AI Peptide Drug Discovery: SurfFlow, ApexGen, and RFpeptides Explained

For decades, discovering a new peptide drug meant the same grind: screen millions of random candidates, test a few thousand in the lab, and hope something sticks. It was slow, expensive, and most candidates failed.

That era is ending. A wave of AI-powered tools — published in just the last few months — is transforming how peptide therapeutics are designed. These systems don't just speed up the old process. They replace it with something fundamentally different: computational design from first principles, where AI generates peptides tailored to specific protein targets with atomic precision.

Here's what's happening, why it matters, and what it could mean for the peptides people actually care about.

The Big Three: SurfFlow, ApexGen, and RFpeptides

Three AI platforms have emerged in rapid succession, each attacking peptide design from a different angle.

RFpeptides: The Nobel Prize Lab's Breakthrough

RFpeptides came out of the Institute for Protein Design at the University of Washington — the lab led by David Baker, who won the 2024 Nobel Prize in Chemistry for computational protein design. Published in Nature Chemical Biology in June 2025, RFpeptides uses deep learning to design macrocyclic peptides (ring-shaped molecules) that bind to disease-associated proteins.

What makes it remarkable is the efficiency. For each protein target, the team synthesized and tested only about 20 designed macrocycles — and found high-affinity binders. Compare that to traditional screening, where you might test billions of random peptides to find a handful of hits.

Macrocyclic peptides are particularly interesting therapeutically. Their ring structure makes them more resistant to degradation in the body than linear peptides, improving their potential as drugs. RFpeptides can generate these complex structures computationally, targeting specific proteins using only their structure or sequence as input.

ApexGen: Sequence and Structure in One Shot

ApexGen, published in November 2025 by researchers at the University of Pennsylvania, takes a different approach. Where most AI tools design either a peptide's amino acid sequence or its 3D structure, ApexGen does both simultaneously.

It uses a flow-matching sampler — a type of generative AI model — that couples sequence and structure design at every step. The result: for any given protein target, ApexGen produces a full atomic-resolution peptide model in just a few computational steps. In testing across hundreds of protein targets, the designed peptides showed tight surface complementarity and strong predicted binding affinity.

The speed matters. Because ApexGen generates both sequence and structure together, it skips the iterative back-and-forth that slows down other approaches. This could dramatically compress the early-stage design timeline from months to days.

SurfFlow: Designing From the Surface Up

SurfFlow, published in January 2026, introduces something the others miss: molecular surface properties. Most peptide design tools focus on backbone geometry and amino acid sequence. SurfFlow adds a third dimension — the electrostatic potential, hydrophobicity, and surface geometry that actually determine how tightly two molecules stick together.

Using multi-modal conditional flow matching, SurfFlow co-designs a peptide's sequence, structure, and surface simultaneously. On the PepMerge benchmark — a comprehensive test suite for peptide design — SurfFlow outperformed all existing full-atom baselines across every metric.

This matters because protein-protein interactions ultimately happen at surfaces. The grooves, clefts, and charge distributions of a molecular surface dictate binding specificity. By making surface properties a first-class design target, SurfFlow generates peptides that interact with their targets the way natural peptide binders do.

The "Undruggable" Problem Is Dissolving

Perhaps the most exciting implication of these tools is what they mean for so-called undruggable targets — the estimated 80% of disease-associated proteins that current drugs can't effectively hit.

Many of these proteins are "undruggable" because they lack the well-defined binding pockets that traditional small-molecule drugs need. They're too flat, too flexible, or too disordered for conventional approaches.

Peptides don't have this limitation. They interact through large surface areas rather than small pockets, making them natural candidates for targeting these difficult proteins. And the new AI tools are proving it works.

In July 2025, the Baker Lab published two landmark studies:

A study in Science introduced a design strategy called "logos" for building binders to disordered proteins and peptides. By assembling binding proteins from a library of 1,000 pre-made parts, the team created tight binders for 39 of 43 targets tested — including proteins involved in cancer and neurodegeneration that were previously considered impossible to drug.
A companion study in Nature used RFdiffusion to generate proteins that wrap around flexible targets, producing high-affinity binders (3-100 nM) for amylin, C-peptide, and other disordered molecules. Notably, the amylin binders dissolved amyloid fibrils linked to type 2 diabetes in lab tests.

These aren't incremental improvements. They represent a category shift: targets that the pharmaceutical industry had essentially given up on are now accessible through computational design.

Why This Feels Like an "AlphaFold Moment"

In 2020, DeepMind's AlphaFold solved the protein structure prediction problem — a grand challenge that had stumped biology for 50 years. It didn't just improve existing methods; it made them obsolete overnight.

The current wave of AI peptide design tools has a similar feel. Consider what's changed in just 12 months:

From screening billions to designing dozens. RFpeptides found high-affinity binders by testing ~20 candidates per target, not billions.
From months to minutes. ApexGen generates full atomic models in a handful of computational steps.
From druggable targets only to (almost) anything. The Baker Lab's undruggable protein work shows that 90%+ success rates are achievable against targets the industry had abandoned.
From one property at a time to everything at once. SurfFlow co-optimizes sequence, structure, and surface properties in a single generative pass.

The common thread: AI isn't accelerating the old pipeline. It's replacing trial-and-error with rational design.

That said, this is still the AlphaFold moment, not the AlphaFold aftermath. AlphaFold predicted protein structures computationally, but turning those predictions into actual drugs has been slower than the hype suggested. The same gap will exist here. Computationally designed peptides still need to be synthesized, tested in cells, validated in animals, and run through clinical trials. No AI-discovered drug has received FDA approval as of early 2026.

But the starting point has fundamentally changed. If you can generate high-quality candidates in days instead of years, the entire downstream process compresses.

What This Means for the Peptide Space

For anyone following the peptide world, here's the practical significance:

More Peptide Therapeutics, Faster

The AI-driven peptide drug discovery market hit $1.08 billion in 2025 and is projected to reach $2.44 billion by 2032 (12.3% CAGR). That growth reflects a real shift in pharma R&D spending toward AI-first approaches, especially for peptide and protein therapeutics where the design space is well-suited to generative models.

New Peptides for Old Problems

Many conditions currently treated with broad-acting drugs could eventually be addressed with precisely designed peptides. AI tools like ApexGen and SurfFlow make it feasible to design binders for specific protein targets — including those involved in autoimmune conditions, metabolic disorders (see our guide to GLP-1 peptides for weight loss), and neurodegeneration — that were previously out of reach.

Faster Optimization of Existing Peptides

The same tools used for de novo design can also optimize existing peptides. Want a version of BPC-157 with better oral bioavailability? A GHK-Cu analog with improved skin penetration? A Semaglutide analog with fewer GI side effects? AI can explore billions of variants computationally and identify the most promising modifications before a single molecule is synthesized.

The Gap Between Research and Reality

It's important to be honest about timelines. Most of the tools discussed here are 2025-2026 publications. The peptides they design are validated computationally and sometimes in vitro, but clinical trials are years away. The AI revolution in peptide design is real, but the AI revolution in peptide medicine is still in its early chapters.

The Bottom Line

We're watching the emergence of a fundamentally new capability: the ability to design peptide therapeutics from scratch using AI, with atomic precision, in a fraction of the time and cost of traditional methods. SurfFlow, ApexGen, and RFpeptides represent different but complementary approaches — surface-aware design, joint sequence-structure generation, and deep learning-guided macrocycle engineering.

The "undruggable" 80% of the human proteome is shrinking. The design cycle is compressing from years to days. And the economic incentives are aligning — a billion-dollar market that's doubling every six years.

This is what an AlphaFold moment for peptides looks like. Not the end of the story, but the beginning of a very different one.

Want to learn more about specific peptides mentioned in this article? Explore our peptide directory for evidence-based profiles on BPC-157, GHK-Cu, Semaglutide, and dozens more.