Unraveling Renal Cell Carcinoma Subtypes and Prognoses by Integrative Histopathology-Genomics Analysis

This repository contains the source codes and the trained convolutional neural networks for predicting renal cell carcinoma subtypes, molecular profiles, and prognoses.

All models are trained using Talos with keras and tensorflow-gpu backend.

Tumor identification

Under the tumor_normal/ directory. Our approaches achieved areas under the receiver operating characteristic curve (AUC) in the independent validation cohort of 0.964-0.985.

Subtype classification

Under the subtypes/ directory. Our models attained an accuracy of 0.935 in the independent validation set.

Overall survival prediction

Under the survival/ directory. Our prediction models distinguished the longer-term survivors from shorter-term survivors among stage I clear cell renal cell carcinoma (ccRCC) patients (log-rank test p = 0.02).

Mutation status prediction

Under the mutations/ directory. We showed that morphological patterns weakly predicted somatic mutation profiles in the three major subtypes of renal cell carcinoma.

Copy number alteration prediction

Under the copy_number_alteration/ directory. We predicted the copy number alterations in ccRCC patients in multiple genes (e.g., VHL, EGFR, KRAS, and WT1) with AUC > 0.7.

Tumor mutation burden prediction

Under the tumor_mutation_burden/ directory. We demonstrated that the histopathology-based features are moderately correlated with tumor mutation burden measured by whole-exome sequencing (Spearman's correlation 0.419; correlation test p = 0.0003).

Ploidy prediction

Under the ploidy/ directory. We trained a binary classification model to predict ploidy > 2 with AUC = 0.633 in papillary renal cell carcinoma.

CDKN2A prediction

Under the cdkn2a/ directory. We trained a binary classification model to predict CDKN2A deletion with AUC = 0.713 in clear cell renal cell carcinoma and AUC = 0.639 in papillary renal cell carcinoma.

9p deletion prediction

Under the 9p_deletion/ directory. We trained a binary classification model to predict 9p deletion with AUC = 0.548 in clear cell renal cell carcinoma and AUC = 0.678 in papillary renal cell carcinoma.

Instructions on using talos

The full documentation of talos could be found here.

Quick installation

pip install talos

Usage and examples

A short example.

A comprehensive example.