2026 ANTHROPIC
ИИ_ДЛЯ_
НАУКИ_
ФАРМА.

Anthropic AI for Science life sciences

Hardcore summary: сегодня Anthropic проводит крупнейший life sciences moment. Nobel laureate John Jumper — архитектор AlphaFold — уходит из Google DeepMind в Anthropic. Claude Mythos 5 проектирует drug candidates 10× быстрее human scientists. Novo Nordisk сократил CSR writing time на 90%. Novartis, BMS, Genentech on stage. Этот tech-разбор покрывает event lineup, Jumper profile, 18-month timeline, Claude for Life Sciences platform, Mythos 5 benchmarks, pharma deployments, Coefficient Bio ~$400M acquisition, export controls, 5-step evaluation path, FAQ.

1. Pain Points: почему briefing переписывает AI × biotech narrative

1) Talent signal: архитектор AlphaFold уходит с DeepMind на career peak — 11 дней до event. 2) Numbers, not slogans: 9/14 protein targets → candidates; Mythos 5 beat dedicated protein language models на AAV capsid prediction. 3) Enterprise lock-in: Novartis CEO Vas Narasimhan (Anthropic board), BMS CEO, Genentech R&D head on stage — penetration глубже, чем assumed. 4) Compliance cloud: Mythos 5 offline с 12 июня; только ~100 US orgs regained partial access 26 июня — international pharma teams в operational uncertainty.

2. Что такое event «AI for Science»?

ПолеДеталь
EventThe Briefing: AI for Science
Date30 июня 2026 — 10:00 AM PST
FormatSan Francisco in-person + global livestream
HostAnthropic
AgendaLife sciences vision, product demos, customer spotlights
SpeakerRole
Vas NarasimhanCEO Novartis; Anthropic board member
Chris Boerner, PhDCEO Bristol Myers Squibb
Aviv RegevEVP & Head of R&D, Genentech
Lotte Bjerre KnudsenFormer Novo Nordisk Chief Scientific Adviser
Eric Kauderer-AbramsHead of Life Sciences, Anthropic
Jonah CoolHead of Life Sciences Partnerships, Anthropic
Matthew HerperSenior Writer, STAT News (moderator)

3. John Jumper: кто он и почему move matters

3.1 Человек, решивший 50-летнюю biology problem

John Michael Jumper (b. 1985, Little Rock, Arkansas): Vanderbilt BS math/physics → Cambridge MPhil physics (Marshall Scholar) → UChicago PhD theoretical chemistry (2017). Six months post-PhD — DeepMind secret protein-folding project.

CASP14 (2020): AlphaFold 2 — experimental-grade accuracy in hours. Impact: 214M+ structures; 2M+ researchers, 190+ countries. 2024 Nobel Chemistry (Hassabis; Baker shared). At 39 — youngest chemistry laureate in 70+ years.

3.2 Почему уход с DeepMind на пике?

19 июня 2026 на X: «After nearly nine years, I have decided to leave Google DeepMind and join Anthropic.» Timing: 11 days before briefing. Strategic fit: April Coefficient Bio acquisition (~$400M) — Genentech computational drug discovery talent. Bloomberg 24 июня: AlphaFold co-authors Adler & Pritzel may follow (unconfirmed).

4. Anthropic Life Sciences Timeline

DateEvent
Oct 2025Claude for Life Sciences launch
Feb 2026Allen Institute + HHMI Janelia partnerships
Apr 2026Coefficient Bio acquired (~$400M)
May 19, 2026Andrej Karpathy joins pre-training team
Jun 9, 2026Claude Fable 5 + Mythos 5 launch
Jun 12, 2026US export controls → models offline
Jun 19, 2026John Jumper → Anthropic
Jun 26, 2026Mythos 5 partially restored (~100 US orgs)
Jun 30, 2026AI for Science briefing

5. Claude for Life Sciences: platform integrations

PlatformCapability
BenchlingSOPs, informed consent from ELN/LIMS
10x GenomicsSingle-cell & spatial transcriptomics
PubMed / bioRxiv / medRxivLiterature search & synthesis
Open TargetsTarget identification & prioritization
MedidataTrial enrollment & site monitoring
ClinicalTrials.govCompetitive landscape queries
Wiley Scholar GatewayFull-text journal access
BioRenderScientific figure workflows

Use cases: discovery → preclinical → clinical → regulatory — literature synthesis, hypothesis generation, experimental design, genomics QC, protocol drafting (FDA/NIH-aware), submission gap analysis.

6. Claude Mythos 5 Benchmarks

10x_FASTER | 9/14_TARGETS | 80%_HYPOTHESIS_WIN | 100x_SMALLER_MODEL

Drug design: ~10× acceleration; fully autonomous workflow (binding sites → tool selection → design runs → self-recovery). 9 of 14 targets — immune checkpoints, growth-factor signaling, neurodegeneration, muscle disease, complex structural targets.

AAV capsid (Dyno Therapeutics data): outperformed dedicated protein language models.

Hypothesis generation: ~80% blind-review win rate vs prior Opus. One E. coli antimicrobial hypothesis validated in lab.

Autonomous genomics (one week, unsupervised): single-cell data from 138 animal species; custom ML model 100× smaller than recent Science benchmark, better performance.

7. Real-world pharma deployments

7.1 Novo Nordisk (Ozempic maker)

Problem: CSRs bottlenecked approvals. Solution: NovoScribe on Amazon Bedrock + Claude + RAG, expert-approved templates.

«Claude has helped us cut writing times on CSRs by 90% so we can get documentation directly into human hands for review and approval.» — Waheed Jowiya, Digitalization Strategy Director

Expanding: device protocols, patient materials, full CTD automation.

7.2 Other deployments

Sanofi, AbbVie, AstraZeneca, Genmab, BMS; Komodo Health, Axiom (toxicity prediction via Claude Code + MCP).

8. Coefficient Bio: ~$400M acquisition logic

Stealth startup (<10 employees); founders from Genentech Prescient Design. Mission: ASI for Science. Investor Dimension: 38,513% IRR. Team under Eric Kauderer-Abrams — protein design + biomolecule modeling, bridge от Claude assistants к true AI drug discovery engines.

9. Industry context & Anthropic edge

MetricTraditional R&DAI leverage
Timeline & cost12–15 years, $2.6B+ (2024)Target ID: months → hours
Success rate~10% clinical candidates approveCompound design: 10×–100× faster simulation
Regulatory docsCSR/CTD bottlenecksNovo case: 90% time reduction

Three edges vs OpenAI/DeepMind: Constitutional AI для regulatory trust; vertical stack (connectors + Coefficient + Jumper); locked-in top pharma customers.

10. Export controls & Jumper unknown

Jun 12: Fable 5 + Mythos 5 offline per US export control. Jun 26: Mythos 5 partially restored для ~100 US critical-infrastructure orgs. Fable 5 negotiations ongoing. International teams — including non-US staff at showcased companies — operational uncertainty.

Can Jumper replicate AlphaFold at Anthropic? Genuinely unknown. AlphaFold: DeepMind infra, institutional partnerships, CASP-defined problem. Anthropic — commercial LLM company pivoting into scientific AI.

11. Five steps: evaluate Claude Life Sciences workflows

  1. Map toolchain: Benchling, PubMed, 10x — MCP connector gaps.
  2. Pilot literature RAG: benchmark synthesis time vs manual baselines.
  3. Run scRNA-seq QC: 10x workflows + scverse best practices.
  4. Regulatory doc sandbox: redacted CSR templates, RAG architecture.
  5. Access compliance audit: Mythos 5/Fable 5 eligibility; plan local MLX bioinformatics fallback on Metal.

12. Deep dive: NovoScribe & regulatory document pipeline

NovoScribe = enterprise RAG + expert templates + case-variable injection — not generic chat. CSRs mandatory for approval; traditional drafting — months. NovoScribe routes Claude через Bedrock, structured trial metadata in MongoDB Atlas, AI drafts → human review. 90% time savings shifts approval window — direct commercial value для blockbusters like Ozempic. CTD full automation extends pattern globally.

Non-US researchers: paths to Mythos 5-tier capabilities unclear. Local open models (ESMFold on MLX/Metal) + regional AI stacks = critical backup.

13. Five signals to watch today

  • Will Jumper appear — official role?
  • Mythos 5 biology trusted-access expansion?
  • New Claude for Life Sciences connectors or Agent Skills?
  • Fable 5 restoration timeline?
  • International access path for non-US researchers?

14. FAQ

Q: Что такое AI for Science event?
A: 30 июня 2026 livestreamed briefing — Claude in life sciences, pharma leader case studies.

Q: Кто такой John Jumper?
A: AlphaFold 2 lead, Nobel 2024, joined Anthropic 19 июня after nine years DeepMind. Role undisclosed.

Q: Drug design speed?
A: ~10× key steps; 9/14 autonomous targets → viable candidates.

Q: Mythos 5 доступен?
A: ~100 US orgs only на 30 июня. Fable 5 suspended since 12 июня.

Q: Claude for Life Sciences?
A: Enterprise stack (Oct 2025): Benchling, PubMed, 10x, Medidata, etc.

Q: Coefficient Bio?
A: Anthropic acquired April 2026, ~$400M. Former Genentech computational biology team.

15. Closing: life sciences AI pipelines → unified memory на Metal

Anthropic bet: AI compresses drug discovery timelines at bedside — not just benchmarks. Windows/Linux cloud boxes hit Claude APIs, но stumble на local scRNA-seq preprocessing, literature RAG + MLX fine-tuning in parallel, BioRender/terminal toolchain synergy. Export controls block Mythos 5 → нужен predictable local bioinformatics backup: MACGPU remote Mac nodes — unified memory для scRNA-seq pipelines, ESMFold inference через Metal API, seamless Jupyter/Cursor integration. Controlled compute hedge до того, как Jumper перепишет rules биологии.