AI Could Dramatically Accelerate the Search for New Therapeutic Peptides

Peptides occupy an awkward and valuable middle ground in medicine. They are larger and more specific than small molecules, smaller and cheaper to make than antibodies, and they can engage the flat, featureless protein-protein interfaces that classical drugs struggle to touch. The commercial case is no longer speculative: GLP-1 receptor agonists such as semaglutide are peptides, and they have become some of the best-selling drugs in the world.

The bottleneck is not enthusiasm. It is the cost of finding out which peptide, out of an effectively infinite number, is worth making at all. In January 2026, a group led by Pranam Chatterjee at the University of Pennsylvania published PeptiVerse in Nature Communications — an open-source platform that predicts a peptide's developability properties from either its amino acid sequence or its chemical structure (Zhang et al., Nature Communications). It is a useful lens on a broader shift in the field, and on the limits of that shift.

Why Peptide Drug Discovery Is So Difficult

A peptide that binds its target beautifully in a test tube can still fail for a dozen unrelated reasons. Native peptides are cleaved by proteases within minutes. Many aggregate or precipitate at therapeutic concentrations. Charged, hydrophilic sequences rarely cross the plasma membrane, which rules out intracellular targets entirely. Cationic amphipathic peptides — the same chemistry that makes antimicrobial peptides effective — frequently lyse red blood cells. Oral bioavailability is close to nonexistent without heavy chemical modification.

Chemists respond by modifying the molecule: cyclisation, D-amino acids, N-methylation, stapling, lipidation, non-canonical residues. This works, but it pushes the molecule outside the alphabet of the twenty natural amino acids, which is exactly where most sequence-based prediction tools stop working. Each modified analogue must then be synthesised and assayed individually, at a cost of days of work and hundreds to thousands of dollars per compound.

The Sequence Space Is Larger Than Any Screen

The arithmetic is unforgiving. A ten-residue peptide built from canonical amino acids has 20¹⁰ — roughly 10¹³ — possible sequences. At thirty residues, still a short peptide, the count reaches 20³⁰, around 10³⁹. Allow non-canonical residues, backbone modification and cyclisation, and the space expands by many further orders of magnitude.

No screening technology closes that gap. Phage and mRNA display libraries reach roughly 10⁹ to 10¹³ members, which sounds enormous until you notice they select almost exclusively for binding, and say nothing about stability, toxicity or permeability. Conventional medicinal chemistry campaigns test hundreds of compounds. The search is therefore not a problem of testing more. It is a problem of choosing better before testing at all.

The Properties That Must Be Optimised Together

Peptide design is a multi-parameter optimisation problem, and the parameters actively conflict with one another.

Increasing hydrophobicity to improve permeability tends to worsen solubility and haemolysis. Adding cationic charge to improve cell entry often increases toxicity. A candidate optimised on one axis routinely collapses on another, which is precisely why late-stage attrition is so expensive.

How Machine Learning Ranks and Generates Candidates

Two complementary uses of machine learning have matured in parallel. The first is discriminative: train a model on measured data so it can score a proposed sequence on a given property. The second is generative: sample entirely new sequences from a model conditioned on a target, or on a desired property profile.

Modern discriminative models rarely learn peptide chemistry from scratch. Instead they exploit foundation models pretrained on enormous unlabelled corpora — protein language models such as ESM-2 for amino acid sequences, and chemical language models such as PeptideCLM and ChemBERTa for SMILES strings. These produce rich numerical embeddings of a molecule, and a comparatively small predictor is then trained on top of a labelled dataset that may contain only a few hundred to a few thousand measurements. This transfer-learning strategy is what makes prediction feasible at all in a field where experimental data are chronically scarce.

What PeptiVerse Attempts to Predict

PeptiVerse applies that recipe deliberately. It freezes ESM-2 (650M) embeddings for amino acid sequences and PeptideCLM embeddings for SMILES, then trains lightweight, heavily regularised prediction heads — XGBoost, elastic net, support vector machines, MLPs, CNNs and transformer heads — chosen per property rather than applying one architecture everywhere. Accepting both representations is the entire point: it lets the same platform evaluate a canonical peptide and its chemically modified analogue side by side.

The reported tasks are haemolysis, non-fouling, solubility, permeability (cell penetrance, PAMPA and Caco-2), toxicity, half-life and binding affinity to protein targets. The training data are modest by machine-learning standards: roughly 17,000 non-fouling entries, 18,000 solubility entries and 11,000 toxicity entries, but only around 1,400 to 1,600 binding pairs and a few hundred half-life measurements.

Reported performance varies accordingly, and the honest reading is that some tasks are far better solved than others.

The authors evaluate under similarity-aware data splits, a stricter and more honest test than random splitting because it stops a model being rewarded for recognising near-duplicates of its own training set. They are also explicit about the weak points. Half-life prediction is constrained by the data, not the model. And they report that structure-prediction confidence (ipTM) correlated essentially not at all with measured peptide-protein binding, at |ρ| ≈ 0.05 — a pointed caution against treating a confident predicted structure as evidence of affinity.

Where the Acceleration Actually Comes From

The realistic benefit is triage, not oracle-like foresight. A generative model or a display campaign can propose 10⁵ sequences; a predictive platform can score them in minutes and rank the few dozen that plausibly survive every constraint at once. Instead of discovering at week twelve that a potent binder is haemolytic, a team can deprioritise it before ordering synthesis. Chatterjee's own framing is deliberately modest: the platform gives researchers a way to check many of those make-or-break properties earlier.

That is filtering, not deciding. A model trained on 1,400 binding pairs is extrapolating heavily when it scores a genuinely novel chemotype, and false negatives quietly discard good molecules while leaving no experimental trace behind.

Why Laboratory Validation and Clinical Trials Remain Necessary

A prediction is a hypothesis with a confidence interval, generated from historical measurements that were themselves collected under inconsistent assay conditions across different laboratories. It is not evidence.

Nothing in a model captures the immunogenicity of a modified backbone, the metabolite profile in a living animal, tissue distribution, manufacturability at scale, or the interaction between a compound and a specific human disease biology. Only synthesis and assay establish that a peptide binds what it is supposed to bind. Only animal work establishes exposure and tolerability. Only controlled clinical trials establish that any of this translates into benefit for patients — and peptides carry real clinical risks, including immunogenicity and injection-site reactions, that no in-silico score anticipates.

AI changes which molecules enter that pipeline. It does not shorten or replace it, and no peptide is an effective medicine by virtue of having scored well in a model.

Other Emerging Generative Approaches

PeptiVerse sits alongside a fast-moving generative literature. PepMLM, from the same laboratory and published in Nature Biotechnology in 2025, designs binders conditioned on a target's sequence alone, with no structural input required, and reports experimental validation against targets including NCAM1 and AMHR2 (Chen et al., Nature Biotechnology). PepTune uses masked discrete diffusion with Monte Carlo tree guidance to optimise binding, solubility, permeability, haemolysis and non-fouling simultaneously rather than one property at a time. Structure-based diffusion tools — RFdiffusion and its cyclic-peptide descendant RFpeptides, along with DiffPepBuilder and PepGLAD — generate backbones that are geometrically complementary to a target surface.

The natural pairing is obvious: generators propose, property predictors filter, and only the survivors go to the bench. The PeptiVerse authors explicitly position their platform for that role.

What the Coming Decade Plausibly Looks Like

The credible trajectory is not autonomous drug design but a tightening loop. Expect richer standardised datasets, active-learning cycles in which automated synthesis platforms feed fresh measurements back into models within days rather than years, and property predictors trained natively on modified chemistry instead of adapted to it after the fact.

The most plausible near-term win is attrition: fewer molecules synthesised per validated hit, and failure modes caught in week one instead of month six. That is a real and substantial gain, and it is a far narrower claim than the one usually made in press coverage.

Data, not model capacity, is the binding constraint — the PeptiVerse authors say so directly. Progress over the coming decade will come from measuring more and sharing it openly, at least as much as from better architectures.

Frequently Asked Questions

Does PeptiVerse design new peptides?

No. It is a property-prediction platform. It scores peptides you supply, and it is designed to be paired with generative models such as PepMLM or PepTune that propose the candidates in the first place.

Which properties does it predict?

Haemolysis, non-fouling, solubility, permeability (cell penetrance, PAMPA and Caco-2), toxicity, half-life and binding affinity to protein targets — from either amino acid sequences or SMILES structures.

Are AI-designed peptides effective medicines?

Not automatically, and not by virtue of being AI-designed. A high predicted score is a prioritisation signal and nothing more. Efficacy is established only by experiment and, ultimately, by controlled clinical trials.

How accurate are the predictions?

Strongly task-dependent. Cell penetrance classification reaches an F1 of about 0.93, while binding affinity regression reaches a Spearman ρ of about 0.56. Half-life is the weakest area, limited by data scarcity rather than by model design.

Can it handle chemically modified peptides?

Yes, and that is a central design goal. SMILES input covers cyclisation, non-canonical residues and other modifications that sequence-only tools simply cannot represent.

Is it freely available?

Yes. The authors released an open-source implementation alongside a no-code web interface hosted on Hugging Face by the Chatterjee lab.

Sources

• •PeptiVerse: A unified platform for therapeutic peptide property prediction, Nature Communications (2026)
• •Open-access full text (PMC12773018)
• •Target sequence-conditioned design of peptide binders (PepMLM), Nature Biotechnology (2025)
• •Peptide-based drug design using generative AI, Chemical Communications
• •AI Platform Could Speed Peptide Drug Discovery, Penn Engineering

Disclaimer: All information is drawn from published research literature and is provided for educational purposes only. This article does not provide dosing, route, or administration guidance. All products are supplied strictly for in-vitro laboratory and research use, not for human consumption. Not medical advice.