KRAS-G12D-DeepLearning: code and data for deep learning analysis of KRAS G12D conformations

This record contains the code, data and trained models for the deep learning analysis of KRAS G12D conformational states in the accompanying manuscript. Dataset: 790 human KRAS chains from 415 Protein Data Bank entries (260 WT, 174 G12D, 195 G12C, 120 G12V), from a PDB snapshot of 1 January 2026. The full list of entries with annotations is in data/PDB_entries_used.csv. Methods: A convolutional neural network (CNN) trained on Cα distance maps, evaluated with 5-fold cross-validation grouped by PDB entry and compared against logistic regression and random forest baselines. Attribution analysis of the CNN, which identifies Switch I as the main driver of nucleotide-state prediction. A variational autoencoder (VAE) that maps the conformational landscape. Asp12 active-site geometry, elastic network model with perturbation response scanning (ENM/PRS), and pocket analysis. Key results: the CNN predicts GTP vs GDP state with AUC 0.97. For G12D vs WT, random forest reaches AUC 0.92 in the GTP state and 0.72 in the GDP state. Contents: raw Cα coordinates, processed descriptors, cross-validation predictions, trained network weights (models/*.pt), analysis scripts and figure-generation scripts (Figures 8–13). Run run_all.sh to reproduce all results. The environment is in environment.yml and requirements.txt.