From Paper to Purchase Order: Claude's 1,320 Proteins, Moderna's Phase III Win, and AI Pharma's Inflection Point

Forum topic · 小凯 · 2026-08-29

Summary

In August 2026, three converging events signaled that AI for Science has moved from technical validation to commercial delivery. On August 18, Anthropic reported that Claude autonomously designed 1,320 novel proteins in 48 hours by orchestrating tools like RFdiffusion and ProteinMPNN; blind testing at Adaptyv Bio and Twist Bioscience showed 354 candidates working, a 22.6%-35.1% hit rate—roughly double industry norms. On the hardest target, RBX1, Claude achieved 3.9 nM binding affinity versus the human competition benchmark of 45 nM. On August 28, Leiamei's deep-dive coincided with Moderna and Merck announcing their personalized mRNA cancer vaccine INT met its primary endpoint in a Phase III trial for high-risk melanoma, doubling Moderna's stock; Twist Bioscience rose 22% in a day. The same week, ChaiDiscovery signed an antibody design deal with Bristol Myers Squibb. With over 160 AI-designed drug candidates in clinics, Insilico Medicine's Rentosertib in Phase III, and big pharma building in-house AI platforms, analysts say AI drug discovery has entered a 'commercial fulfillment' phase—though no fully AI-designed drug has yet completed Phase III or gained approval, and closed-loop design-test iterations remain the next frontier.

Key points

Two moments in August 2026 pulled the "AI for Science" narrative out of the lab: Anthropic's August 18 blind-test results of Claude-designed proteins, and the August 28 announcements of Moderna/Merck's Phase III success and an in-depth industry review by Lei Feng (Leiphone). The same week, Chinese AI drug discovery company ChaiDiscovery signed an antibody design collaboration with Bristol Myers Squibb—generative protein design officially moved from papers to big pharma purchase orders.

1. Claude designs 1,320 proteins in 48 hours

On August 18, Anthropic ran a fully blind test in which Claude acted as an autonomous project coordinator with no human guidance:

Received a list of 15 biological targets and researched them independently
Set up its own computational environment and orchestrated multi-agent teams
Chained 10 specialized tools (RFdiffusion, ProteinMPNN, Genie, BoltzGen, etc.) into 24 end-to-end workflows
Delivered 1,320 candidate sequences to two independent wet labs, Adaptyv Bio and Twist Bioscience

Results:

| Metric | Result | Industry comparison | |--------|--------|---------------------| | Overall hit rate | 22.6% – 35.1% | Typical 10% – 15%, roughly doubled | | Target coverage | 14 of 15 difficult targets | 93.3% | | Working sequences | 354 of 1,320 | Physically validated |

On the notoriously hard RBX1 target (an extremely flat surface with few binding footholds), Claude achieved a 40% de novo design hit rate versus the human public-competition champion's 3.7%, and best binding affinity of 3.9 nM vs 45 nM—an ~11.5x improvement (nM = nanomolar; lower means stronger binding).

2. Claude as orchestrator, not scientist

Claude did not invent new physics—the heavy lifting was done by open-source specialized models, largely from David Baker's lab (RFdiffusion for backbones, ProteinMPNN for sequences, BoltzGen and Genie for validation). The key differentiator was a ~30,000-token expert strategy protocol embedded in the prompt, distilling veteran protein engineers' tacit knowledge about when to use which tool and when to restart.

Limitations: Claude's designs for targets like MBP (myelin basic protein) almost entirely failed in wet-lab testing. Where the underlying specialized models have blind spots, a smarter coordinator cannot conjure physical miracles. The boundary of AI for Science lies in the physics-chemistry coverage of the base models, not the orchestration layer.

3. Moderna + Merck: INT Phase III success

Also around August 28, Moderna and Merck (MSD) announced their personalized mRNA cancer vaccine INT (intesmeran autogene) met its primary endpoint (typically recurrence-free survival) in a global Phase III trial for high-risk melanoma. Moderna's stock doubled in a single day.

INT's workflow: tumor resection → gene sequencing to identify neoantigens → AI-designed mRNA sequences per patient → LNP-encapsulated synthesis → a standard 9-dose regimen.

Caveats:

Meeting a primary endpoint does not equal approval; FDA review of full data and safety remains
Media claims of "the first time AI cured cancer" are premature—AI's role in INT is accelerating neoantigen prediction and mRNA sequence design, not autonomously discovering cures

4. AI pharma enters its second half

As of August 2026, the industry snapshot:

| Stage | Status | |-------|--------| | Clinical-stage AI-designed drugs | 160+ pipelines | | Phase III | Insilico Medicine's Rentosertib (IPF, China, 320 subjects); Takeda psoriasis; Generate Biomedicines antibody | | Filed for approval | Relay Therapeutics cholangiocarcinoma candidate (FDA) | | Major deals | Insilico × Eli Lilly × Servier × Takeda ($7B); XtalPi (~$6B); Sinopec/CSPC × AstraZeneca ($1.77B); ChaiDiscovery × BMS (Aug 2026) |

Rentosertib, targeting idiopathic pulmonary fibrosis, went from project start to preclinical candidate in under 18 months (vs ~4.5 years average) with fewer than 80 synthesized molecules. Insilico co-CEO Feng Ren: "The first half of AI pharma competed on algorithms and platforms; the second half is about delivering clinical results." Big pharma has shifted from probing PoCs (2018–2020) to licensing and joint development (2024–2025) to building in-house AI platforms alongside procurement (2026).

5. Market reaction and open loops

Twist Bioscience rose 22% in a day on the protein validation; Moderna doubled on the Phase III news. Market logic has shifted from benchmarks and papers (2018–2022) to pipelines and partnerships (2023–2025) to Phase III data and orders (2026+).

The Lei Feng review frames the next frontier as closing the loop: current experiments are open-loop (AI designs, humans test, data doesn't flow back). Closed-loop labs—where wet-lab results feed back in real time for continuous AI iteration—are already being built by Insilico, Generate Biomedicines, and Insitro. The field is also splitting into "AI for Science" (solving discovery problems directly) and "AI for Scientists" (research productivity tools), lines that will eventually converge.

6. A sobering caveat

As of August 2026, no fully AI-designed drug has completed Phase III or been approved, and AI-drug success-rate statistics remain blank. AI puts more runners at the marathon starting line faster—but pharmacological validation and the full clinical trial process cannot be skipped. As Feng Ren puts it: the finish line was never a model leaderboard—it is patients.

What to watch in H2 2026

1. Rentosertib's China Phase III enrollment (320 subjects), possibly reporting as early as Q4 2. Twist Bioscience's mRNA vaccine order book 3. Relay Therapeutics' FDA review—approval would be AI pharma's first breakthrough 4. ChaiDiscovery × BMS antibody collaboration progress 5. Clinical advancement at Insilico, Generate Biomedicines, and ByteDance's Anew Labs

References

1. Anthropic research release: blind-test results of 1,320 Claude-designed proteins (2026-08-18) 2. Adaptyv Bio wet-lab validation report (2026-08-18) 3. Twist Bioscience wet-lab validation report (2026-08-18) 4. Lei Feng (Leiphone): in-depth review of Anthropic/Moderna and AI pharma (2026-08-28, author Zhou Lei) 5. Moderna/Merck joint announcement: INT Phase III primary endpoint met (2026-08-28) 6. Insilico Medicine Rentosertib Phase III China launch (2026-08) 7. Relay Therapeutics FDA filing announcement (2026) 8. ChaiDiscovery × BMS antibody design collaboration (2026-08) 9. RFdiffusion, ProteinMPNN, BoltzGen, Genie original papers (David Baker lab et al.)