scienceclaw-prediction

Scientific Prediction Skill

Predict properties, trends, and outcomes across scientific disciplines.

When to Use

• "Predict the solubility of this compound"
• "What's the expected trend for..."
• "Estimate the effect size for this intervention"
• "Forecast the trajectory of..."
• "What properties would this material have?"
• Model-based estimation tasks

When NOT to Use

• Looking up known properties (use literature-search)
• Running actual computations (use code-execution)
• Verifying existing predictions (use scienceclaw-verification)
• Real-time data monitoring or alerts

Prediction Categories

1. Property Prediction

• Chemistry: Molecular properties (logP, solubility, toxicity, pKa, boiling point)
• Materials: Mechanical (strength, hardness), thermal, electrical properties
• Biology: Protein function, binding affinity, gene expression levels
• Physics: Material behavior under conditions (temperature, pressure)

2. Trend Analysis

• Time-series extrapolation with confidence intervals
• Growth/decay curve fitting (exponential, logistic, polynomial)
• Seasonal pattern identification
• Regime change detection

3. Outcome Prediction

• Clinical trial outcome estimation
• Experimental result prediction
• Treatment response probability
• Environmental impact forecasting

4. Model-Based Estimation

• QSAR/QSPR (quantitative structure-activity/property relationships)
• Pharmacokinetic modeling (ADME)
• Population dynamics modeling
• Economic indicator forecasting

Output Format

All predictions must include:

**Prediction**: [Value or range]
**Confidence Interval**: [Lower - Upper] at [confidence level]%
**Method**: [Approach used]
**Key Assumptions**: [List]
**Uncertainty Sources**: [List]
**Validation**: [How to verify this prediction]
**Caveats**: [Known limitations]

Guidelines

1. Always quantify uncertainty — never provide point estimates without ranges
2. State assumptions explicitly — hidden assumptions undermine predictions
3. Distinguish extrapolation from interpolation — flag when predicting outside training data range
4. Consider domain constraints — physical laws, biological limits, economic boundaries
5. Recommend validation approaches — suggest experiments or data to verify predictions
6. Use appropriate models — match model complexity to data availability
7. Flag low-confidence predictions — be transparent about reliability

Discipline-Specific Methods

Domain	Common Methods
Chemistry	QSAR, DFT calculations, molecular dynamics
Biology	Sequence-based prediction, network analysis
Medicine	Cox regression, Kaplan-Meier, NNT/NNH
Physics	Theoretical models, scaling laws
Economics	Econometric models, agent-based simulation
Climate	GCM projections, statistical downscaling
Materials	Phase diagrams, computational screening
Sociology	Panel data models, social network evolution

Computational Prediction Tools

When predictions require computation, integrate with these skills:

Molecular Property Prediction

• Use rdkit-chemistry for descriptor-based QSAR models
• Use pubchem-compound to retrieve experimental property values for training data
• Use scikit-learn-ml to build/evaluate prediction models

Materials Property Prediction

• Use materials-project to retrieve DFT-computed properties
• Use pymatgen-materials for structure-property analysis
• Use scipy-analysis for interpolation and regression

Biological Outcome Prediction

• Use biopython-bio for sequence-based feature extraction
• Use transformers-inference for protein language models (ESM, ProtTrans)
• Use scanpy-singlecell for cell-type and trajectory prediction

Geospatial/Climate Prediction

• Use geopandas-spatial for spatial feature engineering
• Use copernicus-climate for historical climate data as training input
• Use statsmodels-stats for time-series forecasting models

Zero-Hallucination Rule

ALL factual claims, citations, database results, and scientific data presented to the user MUST come from actual tool results (API calls, code execution, web search) in this conversation. NEVER fabricate or "fill in" details from training data. If a tool returns no results or partial data, report exactly what happened.