MultiSurv 📄

Multimodal Deep Learning-based pan-cancer Survival prediction

Using the code • Repo structure • License

---

This repository contains all code developed for the MultiSurv paper:

Long-term cancer survival prediction using multimodal deep learning | scientific reports

Usability note: This is experimental work, not a directly usable software library. The code was developed in the context of an academic research project, highly exploratory and iterative in nature. It is published here in the spirit of open science and reproducibility values.

Using the code

This project was built using the Anaconda distribution of Python 3.6.8. To run a fresh copy of the code, clone this repository and use the Anaconda environment manager to create an environment (from the environment.yml file provided here) to install the MultiSurv code dependencies.

For instructions on how to create and use conda environments, please refer to the docs at conda.io.

Repo structure

The source code is found in the src directory. Higher-level or exploratory analyses, as well as paper figures and tables, are typically run using Jupyter notebooks.

Non-exhaustive lists of links to some key parts of the project are found below.

Basic components

Content	Code source	Description
Data preprocessing code	data directory	Separate Jupyter notebook files for different data modalities
MultiSurv model	multisurv.py	Pytorch model
MultiSurv training	model.py Jupyter notebook	Python Class handling model training Code to run training
MultiSurv evaluation	Jupyter notebook	Evaluation of trained models after loading weights
Baseline model evaluation	Jupyter notebook	Fit and evaluate baseline models

Paper figures and tables

Content	Files	Description
Table 1 - Unimodal data results	Jupyter notebook Jupyter notebook results.csv	Baseline model evaluation MultiSurv evaluation Result table
Table 2 - Multimodal data results	Jupyter notebook results.csv	MultiSurv evaluation Result table
Table 3 - Data summary	Jupyter notebook	Overview of the different data modalities
Fig. 1 - Model architecture	MultiSurv.png	Schematic overview of MultiSurv
Fig. 2 - Survival curves	Jupyter notebook	Predicted survival curves
Fig. 3 - Feature representations	Jupyter notebook	Feature representations and patient survival curves

Code used to generate all additional material for the paper can be found in the figures_and_tables directory.

License

This project is licensed under the terms of the MIT license. See LICENSE file for details.