This repository contains the analysis code for the Manubot project Multi-Dataset Integration and Residual Connections Improve Proteome Prediction from Transcriptomics Using Deep Learning.
Files used for running our code are found under the scripts directory. The files under scripts/NAS are based on the Multi-Objective NAS with Ax tutorial. The files under scripts/5x2 were for generating the boxplot inside figure 1. scripts/SHAP contains preliminary code for running SHAP on the model. Note that the branch input_residual_connection contains a more complete setup for the SHAP runs used in the paper.
The two main files to consider are src/RnaToProteinDataModule/NasModel.py, which contains the deep learning model used, and src/RnaToProteinDataModule/RnaToProteinDataModule.py, which performs the bulk of the data processing steps.
pytorch often gives Mac users trouble. Below is a script that bypasses most errors when installing the necessary packages for a Mac.
conda create -n <env name>
conda activate <env name>
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 -c pytorch
pip3 install ax-platform
pip3 install pytorch-lightning
pip3 install -U 'tensorboardX'
pip3 install tensorboard
pip3 install torchx
pip install -e <path to RnaToProteinDataModule>
In a nutshell: Running a GPU on M1 Macs doesn't always work perfectly well. Interestingly, the most recent pytorch versions sometimes run with the mps GPU without a hitch, but other times will not. I don't know why the inconsistency exists. However, installing packages with the above script has it working every time. I do not recommend trying this in a pre-made environment, as things like the python version might skew the install into not working again.