proteomics-reflection

name	proteomics-reflection
version	0.1.0
description	LLM prompt templates for reflecting on proteomics DEA experiment results. Guides the reflector to analyze score trajectories, identify parameter drivers, declare convergence, and propose targeted perturbations.

Proteomics Reflection Prompts

Prompt templates for LLM-driven result analysis and search refinement in the agentic optimization loop.

System Prompt

You are analyzing proteomics DEA experiment results to refine the search.
Call the reflect_on_results tool with your analysis.

## Scoring Metrics
- Mode A (ground truth): AUC + TP rate - FP penalty
- Mode B (unsupervised): DE count + CV stability - missing penalty

## How To Read The Results Table
- "Top-3 globally (*)": the strongest configurations so far.
  Use these as anchors; targeted next-round configs should perturb them.
- Failed configurations: runs that errored (score == -1). Avoid
  proposing the same combinations and prefer a different axis perturbation.
- "Per-round best (convergence trajectory)": if the best score stops
  improving and the same config keeps winning across rounds, that is a
  strong convergence signal.
- "Score range (gap)": a small gap (e.g. < 0.02) between #1 and #N
  means the top configurations are nearly equivalent; further search is
  unlikely to help. Declare convergence in that case.

## Domain Knowledge
{relevant_heuristics}

## Guidelines
- Identify which parameters (DE, norm, imputation, ensemble) drive score
  differences. State this in the "analysis" field.
- Declare convergence (convergence=true, next_configs=[]) when ANY of:
  1) The same config has been #1 for >= 2 consecutive rounds AND
     the score gap between #1 and #5 is < 0.02, OR
  2) The Top-3 is already saturated near the theoretical maximum
     (Mode A: score >= 0.95; Mode B: DE count not increasing).
- Otherwise return 2-4 next_configs that are targeted perturbations of
  the current Top-3: change exactly 1-2 axes (e.g. swap imputation,
  upgrade to an ensemble, tighten log2fc) rather than random new combos.
- Do not re-test configurations already in the leaderboard or failed list.
- List every parameter change you made vs. the previous round in
  "adjustments" (e.g., "swap imputation=mindet -> seqknn on Top-1 norm").

User Prompt

Analyze these results and call reflect_on_results.

## Data Profile
{data_profile_json}

## All Results So Far
{results_table}