PRAXIS: An AI Agent Framework for Case Distillation and Code Validation in Biological Research

This article introduces the PRAXIS framework, an AI agent system driven by literature learning and case distillation for biological research. Addressing the strict requirements for object validation, method applicability, reproducibility, and auditability in the biological research field, the framework converts research experience, failure boundaries, domain rules, and executable programs into structured long-term memory, supporting the entire workflow including problem definition, object validation, and method selection. Experiments show that case learning significantly improves method selection and error suppression capabilities in complex biological research tasks.

The high complexity and professionalism of the biocomputing field impose special requirements on AI agents:

1. Object Validation: Accurately understand the characteristics and constraints of biological entities; incorrect identification will invalidate the research chain.
2. Method Applicability: Choosing inappropriately among massive analysis tools wastes time or produces misleading conclusions.
3. Reproducibility: Any omission or parameter deviation in complex experimental workflows may lead to irreproducible results.
4. Auditability: The biomedical field requires agent behaviors to be reviewable and verifiable by human experts.

PRAXIS takes case distillation as its core methodology, converting research experience into executable, auditable, and transferable capabilities. Its memory system includes four structured components:

• Success Case Library: Stores validated workflows and their decision contexts.
• Negative Case Library: Records failed attempts to help avoid repeated mistakes.
• Domain Rule Library: Encodes basic constraints and best practices of biocomputing.
• Skill Library: Contains executable atomic operations (e.g., tool calls, code snippets). The coordination mechanism supports the entire workflow: Problem Definition → Object Validation → Method Selection → Workflow Execution → Result Interpretation → Review Feedback.

PRAXIS has been instantiated as a biomedical computing agent suite, covering core scenarios:

• Genomics Analysis: Handles the complete workflow from raw sequencing data to variant annotation, automatically selecting tools and parameters while explaining the logic.
• Protein Structure Prediction: Coordinates multiple tools and databases, integrating AlphaFold predictions and experimental data to generate reliability assessments.
• Drug Repurposing: Integrates gene expression, pathway information, and drug databases to propose candidate drug hypotheses and validation schemes.

Experiments validate the effectiveness of PRAXIS:

• Object Validation: Accurately identifies biological data types, with better boundary case handling than the baseline.
• Case Retrieval: Multi-factor similarity matching outperforms keyword search, supporting cross-domain transfer.
• Ablation Study: Removing the success case library reduces method selection accuracy by 35%; removing the negative case library increases error rate by 42%; missing domain rules leads to more constraint violations.
• Benchmark Tests: Achieves or exceeds the performance of professional workflows in public bioinformatics benchmarks, with method selection decision quality superior to the prompt engineering baseline.
• Multi-agent Collaboration: Agents from different professional domains collaborate to complete complex tasks (e.g., collaboration between genomics and statistics agents).

Implications from PRAXIS:

1. Case Learning Outperforms Prompt Engineering: Structured cases contain decision contexts and support analogical reasoning.
2. Auditability Requires Architectural Support: Explicit memory and trajectory records ensure every step is traceable.
3. Failure Experience Has Significant Value: The negative case library improves reliability.
4. Human-Machine Collaboration Is Key: Agents assist in execution and analysis, while humans retain decision-making authority. Outlook: PRAXIS represents a new paradigm for AI-assisted scientific research, which is expected to become a standard tool in biological research, promoting knowledge accumulation and dissemination.