AI & Machine Learning in Peptide Discovery

From Bench to Algorithm: A New Era in Drug Design
How AI Reshapes Peptide Discovery
AlphaFold and Its Ripple Effects
De Novo Peptide Design: Building From Scratch
Key AI Tools Driving the Field
Notable AI-Designed Peptides
Company Profiles: Who Is Leading the Charge
Where the Field Stands Today
What Comes Next
FAQ
The Bottom Line
References

From Bench to Algorithm: A New Era in Drug Design

Traditional peptide drug development is slow, expensive, and riddled with failure. The standard pipeline — identify a target, screen thousands of candidates, optimize hits through rounds of synthesis and testing — typically takes three to four years just to reach a preclinical candidate. Most of those candidates never make it to human trials.

Artificial intelligence is compressing that timeline. AI-enabled workflows have demonstrated the ability to cut discovery timelines from five years to as little as 12–18 months and reduce costs by up to 40%, according to a 2025 review in Discover Molecules. That shift is particularly meaningful for peptide therapeutics, where the design space is vast — a 20-amino-acid peptide has 20^20 possible sequences, a number that exceeds the atoms in the observable universe.

The AI in pharmaceuticals market, valued at $1.8 billion in 2023, is projected to reach $13.1 billion by 2030. Within that broader market, peptide-focused AI work represents one of the fastest-growing segments, driven by the commercial success of drugs like semaglutide and tirzepatide.

How AI Reshapes Peptide Discovery

AI contributes to peptide drug development across several stages, each with its own set of tools and challenges.

Structure prediction uses trained models to forecast how a peptide will fold in three-dimensional space. Knowing the structure tells you how the peptide might interact with its target — a protein receptor, an enzyme active site, or a cell membrane.

Generative design goes further. Instead of predicting the structure of an existing peptide, generative models create entirely new sequences optimized for specific properties: binding affinity, solubility, stability, or membrane permeability. Think of it as the difference between reading a book and writing one.

Interaction modeling predicts how a designed peptide will bind to its biological target. Will it stick? Will it stick in the right place? Will it actually change the target's function in the way you want?

Property optimization uses reinforcement learning and multi-objective optimization to tune peptides for real-world drug attributes — half-life, immunogenicity, manufacturability — simultaneously. A peptide that binds perfectly but degrades in 30 seconds is useless as a drug.

Each of these steps used to require months of wet-lab work. AI can now perform millions of virtual experiments in hours, narrowing the field to a handful of candidates worth synthesizing and testing.

AlphaFold and Its Ripple Effects

No discussion of AI in biology is complete without AlphaFold, the protein structure prediction tool developed by DeepMind. When AlphaFold2 demonstrated near-experimental accuracy in the 2020 CASP14 competition, it sent shockwaves through structural biology. The work earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry.

AlphaFold has generated over 200 million predicted protein structures and accumulated more than 20,000 citations. But its impact on peptide discovery specifically comes through several downstream applications.

AlphaFold 3, released in 2024, improved predictions of protein complexes — including protein-peptide interactions. That capability matters because most peptide drugs work by binding to protein targets, and understanding those binding interfaces is central to rational design.

AfCycDesign adapted AlphaFold2's architecture with cyclic positional encoding, enabling accurate structure prediction and sequence redesign of cyclic peptides. Several de novo designs yielded experimental confirmation with root-mean-square deviations (RMSDs) under 1 Angstrom — meaning the predicted and actual structures were nearly identical. Some of these designed cyclic peptides served as scaffolds for nanomolar binders against targets like MDM2, a protein implicated in cancer.

EvoBind took a complementary approach. While AlphaFold excels at predicting interactions from structural data, EvoBind generates peptide binders using only sequence information. That distinction matters for targets where high-resolution structures aren't available — which still includes many biologically important proteins.

AlphaFold's real legacy for the peptide field isn't any single prediction. It is the demonstration that machine learning can achieve accuracy comparable to physical experiments at a fraction of the cost and time — a proof of concept that opened the floodgates for specialized peptide design tools.

De Novo Peptide Design: Building From Scratch

De novo design — creating peptide sequences that don't exist in nature — is where AI's creative potential is most visible. Several breakthroughs in 2024-2025 illustrate the field's rapid progress.

Antimicrobial Peptides

Antibiotic resistance is one of the most urgent threats in medicine, and antimicrobial peptides (AMPs) are a promising alternative. The DLFea4AMPGen framework, published in Nature Communications in 2025, used deep learning to extract key features associated with antimicrobial activity and then generate new peptide sequences with those features. The approach identified 12 novel AMPs with validated bioactivity — peptides that kill bacteria but were designed entirely by algorithms.

A separate study published in Nature Materials showcased a family of self-assembling antimicrobial peptides designed through deep learning. These peptides demonstrated potent infection-fighting ability in mouse models against multidrug-resistant bacteria, a result that bridges the gap between computational design and real-world therapeutic potential.

Peptide Inhibitors

EvoPepFold combined a genetic algorithm with molecular docking (using Rosetta) and peptide 3D modeling (using ColabFold) to design inhibitory peptides targeting the SARS-CoV-2 main protease. The framework generated candidate inhibitors that showed strong predicted binding, though wet-lab validation is still in progress.

Multi-Objective Optimization

PepTune, introduced in 2025, uses a masked diffusion language model to optimize multiple drug-like properties simultaneously: binding affinity, solubility, permeability, hemolysis risk, and non-fouling characteristics. That multi-objective approach reflects a maturing field — early AI peptide tools optimized for one property at a time, often creating molecules that excelled on paper but failed in practice.

Key AI Tools Driving the Field

The toolbox for AI-driven peptide design has expanded rapidly. Here are the architectures that matter most.

RFdiffusion

Developed by David Baker's lab at the University of Washington, RFdiffusion uses diffusion models — the same mathematical framework behind image generators like DALL-E — to create entirely new protein and peptide structures. Unlike AlphaFold, which predicts existing structures, RFdiffusion generates novel ones tailored for specific functions. Baker's group has used it to design binders for previously intractable targets.

Protein Language Models

Trained on millions of natural protein sequences, large language models for biology (like ESM-2 from Meta AI, ProtGPT2, and others) learn the "grammar" of amino acid sequences. They can generate new, plausible peptide sequences with desired properties, much the way a language model generates coherent text. These models are particularly useful for initial sequence generation before more computationally expensive structure-based refinement.

Graph Neural Networks

Peptide-protein interactions can be represented as molecular graphs — atoms as nodes, bonds as edges. Graph neural networks excel at learning from this representation, predicting binding energies and interaction geometries without the full computational cost of molecular dynamics simulations.

DiffPepBuilder

This SE(3)-equivariant diffusion model enables de novo design of peptide binders by co-optimizing sequence and three-dimensional conformation simultaneously. That dual optimization is significant because a peptide's function depends on both what amino acids it contains and how they're arranged in space.

Notable AI-Designed Peptides

While no AI-designed peptide has received FDA approval as of early 2026, the pipeline is filling.

Generate Biomedicines' GB-0669, a monoclonal antibody (not a peptide, but designed on the same generative platform), entered Phase 1 clinical trials targeting a region of SARS-CoV-2 previously thought undruggable. The molecule went from concept to clinic in just 17 months. The company's GB-0895, a long-acting anti-TSLP antibody for severe asthma, is advancing toward global Phase 3 studies.

Latent Labs' Latent-X achieved picomolar binding affinities — extremely tight binding — by testing only 30 to 100 candidates per target in the wet lab. Traditional pipelines screen millions of molecules to find hits of comparable quality. That 10,000-fold reduction in screening effort illustrates AI's potential to make drug discovery dramatically more efficient.

The PDCdb database indicates that 78% of peptide-drug conjugates entering clinical trials since 2022 utilized AI-optimized components, up from less than 15% before 2020. That shift suggests AI involvement in peptide therapeutics is becoming the norm rather than the exception.

Company Profiles: Who Is Leading the Charge

Eli Lilly

Lilly committed over $1 billion to AI drug discovery in late 2025, partnering with Nvidia to create a co-innovation lab that connects clinical data with computational power. The company also struck deals with Alphabet's Isomorphic Labs ($45 million upfront, up to $1.7 billion in milestones) and OpenAI for novel antimicrobial discovery. Lilly's Chief AI Officer Thomas Fuchs has stated that the biggest advances will come from combining proprietary experimental data with foundation models. Beyond AI partnerships, Lilly's peptide pipeline includes orforglipron and retatrutide.

Novo Nordisk

Novo Nordisk is building what amounts to a vertically integrated AI platform, partnering with Microsoft Research and Azure AI across regulatory affairs, drug discovery, and trial design. The company has also partnered with Cradle, an AI protein engineering platform, to engineer peptides, enzymes, vaccines, and antibodies. Novo's peptide innovation continues with dual and triple agonists like CagriSema and amycretin — both advancing through clinical development for obesity and diabetes. Learn more in our market analysis of Novo vs. Lilly.

Generate Biomedicines

Founded on the principle of "generative biology," Generate Biomedicines uses machine learning to create novel protein sequences that have never existed in nature. Their platform trains on the entire compendium of known protein structures supplemented with proprietary experimental data. With multiple clinical-stage programs and a technology platform that spans antibodies, peptides, and enzymes, Generate represents the pure-play AI drug design company to watch.

Peptilogics

This Pittsburgh-based company focuses specifically on AI-driven peptide design, with a platform that incorporates molecular simulation, machine learning, and synthetic biology. Their approach emphasizes designing peptides for manufacturability from the start — addressing one of the key failure modes where computationally promising designs prove impossible or prohibitively expensive to synthesize at scale.

Ardigen

Operating as an AI-powered CRO, Ardigen specializes in de novo peptide generation services. Their platform integrates multiple AI architectures — including variational autoencoders, generative adversarial networks, and transformer models — for peptide design projects spanning oncology, anti-infectives, and metabolic disease.

Where the Field Stands Today

It is worth being honest about what AI in peptide discovery can and cannot do right now.

What it does well: AI compresses early discovery timelines by 30–40% and can identify promising candidates from astronomically large sequence spaces. It excels at suggesting sequences with desired binding properties, predicting structures with near-experimental accuracy, and narrowing the experimental funnel from millions of candidates to dozens.

Where it still struggles: Clinical trial duration, regulatory review, and manufacturing scale-up remain bottlenecked by biology, patient enrollment, and regulatory requirements — factors that AI cannot accelerate. AI models also depend heavily on training data quality, and for many peptide targets, that data remains sparse. The field has seen cases where computationally predicted "hits" failed to perform in wet-lab validation, a reminder that in-silico predictions are hypotheses, not guarantees.

As of December 2025, no AI-discovered drug of any type has achieved FDA approval. The most advanced AI-designed drugs are entering late-stage Phase 3 trials, with multiple readouts expected throughout 2026. The field's track record, according to a Drug Target Review analysis, is "validation and disappointment in roughly equal measure."

What Comes Next

Several trends will shape the next two to three years of AI-driven peptide discovery.

Foundation models for biology are getting larger and more capable. Just as GPT-4 outperformed GPT-3, the next generation of protein language models will capture subtler sequence-function relationships, enabling more accurate generative design.

Multimodal models that integrate sequence data, structural data, experimental assay results, and clinical outcomes will replace today's siloed approaches. A model that understands not just how a peptide folds but how it behaves in a living organism will produce more clinically relevant candidates.

Automated wet-lab validation using robotic synthesis and high-throughput screening will close the loop between computational prediction and experimental confirmation. Companies like Generate Biomedicines are already building these integrated platforms.

Oral peptide design presents a specific opportunity for AI. Designing peptides that survive the gastrointestinal tract — optimizing for protease resistance, membrane permeability, and oral bioavailability simultaneously — is a multi-objective optimization problem tailor-made for machine learning. As the oral peptide drug market expands, expect AI-designed oral peptides to become a major focus area.

The peptide therapeutics market is projected to reach $41.7 billion by 2030. AI won't be the sole driver of that growth, but it will increasingly determine which companies get drugs to market fastest — and at what cost.

FAQ

Has any AI-designed peptide been approved by the FDA?

Not yet, as of early 2026. Several AI-designed therapeutic molecules (including peptide-drug conjugates with AI-optimized components) are in Phase 2 and Phase 3 clinical trials. The first approvals are anticipated within the next two to three years, depending on trial outcomes.

How does AlphaFold help with peptide drug design?

AlphaFold predicts three-dimensional protein structures from amino acid sequences with near-experimental accuracy. For peptide drug design, this means researchers can model how potential peptide candidates will interact with their biological targets without waiting months for X-ray crystallography or cryo-EM data. AlphaFold 3, released in 2024, improved protein-peptide interaction predictions specifically.

What is de novo peptide design?

De novo peptide design means creating peptide sequences from scratch — not modifying existing natural peptides, but generating entirely new molecules optimized for a specific therapeutic function. AI tools like RFdiffusion and protein language models enable this by learning the "rules" of protein structure and function from natural sequences, then generating novel sequences that follow those rules while meeting designer-specified objectives.

How much faster is AI-driven peptide discovery compared to traditional methods?

Current evidence suggests AI compresses early discovery timelines by 30–40% and can reduce preclinical candidate development to 13–18 months versus the traditional three to four years. However, clinical trials and regulatory review — which represent the majority of total development time — are not significantly accelerated by AI.

Which companies are leading AI peptide drug discovery?

Major pharmaceutical companies (Eli Lilly, Novo Nordisk) are investing heavily in AI platforms, while specialized companies like Generate Biomedicines, Peptilogics, and Ardigen focus exclusively on AI-driven design. The field also includes technology partnerships — Lilly with Nvidia and Isomorphic Labs, Novo Nordisk with Microsoft Research and Cradle.

Is AI replacing human scientists in drug discovery?

No. AI functions as a tool that dramatically expands what human scientists can explore. The algorithms generate candidates and predictions, but human researchers still design experiments, interpret results, make strategic decisions about which candidates to advance, and navigate the regulatory process. The most successful programs combine computational power with deep biological expertise.

The Bottom Line

AI and machine learning are rewriting the early chapters of peptide drug discovery. The ability to design novel peptide sequences from scratch, predict their structures and binding properties, and optimize multiple drug-like characteristics simultaneously represents a genuine shift in how therapeutic peptides are discovered and developed.

But the hype needs tempering. AI compresses timelines and reduces costs at the front end of the pipeline. The clinical trial process, manufacturing challenges, and regulatory hurdles that follow discovery remain stubbornly resistant to computational shortcuts. The first AI-designed peptide to reach FDA approval will mark a milestone, but it won't erase the fundamental difficulty of turning molecules into medicines.

What is clear: companies that fail to integrate AI into their peptide discovery pipelines will increasingly find themselves at a competitive disadvantage. The question is no longer whether AI will transform peptide drug development, but how quickly — and which organizations will capture the value.

References

AlphaFold and protein structure prediction. DeepMind / EMBL-EBI. https://alphafold.ebi.ac.uk/
DLFea4AMPGen de novo design of antimicrobial peptides. Nature Communications (2025). https://www.nature.com/articles/s41467-025-64378-y
AI-enabled drug and molecular discovery. Discover Molecules (2025). https://link.springer.com/article/10.1007/s44345-025-00037-5
Peptide-based drug design using generative AI. Chemical Communications (2026). https://pubs.rsc.org/en/content/articlelanding/2026/cc/d5cc04998a
A hybrid evolutionary and structural method for AI-guided peptide inhibitor design. Scientific Reports (2025). https://www.nature.com/articles/s41598-025-28061-y
Artificial intelligence in peptide-based drug design. Drug Discovery Today (2025). https://pubmed.ncbi.nlm.nih.gov/39842504/
AI in drug discovery: predictions for 2026. Drug Target Review. https://www.drugtargetreview.com/article/192962/ai-in-drug-discovery-predictions-for-2026/
Nvidia and Eli Lilly AI drug discovery lab announcement. Eli Lilly Investor Relations (2026). https://investor.lilly.com/news-releases/news-release-details/nvidia-and-lilly-announce-co-innovation-ai-lab-reinvent-drug
Novo Nordisk AI and Azure partnership. Microsoft Customer Stories. https://www.microsoft.com/en/customers/story/18752-novo-nordisk-azure
Trends in peptide-drug conjugate research and AI-aided design. Frontiers in Pharmacology (2025). https://www.frontiersin.org/journals/pharmacology/articles/10.3389/fphar.2025.1553853/full
Generate Biomedicines — Generative Biology platform. https://generatebiomedicines.com/
Ardigen: De novo peptide generation tools and progress. https://ardigen.com/latest-progress-and-tools-for-de-novo-generation-of-peptides/

Table of Contents