hypothesis-generation
GitHub基于科学方法，将观察转化为可测试的机制假设。涵盖区分观测与预测、评估假设质量（如可证伪性）、设计实验及制定预测。支持从具体数据或广泛问题出发，结合文献综合生成竞争性假说。
触发场景
安装
npx skills add jaechang-hits/SciAgent-Skills --skill hypothesis-generation -g -y
SKILL.md
Frontmatter
{
"name": "hypothesis-generation",
"license": "CC-BY-4.0",
"description": "Structured hypothesis formulation: turn observations into testable hypotheses with predictions, propose mechanisms, design experiments. Follows the scientific method. Use scientific-brainstorming for open ideation; hypogenic for automated LLM hypothesis testing on datasets."
}
Scientific Hypothesis Generation
Overview
Hypothesis generation is a systematic process for developing testable mechanistic explanations from observations. This knowhow covers the full cycle: from understanding a phenomenon through literature synthesis, generating competing hypotheses, evaluating hypothesis quality, designing experimental tests, and formulating testable predictions.
Key Concepts
1. Hypothesis vs Observation vs Prediction
- Observation: A factual statement about what was measured or seen (e.g., "Drug X reduces tumor size in mice")
- Hypothesis: A proposed mechanistic explanation for the observation (e.g., "Drug X inhibits angiogenesis via VEGF pathway blockade, reducing tumor nutrient supply")
- Prediction: A testable consequence of the hypothesis (e.g., "VEGF levels should decrease after Drug X treatment; tumors in VEGF-knockout mice should show no additional effect")
Good hypotheses are mechanistic (explain HOW/WHY), not descriptive (restate WHAT).
2. Hypothesis Quality Criteria
| Criterion | Definition | Example of Strong | Example of Weak |
|---|---|---|---|
| Testability | Can be empirically investigated | "Protein X binds to receptor Y" (can test with co-IP) | "Life force drives cellular growth" (untestable) |
| Falsifiability | Specific observations would disprove it | "If X is absent, effect disappears" | "X contributes to the effect somehow" |
| Parsimony | Simplest explanation fitting the evidence | Single mechanism | Multi-step chain without evidence |
| Explanatory Power | Accounts for observed patterns | Explains dose-response and tissue specificity | Explains only one observation |
| Scope | Range of phenomena covered | Applies across related systems | Limited to single dataset |
| Consistency | Aligns with established knowledge | Consistent with known pathway biology | Contradicts thermodynamics |
| Novelty | Offers new insight | Proposes unexplored mechanism | Restates established knowledge |
3. Levels of Mechanistic Explanation
Hypotheses can operate at different scales. Strong hypothesis sets include explanations at multiple levels:
- Molecular: Protein interactions, gene regulation, enzymatic activity
- Cellular: Signaling pathways, cell fate decisions, metabolic changes
- Tissue/Organ: Microenvironment, cell-cell communication, organ function
- Organismal: Systemic responses, physiological adaptation
- Population: Evolutionary pressures, epidemiological patterns
Decision Framework
What is your starting point?
├── Specific observation / data → Follow the full 8-step Workflow below
├── Broad research question → Start with Step 2 (literature search) to narrow scope
├── Existing hypothesis to refine → Start at Step 5 (evaluate quality) and iterate
└── Need creative ideation first → Use scientific-brainstorming skill, then return here
| Starting Situation | Approach | Key Steps |
|---|---|---|
| Unexpected experimental result | Phenomenon-driven | Steps 1→2→3→4 (focus on competing explanations) |
| Literature gap identified | Gap-driven | Steps 2→3→4→5 (focus on novelty criterion) |
| Cross-domain analogy noticed | Analogy-driven | Steps 1→4→5→6 (focus on translating mechanism) |
| Contradictory findings in literature | Conflict-driven | Steps 2→3→4→7 (focus on discriminating predictions) |
| Large dataset patterns | Data-driven | Use hypogenic first, then Steps 5→6→7 here |
Best Practices
-
Always generate competing hypotheses (3–5): A single hypothesis is a confirmation trap. Multiple competing explanations force you to design experiments that discriminate between alternatives, not just confirm your favorite.
-
Start with mechanism, not correlation: "X is associated with Y" is not a hypothesis. "X causes Y via mechanism Z" is. Always include the mechanistic link (HOW the cause produces the effect).
-
Make predictions that differ between hypotheses: The most valuable predictions are those where Hypothesis A predicts outcome X and Hypothesis B predicts outcome Y. This is called a "crucial experiment" — design your tests around these discriminating predictions.
-
Ground every hypothesis in evidence: Cite existing literature for each hypothesis. "It is known that pathway X can regulate process Y [Author, 2023]; therefore, we hypothesize that..." Unsupported hypotheses are speculation, not science.
-
State falsification criteria explicitly: For each hypothesis, write "This hypothesis would be falsified if..." before designing experiments. If you cannot state falsification criteria, the hypothesis is untestable.
-
Consider the null hypothesis: The simplest explanation — that there is no novel mechanism and observed effects are due to known processes, artifact, or chance — should always be included as one of the competing hypotheses.
-
Scale predictions quantitatively when possible: "Expression should increase" is weaker than "Expression should increase 2–5 fold (based on known pathway kinetics)." Quantitative predictions enable power analysis for experimental design.
Common Pitfalls
-
Confirmation bias in hypothesis selection: Generating one "main" hypothesis and 2-3 weak alternatives to make the main one look good. How to avoid: Generate hypotheses independently, then rank them by quality criteria. Have someone else review whether alternatives are genuinely competitive.
-
Untestable "just-so" stories: Hypotheses that sound plausible but cannot be empirically tested with current technology. How to avoid: For each hypothesis, immediately write the experiment that would test it. If you cannot design an experiment, the hypothesis needs revision.
-
Confusing correlation-based claims with mechanistic hypotheses: "Gene X is upregulated in disease Y" is not a hypothesis. How to avoid: Always include HOW and WHY in the hypothesis statement. Use the template: "[Mechanism] leads to [effect] because [rationale]."
-
Ignoring contradictory evidence: Cherry-picking literature that supports your hypothesis while ignoring opposing data. How to avoid: In Step 3 (Synthesize Evidence), explicitly section contradictory findings. Each hypothesis must address how it handles conflicting data.
-
Scope creep in hypothesis evaluation: Trying to make one hypothesis explain everything. How to avoid: A hypothesis does not need to explain all observations — it needs to explain the specific phenomenon under investigation. State scope boundaries explicitly.
-
Designing experiments that can only confirm: If your experiment cannot produce a negative result, it does not test your hypothesis. How to avoid: For each experiment, write down what "failure" looks like. Include negative and positive controls.
-
Neglecting feasibility in experimental design: Proposing experiments requiring technology, samples, or timelines that are unrealistic. How to avoid: Include feasibility assessment (available reagents, equipment, sample access, timeline) alongside each experimental proposal.
Workflow
Structured Hypothesis Generation Process (8 Steps)
-
Understand the phenomenon: Clarify the core observation, define scope and boundaries, note what is known vs uncertain, identify the relevant scientific domain(s)
-
Conduct literature search: Search PubMed (biomedical) and general databases for reviews, primary research, related mechanisms, and analogous systems. Look for gaps, contradictions, and unresolved debates
-
Synthesize existing evidence: Summarize current understanding, identify established mechanisms that may apply, note conflicting evidence, recognize knowledge gaps, find cross-domain analogies
-
Generate 3–5 competing hypotheses: Each must be mechanistic (explain HOW/WHY), distinguishable from others, evidence-grounded, and consider different levels of explanation (molecular → population)
-
Evaluate hypothesis quality: Score each hypothesis against the 7 quality criteria (testability, falsifiability, parsimony, explanatory power, scope, consistency, novelty). Note strengths and weaknesses explicitly
-
Design experimental tests: For each viable hypothesis, propose specific experiments with: measurements, controls, methods, sample sizes, statistical approaches, and potential confounds
-
Formulate testable predictions: State what should be observed if the hypothesis is correct, specify expected direction and magnitude, identify conditions where predictions hold, distinguish predictions between competing hypotheses
-
Present structured output: Organize findings into: executive summary, competing hypotheses with evidence, testable predictions, critical comparisons, and detailed appendices (literature review, experimental protocols, quality assessments)
Further Reading
- Platt, JR (1964) "Strong Inference" — Science 146:347-353. Classic paper on designing experiments to discriminate between competing hypotheses
- Popper, KR (1959) "The Logic of Scientific Discovery" — foundational framework for hypothesis falsification
- Chamberlin, TC (1890) "The Method of Multiple Working Hypotheses" — Science 15:92-96. Original argument for generating competing explanations
- NIH Guide to Hypothesis Development — practical guidance for grant-writing hypothesis sections
- Kell, DB & Oliver, SG (2004) "Here is the evidence, now what is the hypothesis?" — BioEssays 26:99-105. Data-driven hypothesis generation
Related Skills
- scientific-brainstorming — open-ended creative ideation when you need divergent thinking before structured hypothesis formulation
- scientific-critical-thinking — evaluating evidence quality and logical reasoning; complements hypothesis quality assessment
- literature-review — systematic evidence gathering; feeds into Steps 2–3 of this workflow
- statistical-analysis — power analysis and experimental design statistics for Step 6
- scientific-writing — structuring hypothesis-driven manuscripts for publication
版本历史
- 02745ef 当前 2026-07-19 09:23
