hayatlab/cart_llms

Purpose and entry points

This repository implements a CART nomination workflow: score candidate genes using expression, protein-atlas, clinical-trial, and differential-expression evidence, then use LLM prompts to prioritize candidates. The main executable entry points are Python scripts for model/prompt variants and R scripts/notebooks for evidence generation.

Evidence-generation layer

Expression and tissue-specific evidence are produced from GTEx and protein-expression sources. gtex_scoring.R and protein_exp.py appear to generate inputs used by downstream aggregation, with persisted outputs under reproduce/gtex_results/median_expression_results.csv and reproduce/protein_expression/protein_expression_results_with_tissues.csv.

Protein Atlas evidence is handled separately through HumanProteinAtlas.ipynb, Protein_Expression_Scoring.ipynb, and the reproduced score table reproduce/Protein_Atlas_results/gene_scores_skin.csv.

Differential expression and atlas analysis

de_scoring.R provides an R-based differential-expression scoring stage. Skin-focused exploratory/downstream analysis is represented by malignant_skin_atlas_downstream_v2.ipynb and Heatmap.ipynb, with materialized heatmap results at reproduce/heatmap_results/skin_results.csv.

Clinical-trial evidence

clinical_trials_scoring.ipynb derives a distinct clinical-trial signal, persisted as reproduce/clinical_trials/clinical_trial_scores.csv. This is an independent evidence boundary feeding the candidate-ranking process rather than an LLM-generated source.

Score aggregation and prompt preparation

Score_Aggregation.ipynb combines evidence streams into weighted gene-score datasets. Prompt-ready variants encode alternative weighting policies—balanced, equal, expression, heatmap, HPA, and clinical-trial emphasis—under reproduce/llm_prompt_prep/, including reproduce/llm_prompt_prep/gene_scores_expression_emphasis.csv.

The llm_prompt_prep/gene_scores_expression_emphasis.csv path is a working prompt-preparation input outside the reproducibility snapshot.

LLM nomination variants

The LLM stage has four parallel experimental modes: baseline (gemini_model.py, prompt_2_new.py), bulk review (gemini_model_bulk.py, prompt_2_new_bulk.py), no-base-context (gemini_model_no_base.py, prompt_2_new_no_base.py), and zero-shot (gemini_model_zero_shot.py, prompt_2_new_zero_shot.py). Corresponding visualization notebooks, such as llm_prompt1_visual.ipynb and llm_prompt1_visual_zero_shot.ipynb, inspect or compare model outputs.

multi_request.py is the likely orchestration boundary for issuing multiple LLM requests. The model scripts indicate an external Gemini service dependency; credentials, request configuration, and response persistence are not present in the supplied tree.

Reproducibility boundary

reproduce/ is the packaged data/artifact layer: it contains source-like input data, generated evidence scores, prompt-preparation CSVs, and environment definitions in reproduce/environment.yml and reproduce/updated_requirements.txt. The bulk expression input reproduce/bulk_review/gene_reads_v10_skin_sun_exposed_lower_leg.gct.gz supports the bulk-review branch.