AI scientist agents & hypothesis generation

Research teams drown in literature and data, and good hypotheses are bottlenecked by the time scientists have to read, connect and test ideas. AI scientist systems chain reasoning LLMs, literature retrieval, code execution and domain tools to propose ranked hypotheses, run in-silico analyses and draft write-ups. They work best as tireless junior collaborators on well-defined problems; validation, wet-lab confirmation and accountability stay with human scientists.

Problem & value

• Today: scientists scan papers manually, rely on personal networks for ideas, and run analyses in ad hoc notebooks. Cross-disciplinary connections and negative results are routinely missed.
• AI value: broader literature coverage, systematic generation and critique of candidate hypotheses, automated routine analysis (re-analysis of datasets, parameter sweeps), faster iteration before committing lab budget.
• Evidence so far: Google's AI co-scientist (multi-agent, Gemini-based) reported hypotheses for drug repurposing and a bacterial gene-transfer mechanism that were checked in wet-lab collaborations [1]; FutureHouse reported its Robin multi-agent system proposing a candidate therapeutic direction for dry age-related macular degeneration that was then tested experimentally [2]; Sakana AI's "AI Scientist" produced full ML papers end to end, with one workshop-level paper passing peer review, while independent evaluations found frequent flaws in novelty checks and experimental rigour [3][4]. These are vendor/lab-reported results on selected problems, not general productivity guarantees.

Variants

• Pharma / biotech R&D: target and repurposing hypotheses, mechanism-of-action reasoning (links to SCI-002, SCI-017).
• Materials & chemistry: candidate proposal feeding simulation and self-driving labs (SCI-003, SCI-007).
• Data-driven discovery: agents that autonomously analyse large datasets (genomics, social science, economics) and write reports.
• ML research automation: agents that implement and benchmark ideas in code.
• Academia / public research: institutional research assistants, grant and review support; private use: curious individuals using deep-research tools for learning, not discovery.

How it works

1. research goal + constraints
2. literature & database retrieval (RAG over papers, patents, internal reports)
3. hypothesis generation (multiple agents / debate / evolutionary "generate-critique-rank")
4. novelty & plausibility check (search, knowledge graph, tool calls)
5. experiment / analysis planning
6. execution in sandbox (code, simulations, docking, stats) or hand-off to lab automation
7. result interpretation, iteration, provenance log
8. report with citations ⇒ scientist review & decision

• Reasoning LLMs generate and critique ideas; tournament or Elo-style ranking selects candidates.
• Tools: literature search, code interpreter, domain models (e.g. AlphaFold-class, SCI-001), databases.
• Memory / provenance: every claim linked to a source or run; reproducible notebooks.
• Human-in-the-loop: scientists set goals, review ranked hypotheses, approve any wet-lab, spending or data-sharing step, and sign off on publications. Agents must not self-approve experiments.

Data preparation

Agents need the literature (publisher PDFs, preprint servers, patent databases), internal knowledge (electronic lab notebooks, reports, wikis, file shares), and structured research data (LIMS, assay results, instrument files). Typical gaps: scanned or protected PDFs, licence restrictions on text mining, duplicated and undated reports, free-text lab notebooks without identifiers, failed experiments never written down, inconsistent units and ontologies, and unclear access rights between projects.

Preparation steps: inventory sources and rights (PREP-001, PREP-004); check publisher licences for text and data mining; OCR scans (PREP-006); curate an authoritative corpus with metadata (PREP-013, PREP-010); build a glossary and synonym lists (PREP-015); expose structured data via APIs (PREP-005); and create a test set of questions with expert answers (PREP-018). Tools: GROBID, Docling or similar parsers, a vector database. Expect about 40 percent of the effort in data preparation. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (pharma, R&D): Path B. Agent frameworks on frontier model APIs with EU regions give the best reasoning quality, while internal data stay behind own connectors.
• Developer: Path B. Frontier APIs plus open orchestration frameworks are the fastest way to prototype and evaluate agent designs.
• Public (universities, institutes): Path B. Use EU-hosted or sovereign cloud model endpoints; self-hosting frontier-class models is out of reach for most institutions.

Path A — Buy (off-the-shelf)

• Research agents / deep research: FutureHouse platform (Crow, Falcon, Owl, Phoenix agents) and its spin-out Edison Scientific (Kosmos), Google Gemini Deep Research and AI co-scientist (access programme), OpenAI deep research, Elicit, Consensus, FutureHouse-style tools for literature (see KNOW-008).
• Enterprise R&D platforms: Benchling, Dotmatics, Elsevier/Reaxys and Digital Science tools add agentic features; EU vendors include Iktos (FR, molecule design), Owkin (FR, biology/pharma agents), Cradle (NL, protein design) and Bioptimus (FR, biology foundation models).
• Choose A when: you want literature-grounded ideation quickly and have no ML team; accept that hypothesis generation is generic and check data-processing terms for confidential IP.

Path B — Build on cloud APIs (recommended)

• Models: frontier reasoning models (Claude, Gemini, GPT) via EU regions (Vertex AI, Azure OpenAI EU, AWS Bedrock Frankfurt), or Mistral (EU provider).
• Orchestration: LangGraph, Google ADK, OpenAI Agents SDK, Claude Agent SDK, PydanticAI; Model Context Protocol (MCP) servers for PubMed, arXiv, UniProt, ChEMBL; sandboxed code execution (E2B, Docker).
• Retrieval: OpenAlex, Semantic Scholar API, Europe PMC, PubMed E-utilities, CORE for open access full text; vector DB (pgvector, Qdrant (DE)) and optional knowledge graph (Neo4j).
• Choose B when: you have a defined domain, internal data to integrate, and developers.

Path C — Self-host (open source)

• Frameworks: Sakana's AI-Scientist (open code, including v2 with agentic tree search), FutureHouse PaperQA2 and open Aviary environments, AI2 Asta/OpenScholar-style retrieval tooling, AutoGen/smolagents.
• Models: open-weight reasoning models such as DeepSeek-R1-class, Qwen3 reasoning, gpt-oss, Mistral and Llama families; science-tuned variants (e.g. FutureHouse ether0 for chemistry).
• Hardware: HW4 for large reasoning models at useful speed; HW3 with quantised ~70-120B models for pilots. Confidential patents and unpublished data never leave the premises.
• Effort: substantial; agent reliability needs evaluation harnesses and a dedicated team.

Path D — Custom training

Fine-tuning or RL on domain tasks (e.g. chemistry reasoning, experiment-outcome prediction) pays off only with large proprietary labelled histories, such as years of ELN and assay data. Start with retrieval and tools; train later to reduce cost or add capabilities.

State of the art (2026)

• Multi-agent generate-debate-rank loops with test-time compute are the dominant pattern; systems report that more compute improves hypothesis quality as judged by experts and Elo-like self-ratings [1].
• End-to-end paper-writing agents exist (Sakana AI Scientist v1/v2), but independent audits found weak novelty detection, coding errors and inflated claims; most output is incremental [4].
• Benchmarks: LAB-Bench (FutureHouse, biology research tasks), PaperBench and MLE-bench (reproducing and running ML research), ScienceAgentBench; agents remain well below expert level on open-ended tasks [5].
• Failure modes: hallucinated citations, plausible but untestable or already-known ideas ("novelty illusion"), p-hacking through autonomous analysis, evaluation by LLM judges that share the generator's biases, and reproducibility gaps. Scientists must verify sources and rerun key analyses.
• Trend: coupling with robotic labs (SCI-007) to close the hypothesis–experiment loop, and with specialist models (SCI-001, SCI-002, SCI-003).

Shared resources & combinations

• Same stack (llm-reasoning, agent-framework, vector-db, llm-code): KNOW-008 literature assistant, AGENT-013 multi-agent systems, SCI-018 spectral interpretation agents.
• Upstream/downstream: SCI-017 (target identification), SCI-002 (screening), SCI-003 (materials), SCI-007 (autonomous experimentation), SCI-012 (formal reasoning).
• Data pipelines: a shared literature index (OpenAlex/Europe PMC) and ELN/LIMS connectors serve many R&D use cases; a sandboxed code-execution service is reusable across agents.

Effort, cost & KPIs

• Path A: days to weeks; per-seat or credit pricing.
• Path B: 3-6 months for a domain pilot with evaluation harness and integrations; LLM cost scales with agent steps (long reasoning loops can be expensive, so set budgets).
• Path C: 6+ months, GPU node plus an ML/platform team.
• KPIs: share of agent hypotheses rated worthwhile by experts, citation-accuracy rate, hit rate on experimental validation, time from question to ranked plan, scientist hours saved, cost per validated hypothesis, reproducibility rate of agent analyses.

People & change

Involve management for scope and budget, the research department lead as process owner, scientists as daily users, IT and IT security (sandboxed code execution, tool permissions), legal/compliance (licences, IP, publication rules) and the owners of data sources. Bring in the DPO if personal data, such as clinical or participant data, enter the system (GUIDE-005, GUIDE-013).

Works council: depends. Usage logs of prompts, agent runs and outputs can be suited to monitor researchers' behaviour or performance, and objective suitability suffices for §87(1) no. 6 BetrVG; §90(1) no. 3 requires timely information on AI-supported work processes, and §80(3) allows an AI expert. A works agreement limiting performance evaluation is the usual solution (GUIDE-004). Public institutes involve the staff council.

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; train on hallucinated citations and over-confident hypotheses (GUIDE-006). Pilot with one willing team and one question type, with an AI champion. Typical concerns: loss of scientific autonomy and credit. Address them by keeping agents as assistants, requiring human review of every step and claim, and documenting authorship rules (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: typically minimal risk for internal R&D support; the Act excludes AI systems developed and used solely for scientific research and development (Art. 2(6)), while general-purpose model providers carry GPAI duties [6]. Products reused for regulated decisions (medical devices, safety) change the classification; AI literacy (Art. 4) applies to staff.
• GDPR: mostly non-personal, but patient data, clinical datasets and researcher data may enter prompts; need legal basis, DPA, EU processing, data minimisation (see EDPB/EDPS guidance).
• Research integrity: follow the DFG guidance on generative AI and the ALLEA/EU "Living guidelines on responsible use of generative AI in research"; disclose AI use, no AI authorship, human accountability [7][8].
• Dual use / biosecurity: agents that design or plan biological or chemical work need tool-level restrictions, screening of generated sequences and logging; do not give agents unsupervised lab control.
• IP: confidentiality of unpublished data when using external APIs; inventorship requires human contribution.
• Mitigations: provenance logs, citation verification, human sign-off, sandboxing, budget caps.

Getting started

1. Pick one narrow domain question with known ground truth (a past discovery) to benchmark the agent.
2. Build a retrieval layer over OpenAlex/Europe PMC plus your internal reports with citation checking.
3. Prototype a generate-critique-rank loop (Path B) and have two scientists blind-rate outputs.
4. Add a sandboxed analysis tool and full provenance logging; set cost and step limits.
5. Define governance: AI-use disclosure, approval gates before experiments, data-handling rules.

Sources

1. Google Research — Accelerating scientific breakthroughs with an AI co-scientist — https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/
2. FutureHouse — Demonstrating end-to-end scientific discovery with Robin — https://www.futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system
3. Sakana AI — The AI Scientist: towards fully automated open-ended scientific discovery — https://sakana.ai/ai-scientist/
4. Beel et al. — Evaluating Sakana's AI Scientist for autonomous research — https://arxiv.org/abs/2502.14297
5. Laurent et al. — LAB-Bench: Measuring capabilities of language models for biology research — https://arxiv.org/abs/2407.10362
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. European Commission — Living guidelines on the responsible use of generative AI in research — https://research-and-innovation.ec.europa.eu/news/all-research-and-innovation-news/guidelines-responsible-use-generative-ai-research-developed-european-research-area-forum-2024-03-20_en
8. DFG — Good research practice — https://www.dfg.de/en/basics-topics/basics-and-principles-of-funding/good-scientific-practice ; Executive Committee statement on generative AI models (Sept 2023) — https://www.dfg.de/resource/blob/389608/ai-230921-statement-executive-committee.pdf
9. OpenAlex — https://docs.openalex.org/