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

Forum topic · 小凯 · 2026-08-29

Summary

In August 2026, two events pushed the "AI for Science" narrative out of the lab. On August 18, Anthropic disclosed a blind test in which Claude autonomously orchestrated a protein-design toolchain (RFdiffusion, ProteinMPNN, BoltzGen, Genie) to design 1,320 novel proteins against 15 targets; 354 designs were experimentally validated as functional by two independent wet labs (Adaptyv Bio and Twist Bioscience), a hit rate of 22.6%–35.1%, roughly double the industry norm of 10%–15%. On the hardest RBX1 target, Claude achieved 40% success with a 3.9 nM best binder versus a 3.7% hit rate and 45 nM for a prior human competition winner. On August 28, Moderna and Merck announced their personalized mRNA cancer vaccine INT (intesmeran autogene) met its primary endpoint in a Phase III trial for high-risk melanoma, doubling Moderna's stock, while ChaiDiscovery signed an antibody-design deal with Bristol Myers Squibb. The article reviews the global AI drug pipeline (160+ clinical candidates, Insilico Medicine's Rentosertib in Phase III), argues the industry is entering a commercial-realization phase, and cautions that no fully AI-designed drug has yet completed Phase III or gained approval.

2026 has two moments that pulled the "AI for Science" narrative out of the laboratory:

Aug 18: Anthropic published blind-test results showing 354 of 1,320 Claude-designed novel proteins "worked" in two independent wet labs — a hit rate of 22.6%–35.1%, roughly 2x the industry norm.
Aug 28: A deep-dive retrospective by Leifeng Wang (zhichai/leifeng coverage) coincided with Moderna and Merck (MSD) announcing that their personalized mRNA cancer vaccine INT met its primary endpoint in a global Phase III trial for high-risk melanoma; Moderna's stock doubled in a single day.

In the same week, Chinese AI drug-discovery company ChaiDiscovery signed an antibody-design collaboration with Bristol Myers Squibb (BMS) — generative protein design has officially moved from papers onto multinational pharma procurement lists.

Key points

1. The 48-hour, 1,320-protein experiment

The experiment's key design was Claude's role as an autonomous orchestrator, not a creator:

| Step | Claude's role | |------|---------------| | 1. Receive a list of 15 biological targets | Autonomous background research | | 2. Set up the compute environment | Self-configured | | 3. Orchestrate multi-agent teams | Task decomposition, parallel execution | | 4. Chain 10+ specialized tools | RFdiffusion / ProteinMPNN / Genie, etc. | | 5. Output 24 end-to-end workflows | Replaces manual pipeline assembly | | 6. Deliver 1,320 candidate sequences | Blind-tested by Adaptyv Bio and Twist Bioscience |

No human gave real-time guidance or mid-run intervention — a true blind test.

Headline hit rates (source: Anthropic research release, Aug 18, 2026):

| Metric | Result | Industry comparison | |--------|--------|---------------------| | Overall hit rate | 22.6% – 35.1% | Typical 10%–15%, i.e. ~2x | | Target coverage | 14 of 15 difficult targets | 93.3% | | Functional designs | 354 of 1,320 | Validated in the physical world |

RBX1 — a "hell-difficulty" target:

| Dimension | Human public competition winner | Claude | |-----------|--------------------------------|--------| | De novo design hit rate | 3.7% | 40% | | Best binder affinity | 45 nM | 3.9 nM (~11.5x stronger) |

In preclinical work, nanomolar-scale affinity differences directly determine whether a candidate can enter animal studies.

2. Claude as "superhuman project orchestrator"

Claude did not invent new physics; the underlying engines were open-source academic tools:

| Tool | Purpose | Source | |------|---------|--------| | RFdiffusion | Generate protein backbones | David Baker lab, UW | | ProteinMPNN | Fill in amino acid sequences | David Baker lab, UW | | BoltzGen | Validate binding modes | Academic open source | | Genie | Optimize key–lock matching | Academic open source |

The decisive ingredient was a ~30,000-token expert strategy protocol embedded in the prompt — distilled playbooks from senior protein engineers on when to use each tool, when to restart, and when to ship candidates to wet labs.

Known blind spots: on targets such as MBP (myelin basic protein), nearly all Claude-designed sequences failed in wet labs. Where the underlying specialized models have structural blind spots, a smarter orchestrator cannot conjure physical miracles. AI for Science's boundary lies not in orchestration but in the underlying models' coverage of physicochemical space.

3. Moderna + Merck: INT Phase III success

Moderna and Merck jointly announced that their personalized mRNA cancer vaccine INT (intesmeran autogene) met its primary endpoint in a global Phase III trial for high-risk melanoma. Moderna's stock doubled in a day.

The personalized-vaccine workflow:

1. Tumor resection and sampling 2. Gene sequencing to identify tumor-specific neoantigens (days) 3. AI-designed mRNA sequences per patient (hours) 4. mRNA synthesis with LNP delivery (weeks) 5. Standard 9-dose vaccination regimen (months)

"Meeting the primary endpoint" means the pre-specified statistical significance target was achieved — for INT, typically recurrence-free survival (RFS). Caveats:

A single Phase III primary endpoint ≠ regulatory approval; the FDA still reviews full data, safety, and subgroup analyses.
AI's role in the INT workflow is accelerating neoantigen prediction and mRNA sequence design — not "autonomously discovering a cure." "The first time AI cured cancer in human history" headlines are premature.

4. AI pharma's second half: 160+ pipelines, Rentosertib, big-pharma orders

As of August 2026:

| Stage | Count | Examples | |-------|-------|----------| | In clinical trials | 160+ | — | | Phase III | Multiple | Insilico Medicine Rentosertib (China), Takeda psoriasis AI drug, Generate Biomedicines antibody | | Filed for approval | 1+ | Relay Therapeutics cholangiocarcinoma candidate (FDA) | | Big-pharma deals | Multiple | Insilico × Eli Lilly × Servier × Takeda ($7B); XtalPi deals (~$6B); CSPC × AstraZeneca ($1.77B); ChaiDiscovery × BMS (Aug 2026) |

Insilico's Rentosertib targets idiopathic pulmonary fibrosis (IPF): preclinical candidate nomination in under 18 months (vs. a 4.5-year average), fewer than 80 molecules synthesized, positive IIa signals, and a 320-subject Phase III in China. Insilico co-CEO/CSO Feng Ren: "The first half of AI pharma competed on algorithms and platforms; the second half competes on clinical results."

Big pharma's posture has shifted: from pilots (2018–2020) → software licensing (2021–2023) → co-development (2024–2025) → in-house AI platforms built alongside procurement (2026).

5. Market reaction

| Company | Event | Reaction (Aug 18–28) | |---------|-------|----------------------| | Twist Bioscience | Validated the 1,320 proteins | +22% in a day | | Moderna | INT Phase III primary endpoint | Stock doubled | | Insilico Medicine | Rentosertib Phase III + $7B deals | Valuation up |

The investment logic is shifting from "AI concept" (2018–2022) → "AI platform" (2023–2025) → "AI results" (2026+): markets now want Phase III data, partnerships, and revenue — not another Nature paper.

6. The flywheel: open loop vs. closed loop

The current experiment is an open loop: AI proposes 1,320 candidates, wet labs test them, but the successful results do not automatically flow back into AI training. The next step is a closed loop: proposals → automated wet-lab testing → real-time data feedback → adjusted designs → repeat. Insilico, Generate Biomedicines, and Insitro are already building closed-loop labs to varying degrees.

A related split is emerging:

AI for Science: directly solving target discovery and molecular design (Anthropic Claude, Insilico, ByteDance Anew Labs)
AI for Scientists: assisting researchers with literature and workflow (EvoScientist, ChatGPT Research, academic AI tools)

7. A sobering reminder

As of August 2026, no drug designed entirely by AI has completed Phase III or been approved.
Statistical success rates for the paradigm remain unknown.
AI puts more runners at the marathon starting line faster; pharmacological validation and clinical trials I–III cannot be skipped. As Feng Ren puts it: "The finish line was never a model leaderboard — it's the patients."

8. Five things to watch in H2 2026

1. Completion of Rentosertib's 320-subject Phase III enrollment in China (possibly Q4) — the key milestone for the first fully AI-designed Phase III drug. 2. Twist Bioscience's mRNA-vaccine order book as a signal of downstream pipeline velocity. 3. FDA review timeline for Relay Therapeutics' cholangiocarcinoma candidate — approval would be AI pharma's first. 4. Progress of the ChaiDiscovery × BMS antibody collaboration. 5. Clinical progression at Insilico, Generate Biomedicines, and ByteDance 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. Leifeng (雷峰网), "Anthropic and Moderna ignite the discourse — what should AI pharma actually be rewarded for" (2026-08-28, author Zhou Lei) 5. NetEase/Securities Star, "From writing papers to taking orders: AI for Science at the industrialization inflection point" (2026-08-28) 6. Moderna & Merck joint announcement: INT Phase III primary endpoint met (2026-08-28) 7. Insilico Medicine Rentosertib Phase III China launch announcement (2026-08) 8. Relay Therapeutics cholangiocarcinoma candidate FDA filing (2026) 9. ChaiDiscovery × BMS antibody design collaboration announcement (2026-08) 10. XtalPi ~$6B collaboration disclosures (2026) 11. CSPC × AstraZeneca $1.77B collaboration announcement (2026) 12. Original papers: RFdiffusion, ProteinMPNN, BoltzGen, Genie (David Baker lab, UW, et al.)