[Discussion] Adapting Paper Methodology is a Nightmare: Building an Agent to Handle the "Transfer" Problem.

A researcher is developing a novel AI agent to tackle the pervasive 'methodology transfer' problem in scientific research, where adapting a high-impact paper's analytical logic to a new study with different constraints (like n=80 samples and no GWAS access) often fails. Moving beyond typical 'dump-it-in-a-vector-DB' RAG (retrieval-augmented generation) approaches, the prototype centers on a SQLite-backed knowledge base that deconstructs paper architectures to extract the 'Scientific Intent'—the hidden 'why' behind methodological choices like using WGCNA or specific DE analysis constraints. The workflow is designed with caution, using prompt-chained checkpoints that require user validation to prevent the AI from hallucinating perfect but impractical workflows that collapse in real-world application.

The core technical challenge involves navigating a custom 6-layer evidence stack, particularly the difficult transition from 'Cell-type Specificity' (L3) to 'Causal Directionality' (L4). The developer questions whether these should remain in a linear chain or be treated as independent nodes. The most critical bottleneck is the 'Proxy Problem'—finding valid methodological substitutes when a user's resources (like low-depth bulk RNA-seq) cannot directly replicate a paper's methods (like spatial transcriptomics). Instead of naively scaling down, the agent aims to find proxies like CIBERSORTx. The developer is leaning towards a Constraint-Satisfaction model, which explicitly maps a user's actual budget, data type, and sample size against the paper's intent, rather than relying on an LLM's potentially flawed 'common sense'.