SCRIBE: Explainable Machine Learning for Understanding Human Tumors

Abstract

SCRIBE is a Python machine learning pipeline for identifying biomarkers and therapeutic targets from single-cell RNA sequencing (scRNA-seq) data using explainable machine learning. SCRIBE converts raw GEO datasets, corrects batch effects, trains an interpretable Random Forest classifier, and generates publication-quality plots — all from a single CLI. This computational pipeline was presented at the Special Program for High School Students at the 2026 AACR annual meeting in San Diego. SCRIBE uses Random Forest classification on gene expression data to: Convert raw GEO datasets into a standard AnnData format Merge multiple datasets with batch-aware HVG selection Correct batch effects via ComBat or Harmony Classify cells by condition (normal / precancerous / malignant) Identify top discriminating genes through feature importance analysis Analyze differential gene expression with volcano plots and LFC charts Visualize results with UMAP plots and an interactive Marimo explorer SCRIBE has been validated on three published pancreatic scRNA-seq studies from GEO: Dataset Description Conditions GSE154778 Primary vs metastatic PDAC (10+6 samples) primary, metastatic GSE162708 Pancreatic neuroendocrine tumor (24,544 cells) primary_tumor, metastasis, normal GSE165399 Normal pancreas, IPMN, adenosquamous carcinoma normal, IPMN, PASC These are unified into a 3-class scheme: normal / precancerous / malignant.