AI Is Designing Drugs Now: Insilico Medicine, AlphaFold, and the Honest State of AI in Peptide Design

We spend a lot of time on this site at the frontier where biology turns into medicine — the one-time gene edit that lowers cholesterol for life, the oral PCSK9 pill that's secretly a peptide, the gene therapy that wants to replace the weekly GLP-1 shot, the senolytic cream that clears "zombie" skin cells. Underneath all of them is a quieter revolution: the way molecules get discovered and designed in the first place. "AI designed this drug" has gone from a pitch-deck slogan to something with peer-reviewed human data behind it.

In June 2025, a drug whose disease target and whose molecule both came from generative AI published Phase 2a results in Nature Medicine. [1] That's a genuine first. As always on this site, the caveats are the whole story — so here's the honest breakdown of where AI drug design actually is, and what it does (and doesn't) mean for peptides.

The company that designs drugs with AI: Insilico Medicine

Insilico Medicine, founded in 2014 by Alex Zhavoronkov, is the clearest real-world case of "AI did the chemistry." Its platform, Pharma.AI, has three pieces worth understanding in plain language: [7]

• PandaOmics — the target-finder. It mines mountains of biological data (genomics, clinical records, the published literature) to nominate which protein a disease hinges on. That's the question that normally eats years of academic biology: what should we even drug?
• Chemistry42 — the molecule-maker. Given a target, it generates brand-new small molecules from scratch, optimizing on the fly for potency, drug-like absorption, manufacturability, and toxicity flags.
• inClinico — a model that predicts a clinical trial's odds of success, used to triage which programs to push.

The pitch is a two-step AI assembly line: one model decides what to hit, another invents something to hit it with. The interesting question is whether the molecules that come off that line actually hold up in humans.

The flagship: rentosertib, an AI-discovered drug for lung fibrosis

The test case is rentosertib (formerly ISM001-055), a treatment for idiopathic pulmonary fibrosis (IPF) — a progressive, often fatal scarring of the lungs with only two approved drugs, both of which slow decline rather than reverse it. A genuinely high-unmet-need disease.

Here's the AI part. PandaOmics nominated a protein called TNIK as a novel driver of fibrosis — a target that wasn't on the standard IPF list. Chemistry42 then designed a small molecule to inhibit it. Insilico says the program went from naming the target to a preclinical candidate in roughly 18 months, having synthesized fewer than 80 molecules, at a reported cost of about $2.6 million — against an industry norm of several years and thousands to millions of compounds screened. [1][4]

Keep the tier straight: those speed and cost figures are company-reported and not independently audited. [4] They're the headline everyone repeats, and they may well be roughly right — but "we did it in 18 months for $2.6M" is a claim from the company that benefits from the claim. What isn't just a press release is what happened next: the molecule went into actual human trials, and the data got published in a top journal.

What the Phase 2a actually showed — and how much to trust it

The trial was GENESIS-IPF (ClinicalTrials.gov NCT05938920), a randomized, double-blind, placebo-controlled Phase 2a, published in Nature Medicine on June 3, 2025. [1][2][3] The design, precisely:

• 71 patients, across 22 sites — all in China.
• Four arms over 12 weeks: placebo, 30 mg once daily, 30 mg twice daily, and 60 mg once daily.
• The primary endpoint was SAFETY, not efficacy — the percentage of patients with a treatment-emergent adverse event. On that measure it passed: adverse-event rates were broadly similar to placebo, i.e. the drug was tolerable enough to keep developing. [1]

The number that got attention is a secondary endpoint — forced vital capacity (FVC), a standard measure of lung function where IPF patients normally decline. At 12 weeks, the best arm (60 mg once daily) showed FVC +98.4 mL (95% CI 10.9 to 185.9), versus −20.3 mL for placebo. [1] That's roughly a 120 mL separation in the right direction over three months — in a disease defined by relentless lung-function loss, a signal worth taking seriously.

Now the honesty section, because this is exactly the kind of result that gets over-read:

• FVC was a secondary endpoint, in a 71-patient, 12-week, single-country study. That is a hypothesis-strength signal, not proof of efficacy.
• Dropout was high — about 22.5% (16 of 71), concentrated in the higher-dose arms.
• There was a liver-toxicity signal — elevated liver enzymes were among the most common reasons for discontinuation, with ALT increases in a third of the top-dose arm. [1]
• The lead investigator stated plainly that the results "will need to be validated in larger cohort studies." [3]

So: a real, peer-reviewed, AI-origin drug with a promising directional signal — and every reason to wait for a bigger trial before believing the effect size.

The honest counterpoint: AI-designed is not AI-proven

This is the part the hype cycle skips. Medicinal chemist Derek Lowe, reviewing Insilico's disclosed pipeline, pointed out that in nearly every case the biological targets were already known to be implicated in the disease — meaning AI mostly accelerated known biology rather than discovering targets out of nowhere. [5] A second expert, machine-learning researcher Andreas Bender, added the uncomfortable statistical wrinkle: a known target tends to be safer, which helps a drug clear the Phase 1 safety bar — which in turn inflates the "AI-discovered drugs have higher success rates" headline. [5]

The single most honest line on the whole subject comes from Insilico's own Nature Medicine paper, which concedes that AI-discovered drugs have experienced similar levels of phase 2 trial failure as non-AI-discovered drugs, and none has so far progressed through phase 3 trials. [1]

Read that twice. AI is compressing the discovery stage — finding targets and designing molecules — dramatically. It has not yet shown that it de-risks the clinic, which is where the overwhelming majority of drugs die, for reasons (toxicity, human efficacy, immunogenicity) that today's models don't reliably predict. As of mid-2026, no fully AI-designed drug has reached Phase 3 or been approved. [5] Rentosertib is the closest, and it's a 71-patient Phase 2a.

The other half of the revolution: AlphaFold and "seeing" proteins

The Insilico story is about generating molecules. The other half of AI-in-biology is about seeing them. In 2021, DeepMind's AlphaFold2 essentially solved a 50-year-old problem — predicting a protein's 3D shape from its amino-acid sequence — at near-experimental accuracy. In May 2024, AlphaFold3 extended that from single proteins to complexes: how a protein binds a small-molecule drug, a strand of DNA or RNA, or another peptide. [8][9] The field's importance was stamped in October 2024, when the Nobel Prize in Chemistry went half to Demis Hassabis and John Jumper for structure prediction, and half to David Baker for protein design — a split worth remembering, because prediction and design are two different superpowers. [10]

What changed in practice: you can now predict, in silico, how a candidate molecule docks into a target protein — and on standard benchmarks AlphaFold3 beats the classical docking software that drug hunters used for decades. [8] But the AlphaFold team is refreshingly blunt about the limits, and you should be too: the model can "hallucinate" tidy structure in regions that are actually floppy and disordered, it predicts a static snapshot rather than how a protein moves, and — the big one — a predicted structure is a hypothesis, not a function. [8] Knowing the shape of a lock doesn't mean you've made a working key, let alone a safe drug.

Where peptides actually come in

Prediction is half the story. The bigger leap for peptide medicine is generative design — not predicting an existing structure, but inventing a brand-new molecule that does a specific job. This is the Baker-lab half of that Nobel, and it's where peptides get interesting:

• RFdiffusion (2023) generates novel protein backbones from noise; a partner model, ProteinMPNN, designs the amino-acid sequence to fold into them; AlphaFold then filters the winners. [11] This pipeline can invent binders to a chosen target with success rates orders of magnitude better than the hand-design era.
• The most peptide-relevant result came in 2024: Baker's team designed, entirely from random noise and with no lab optimization, protein binders to several bioactive peptide hormones — including glucagon, secretin, parathyroid hormone, and neuropeptide Y — some reaching picomolar affinity, among the tightest computationally designed binders ever made. [12] That's AI designing a molecule to precisely grab a peptide.
• DeepMind's own AlphaProteo (2024) is a generative binder-designer reporting 3× to 300× stronger affinities than prior methods, and the first AI-designed binder to VEGF-A — but, honestly, it flat-out failed to produce any binder for TNFα, a reminder that these tools are not universal solvers. [13]
• For peptide therapeutics specifically, the holy grail is the macrocyclic peptide — a ringed peptide that combines small-molecule stability and oral potential with biologic-grade specificity. In 2025, researchers reported de novo AI design of macrocyclic peptides binding targets at nanomolar affinity. [14] That's aimed squarely at peptides' biggest real-world weaknesses (they get chewed up fast and don't survive the gut — the exact reason enlicitide was such a big deal).
• And it's reaching the clinic: Generate:Biomedicines' AI-engineered anti-TSLP antibody (GB-0895) has advanced into Phase 3 for severe asthma — the most clinically advanced AI-designed biologic so far. [15]

The myth worth killing: your peptides were not designed by AI

Here's the honesty point that matters most for our readers, because the marketing is already getting ahead of the truth.

The famous peptides you can actually research today were not designed by AI. Semaglutide and tirzepatide came out of a decade of empirical medicinal chemistry at Novo Nordisk and Lilly — chemists testing thousands of fatty-acid acylations, linkers, and amino-acid swaps to stretch GLP-1's natural half-life from minutes to a week. No generative model invented them. [16] BPC-157 is a fragment of a naturally occurring protein. TB-500, tesamorelin, the GH secretagogues — all are known or natural sequences, not AI creations.

AI's role in peptides is to design the next generation — new binders, de novo macrocycles, novel sequences — not to have invented the ones already in your fridge. So if a vendor ever markets an "AI-designed peptide" to you, the honest translation today is: at best, a preclinical research tool; at worst, a buzzword stapled to an ordinary compound. The science is real and accelerating; the marketing will outrun it. (It's the same reason we built a COA verification guide — claims are cheap; proof is the thing.)

The bottom line

This is a genuine inflection point, told honestly:

• AI now routinely finds targets and designs molecules, and at least one of those molecules — rentosertib — has peer-reviewed Phase 2a human data. Five years ago that was science fiction.
• But "AI-designed" describes a molecule's origin, not its proof. The clinic is still the clinic; most candidates still fail there, AI or not — and by the field's own admission, none has yet cleared Phase 3. A signal in 71 patients is a reason to run a bigger trial, not to declare victory.
• For peptides specifically, the real action is in de novo design — new binders, macrocycles, and novel sequences. It's early and mostly preclinical, but it's pointed straight at the weaknesses (stability, half-life, oral delivery) that have always limited peptide drugs. The peptides on this site weren't AI-designed; a growing share of the next generation will be.

Honestly, it's a hopeful moment — we're watching the design of medicine itself get faster, cheaper, and more precise. We just keep the evidence tiers straight: a molecule an AI dreamed up still has to survive the same brutal gauntlet as everything else, and the gauntlet is where biology humbles everyone, silicon included.

---

PeptidePrices tracks the research-peptide market and the science around it. We don't sell peptides, and we have no financial relationship with any company named here. This article is educational and is not medical or investment advice. Every efficacy and cost figure above is labeled by evidence tier — peer-reviewed, company-reported, or preprint — for a reason: treat them accordingly.

Sources

1. Insilico Medicine et al. — "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial," Nature Medicine (2025) — https://www.nature.com/articles/s41591-025-03743-2
2. ClinicalTrials.gov — GENESIS-IPF trial record (NCT05938920) — https://clinicaltrials.gov/study/NCT05938920
3. Insilico Medicine — "Nature Medicine Publication of Phase IIa Results Evaluating Rentosertib" (press release, June 2025) — https://www.prnewswire.com/news-releases/insilico-medicine-announces-nature-medicine-publication-of-phase-iia-results-evaluating-rentosertib-the-novel-tnik-inhibitor-for-idiopathic-pulmonary-fibrosis-ipf-discovered-and-designed-with-a-pioneering-ai-approach-302472070.html
4. Chemical & Engineering News — "Artificial intelligence predicts a target for treating fibrosis" (Feb 2021; the original 18-month / $2.6M claim) — https://cen.acs.org/physical-chemistry/computational-chemistry/Artificial-intelligence-predicts-target-treating/99/i7
5. Chemistry World — "As AI-designed drug looks to pass final hurdle, will this tech change drug discovery forever?" (Jul 31, 2025; Lowe + Bender skeptic view) — https://www.chemistryworld.com/news/as-ai-designed-drug-looks-to-pass-final-hurdle-will-this-tech-change-drug-discovery-forever/4021894.article
6. GEN — "Lilly Grows AI Footprint with Up-to-$2.75B Insilico Collaboration" (2026; pipeline + deal terms) — https://www.genengnews.com/topics/artificial-intelligence/lilly-grows-ai-footprint-with-up-to-2-75b-insilico-collaboration/
7. Insilico Medicine — Pharma.AI platform & pipeline — https://insilico.com/pipeline
8. Abramson et al. — "Accurate structure prediction of biomolecular interactions with AlphaFold 3," Nature (2024) — https://www.nature.com/articles/s41586-024-07487-w
9. Google DeepMind — "AlphaFold 3 predicts the structure and interactions of all of life's molecules" (May 8, 2024) — https://blog.google/innovation-and-ai/products/google-deepmind-isomorphic-alphafold-3-ai-model/
10. The Nobel Prize in Chemistry 2024 — press release (Baker; Hassabis & Jumper) — https://www.nobelprize.org/prizes/chemistry/2024/press-release/
11. Watson et al. — "De novo design of protein structure and function with RFdiffusion," Nature 620:1089–1100 (2023) — https://www.nature.com/articles/s41586-023-06415-8
12. Vázquez Torres et al. — "De novo design of high-affinity binders of bioactive helical peptides," Nature 626:435–442 (2024) — https://www.nature.com/articles/s41586-023-06953-1
13. Zambaldi et al. — "De novo design of high-affinity protein binders with AlphaProteo," arXiv:2409.08022 (2024, preprint) — https://arxiv.org/abs/2409.08022
14. Rettie, Juergens, Adebomi et al. — "Accurate de novo design of high-affinity protein-binding macrocycles using deep learning," Nature Chemical Biology (2025) — https://www.nature.com/articles/s41589-025-01929-w
15. Generate:Biomedicines — "Global Phase 3 studies of GB-0895, a long-acting anti-TSLP antibody engineered with AI" (press release) — https://www.prnewswire.com/news-releases/generatebiomedicines-to-initiate-global-phase-3-studies-of-gb-0895-a-long-acting-anti-tslp-antibody-for-severe-asthma-engineered-with-ai-302628189.html
16. Knudsen & Lau et al. — history of GLP-1 drug discovery (semaglutide's empirical origin), PNAS / PMC (2024) — https://pmc.ncbi.nlm.nih.gov/articles/PMC11441540/