type 2 diabetes — a condition responsible for roughly one-quarter of all U.S. healthcare spending as of 2021. A separate 2024 CDC data brief found that 40.3% of all U.S. adults had obesity, a condition that significantly elevates risk for stroke, coronary heart disease, and several cancers.

Current frontline medications — particularly GLP-1 receptor agonists like semaglutide (Ozempic, Wegovy) and tirzepatide (Mounjaro) — have shown impressive results for both blood sugar control and weight loss. But GLP-1 agonists work primarily through one receptor target. Researchers have long hypothesized that drugs capable of simultaneously targeting multiple hormone receptors — specifically GLP-1, glucagon, and glucose-dependent insulinotropic polypeptide (GIP) — could deliver more effective treatment outcomes than single-target approaches.

The challenge: designing effective "triple agonist" peptides that can activate all three receptor types at once is extraordinarily complex. The molecular structure of a peptide determines which receptors it can activate and how strongly. Searching the vastly large space of possible peptide sequences by hand or traditional laboratory methods takes years of iterative testing — which is precisely where AI enters the picture.

The AI Approach: Machine Learning Meets Molecular Design

Wong and his research team — including Sanskruthi Guduri, Tsungyen Chen, MD, and Kunal Patel, MD — took a computational approach to peptide design that would have been impossible without modern AI.

Their methodology combined two powerful AI techniques:

1. A 3D molecular property model The team used an AI model capable of capturing both the three-dimensional shape and chemical properties of peptide molecules. Traditional drug screening often treats molecules as simple sequences of amino acids; this approach models the actual physical geometry that determines how a peptide docks with and activates a hormone receptor. Receptor binding is essentially a 3D spatial problem — shape, charge distribution, and flexibility all matter — and the AI model captures these dimensions in a way linear sequence analysis cannot.

2. Genetic algorithms for peptide sequence generation Building on the molecular model, the team applied genetic algorithms — an AI optimization technique inspired by evolutionary biology — to search the space of possible peptide sequences for candidates most likely to activate all three target receptors simultaneously.

Genetic algorithms work by starting with a population of candidate solutions, evaluating their fitness against a target objective (in this case, predicted receptor activation), selecting the best performers, and iteratively combining and mutating them to produce improved candidates over successive "generations." Applied to peptide design, this process can explore millions of molecular variants in the time it would take a laboratory team to manually synthesize and test dozens.

The result: the team generated 20 new peptide sequences with computational evidence for triple-agonist activity — compounds that no human researcher had previously designed or tested.

The Validation: 80-90% Accuracy

Any computational approach to drug discovery is only as valuable as its predictive validity — the degree to which the AI's molecular predictions translate to actual biological activity in the lab.

Wong's team validated their model rigorously. When the algorithm was tested against known peptide sequences with established biological properties, it passed validation testing with 80 to 90% accuracy — a performance threshold that Wong described as an "aha moment" for the research.

"When we saw it pass validation testing with about 80 to 90% accuracy, it was a real 'aha' moment," Wong said. "At that point, I realized we had built something that could have real potential for drug discovery."

For context, an 80-90% accuracy rate on a molecular property prediction task is scientifically meaningful — it suggests the model has genuinely learned patterns in the structure-activity relationship of these peptides, rather than simply memorizing training examples. The 20 novel peptide sequences the algorithm generated are now candidates for experimental laboratory synthesis and biological testing, the next step in the drug discovery pipeline.

The potential economic impact is significant. Traditional iterative lab-based drug design is expensive precisely because researchers must physically synthesize and test each candidate molecule. A computational pre-screening step that can reliably identify promising candidates with 80-90% accuracy dramatically reduces the number of expensive lab experiments required — compressing years of discovery work into weeks or months of AI-guided computation.

What the AMA Recognition Means for AI in Medicine

The American Medical Association's decision to recognize this research with the Innovative Research Award — given to the finalist who "brings fresh ideas to the table — creative thinking, new approaches and research that moves the field forward" — sends a clear signal about where organized medicine sees AI's role heading.

AMA President Bobby Mukkamala, MD, called all five finalists "an extraordinary group whose work represents the innovative thinking that moves medicine forward." Within that group, Wong's AI-driven computational approach stood out for its methodological novelty and its immediate translational relevance.

Wong himself was characteristically humble about the recognition: "I was honestly very surprised. During the AMA Research Challenge, I saw so many impressive and high-quality projects, so being named a finalist was an honor." The fact that his computational AI approach stood out in a field of medically sophisticated research projects reflects a growing recognition that AI-native methods represent a new class of scientific tool in medicine — not just a supplement to existing techniques, but a fundamentally different approach to certain classes of problems.

For the medical profession broadly, this recognition carries an important message: AI is not a replacement for clinical expertise; it's an amplifier of it. Wong is training to be an internist, not a computer scientist. His ability to apply AI to a medical problem was enabled by domain expertise — understanding which hormone receptors matter, what constitutes a therapeutically viable peptide, and what the clinical endpoints of metabolic disease treatment actually are. The AI handled the computational search; the medical knowledge defined what to search for.

AI in Medical Research: A Rapidly Expanding Frontier

Wong's research sits within a rapidly accelerating field. The application of AI to drug discovery, medical imaging, genomics, and clinical decision support is expanding faster than almost any other domain in medicine.

Key areas where AI is already demonstrating validated impact in medicine include:

• Molecular docking and drug design — AI models like AlphaFold (DeepMind) have transformed protein structure prediction, and successor systems are beginning to predict how small molecules and peptides interact with those structures
• Genomic medicine — Machine learning models identify disease-associated genetic variants and predict individual patient drug response profiles
• Medical imaging — AI systems now match or exceed radiologist performance on specific imaging classification tasks in dermatology, radiology, and ophthalmology
• Clinical trial design — AI is being used to identify optimal patient populations, predict dropout rates, and design adaptive trial protocols that reach statistical significance faster

For peptide therapeutics specifically, the case for AI is particularly compelling. The space of possible peptide sequences is astronomically large — a 20-amino acid peptide can be assembled from 20 natural amino acids in 20²⁰ possible combinations. No laboratory could ever screen more than a tiny fraction of that space empirically. AI provides a principled method for navigating that space intelligently, focusing experimental resources on the most promising candidates.

Wong's research is exactly the kind of case study that illustrates why practitioners with AI skills — even at an early career stage — are increasingly positioned to make disproportionate research contributions. The skills to frame a problem computationally, select appropriate AI methods, and interpret model outputs critically are now as relevant in medicine as in technology.

At the FireStart Applied AI Program, we believe this intersection of AI capability and domain expertise is where the most important work of the next decade will happen — not in pure AI research labs, but in fields like medicine, law, engineering, and finance, where practitioners who understand both the domain and the tools can deploy AI in ways that pure technologists cannot. Anthony Wong's research is a compelling example of exactly that kind of applied AI thinking in action.