Penn’s ApexGO: The Generative AI Tool Accelerating Antibiotic Discovery

The Problem ApexGO Solves

Most AI approaches to drug discovery work like a search engine. They scan massive molecular databases looking for compounds that might have antibiotic properties. It’s useful, but it has a ceiling — you’re limited to what already exists.

ApexGO flips that model entirely.

Instead of searching for a needle in a haystack, it starts with a promising-but-imperfect molecule and asks: how do we make this better? It proposes targeted molecular edits, predicts whether each change will increase antimicrobial activity, and uses those predictions to guide the next round of modifications.

“Antibiotic discovery is fundamentally a search problem across an enormous molecular space,” says César de la Fuente, Presidential Associate Professor at Penn Engineering and co-senior author of the paper. “ApexGO gives us a way to navigate that space with far more direction.”

How ApexGO Actually Works

Understanding the mechanics here matters — because this is where the real innovation lives.

Generative AI + Bayesian Optimization

ApexGO is built on two core components working in tandem.

The first is a generative AI layer that proposes molecular tweaks to antimicrobial peptides — short strings of amino acids that can disrupt bacterial cell membranes. These aren’t random edits. The system learns which types of modifications tend to improve activity and focuses its suggestions accordingly.

The second is Bayesian optimization, a statistical method that helps AI systems explore large solution spaces efficiently. Rather than testing every possible peptide variant (which would be computationally and physically impossible), Bayesian optimization balances two priorities: exploiting regions of molecular space that already look promising, and exploring uncertain regions where hidden improvements might still exist.

“It would be impossible to test every possible peptide,” explains Yimeng Zeng, a doctoral student in Computer and Information Science and co-first author of the paper. “Bayesian optimization helps the model make informed choices about what to try next, balancing candidates that look promising with candidates that could teach the model something new.”

The Role of APEX

ApexGO doesn’t operate in isolation. It builds directly on APEX, a previously published AI model from the de la Fuente lab that predicts whether a given peptide is likely to have antimicrobial properties.

Here’s how the loop works in practice:

1. ApexGO proposes a molecular edit to a candidate peptide
2. APEX evaluates whether that edit is likely to increase antimicrobial activity
3. ApexGO uses that prediction to inform the next round of proposed edits
4. The cycle repeats, progressively converging on stronger candidates

“ApexGO begins with a promising but imperfect peptide, proposes precise edits, predicts whether those changes are likely to enhance antimicrobial activity, and then keeps moving toward versions that are more likely to work when we make and test them,” de la Fuente explains.

What the Lab Results Actually Showed

This is where the story gets genuinely compelling — because AI predictions that don’t hold up in the real world are worthless.

ApexGO’s predictions did hold up.

In laboratory tests against disease-causing bacteria, 85% of AI-generated molecules halted bacterial growth. More impressively, 72% outperformed the original peptides they were derived from. The model wasn’t just finding molecules that looked good on paper — it was finding molecules that actually worked.

In mouse models, two antimicrobial peptides created by ApexGO reduced bacterial counts at levels comparable to polymyxin B — an FDA-approved antibiotic used as a last-resort treatment for drug-resistant infections.

“What is striking is that ApexGO’s predictions held up in the real world,” says Jacob R. Gardner, Assistant Professor in Computer and Information Science and co-senior author. “ApexGO was optimizing against another computer model, so one concern was that it might find molecules that looked good to the model but failed in the lab. Instead, the majority of the molecules it designed actually worked.”

That gap between model performance and real-world performance is one of the biggest challenges in computational drug discovery. ApexGO’s track record here is a meaningful signal.

From Accidental Discovery to Systematic Search

The history of antibiotic discovery is largely a history of accidents. Penicillin — the most famous example — was discovered when Alexander Fleming noticed mold contaminating a petri dish and killing nearby bacteria. That was 1928.

Nearly a century later, the field still relies heavily on serendipity and large-scale screening.

“In a sense, we’ve been incredibly lucky,” says de la Fuente. “ApexGO points to a more systematic way forward.”

The de la Fuente lab has spent years hunting for antibiotic candidates in unconventional places — frog secretions, ancient microbes, even woolly mammoths and giant sloths. Their earlier APEX model helped identify promising candidates buried in enormous biological datasets. ApexGO takes the next step: once you have a candidate, it tells you how to make it better.

“We ran ApexGO for a few months and found hundreds of candidates,” notes Gardner. “If we ran that process for a year, how many thousands of these could we find?”

That’s not a rhetorical question. It’s a research roadmap.

Why This Matters for AI in Drug Development

ApexGO is a clear example of how generative AI can move beyond pattern recognition into active optimization — a shift that has significant implications for the broader drug discovery pipeline.

The Efficiency Argument

Drug development is expensive and slow. The average cost to bring a new antibiotic to market runs into the billions, and the failure rate is brutal. Any tool that helps researchers identify which molecules are worth making and testing before they invest in synthesis and clinical trials has enormous practical value.

ApexGO directly addresses this bottleneck. Instead of synthesizing candidates one by one through trial and error, it narrows the search to molecules statistically more likely to work. That’s not just faster — it’s a fundamentally different way of allocating research resources.

The Generalization Potential

The researchers are explicit that ApexGO’s underlying approach isn’t limited to antibiotics.

“In this case, we wanted to optimize peptides for antimicrobial activity,” says de la Fuente. “But you could imagine applying the same idea to peptides with other biological functions, like modulating the immune system or targeting tumors.”

Gardner’s lab is already exploring related approaches using AI agents capable of reasoning through design choices using broader scientific knowledge. The architecture that powers ApexGO — generative proposals guided by predictive feedback and Bayesian search — is a general-purpose optimization framework. Antibiotics are the proof of concept. The applications extend much further.

What ApexGO Is Not (Yet)

It’s worth being clear about where ApexGO sits in the drug development timeline.

The peptides it generates are early-stage candidates. Before any of them could be used to treat human infections, they would need further optimization for safety, metabolic stability, and pharmacokinetics — how long they remain active in the body and how they’re processed.

The researchers are upfront about this. The value of ApexGO at this stage is in prioritization and direction — helping researchers decide which molecules deserve further investment, not delivering finished therapeutics.

That’s still enormously valuable. The drug discovery pipeline is long, and the earlier you can eliminate weak candidates and focus resources on stronger ones, the better the odds of eventually reaching the clinic.

The Bigger Picture

Antibiotic resistance is accelerating. The World Health Organization has identified it as one of the greatest threats to global health. New antibiotics are urgently needed, but the economics of antibiotic development have historically made it an unattractive investment for pharmaceutical companies.

AI tools like ApexGO don’t solve the economic problem. But they do change the cost structure of early-stage discovery — making it faster, cheaper, and more directed to find candidates worth pursuing.

“ApexGO shows that AI can do more than predict which molecules might work: it can help us improve them,” says de la Fuente. “At a time when antibiotic resistance is rising worldwide, we need technologies that help us move faster from an idea to a real therapeutic candidate. ApexGO is an important step toward that future.”

The shift from screening existing molecules to generating and optimizing new ones is one of the most consequential transitions happening in AI-assisted science right now. ApexGO is a concrete, peer-reviewed example of that shift working in practice — not in theory.

For anyone tracking where generative AI is creating real-world impact, this is exactly the kind of use case worth watching closely.