This repository contains the source code for the TANDEM-DIMPLE project.

Github repository: https://github.com/locitran/tandem.git.

TANDEM-DIMPLE is a DNN model designed to predict the pathogenicity of missense variants. The model is trained on R20000 set obtained from the Rhapsody study, using a wide range of features, including sequence&chemical, structural, and dynamics features.

Comparison of prediction accuracy across general and disease-specific models. (A) Schematic overview of training and evaluation pipelines. 20,361 SAVs from Rhapsody (Ponzoni et al., 2020), referred to as R20000, were split into R20000_train, R20000_val, and R20000_test in a 72:18:10 ratio. DNN-based general disease models (left) were trained on R20000_train using either the TANDEM or Rhapsody feature sets. Model evaluation was conducted on R20000_test, along with two additional SAV sets related to specific diseases: GJB2_knw and RYR1_knw. Subsequently, 30% of the SAVs from each set (GJB2_train/RYR1_train) were applied to fine-tune TANDEM models via transfer learning, while 10% (GJB2_test/RYR1_test) served in evaluating the specific disease models. (B-C) Comparison of prediction accuracy on independent test sets using general disease models and specific disease models.

This repository contains:

1. The code to produce the features
2. TANDEM-DIMPLE model
3. Transfer-learned model for two specific diseases: GJB2 and RYR1.

To install the code, please follow this instruction.

Input format and output format are described in the input_output_format.md file.

Website: https://dyn.life.nthu.edu.tw/TANDEM/

@article{Loci2025,
  author  = {Loci Tran, Chen-Hua Lu, Pei-Lung Chen, Lee-Wei Yang},
  journal = {Bioarchiv},
  title   = {Predicting the pathogenicity of SAVs Transfer-leArNing-ready and Dynamics-Empowered Model for DIsease-specific Missense Pathogenicity Level Estimation},
  year    = {2025},
  volume  = {*.*},
  number  = {*.*},
  pages   = {*.*},
  doi     = {*.*}
}