Title	Authors
Bimodal Neural Network for Cancer Prognosis Prediction	Bo-Run Wu Tsung-Wei Lin Zow Ormazabal

Exploiting Common Patterns in Diverse Cancer Types via Multi-Task Learning

We developed a Python code base to interact with the REST API provided by the GDC Portal for data filtering and download. Due to Python packages' limited functionality compared to those in other languages like the R package TCGAbiolinks, we resorted to using this custom Python code base. We filtered out patients lacking survival status, time, complete RNA-Seq, and clinical data. We downloaded the data of qualified patients using GDC API version 33.1, released on May 31, 2022

System Requirements

This code was tested on Ubuntu 18.04.6, with 46.9 GiB RAM, Intel® Core™ i7-8700 CPU @ 3.20GHz × 12 processor, and NVIDIA GeForce RTX 2080/PCIe/SSE2 graphics card. We used Python 3.10.9 and Pytorch 1.13.1.

Besides the packages listed in requirements.txt, DGL is required.

Installation guide

For running this code, you will need to create en environment in Anaconda using the following command:

conda env create --file environment.yaml

Make sure you are using the same library versions as in the environment.yaml file so that you avoid any errors regarding version incompatibilities. Installing all the dependencies might take between 10 and 30 minutes.

Reproducing results:

For reproducing the results for the single-task and multi-tasks models you need to run:

main.py -c [CONFIG NAME FILE]

Each configuration file will give you the results for each of the modes and ablation studies that we performed. The following table summarizes which table corresponds to which table in our study.

Config Filename	Task	Notes	Table
tcga_brca_coad_multi_dnn_trainer_task_cross_validation_4_bootstrap_tpm_clinical_overall.yaml	Multi	Leave LUAD out	3
tcga_brca_luad_multi_dnn_trainer_task_cross_validation_4_bootstrap_tpm_clinical_overall.yaml	Multi	Leave COAD out	3
tcga_dnn_trainer_cross_validation_4_bootstrap_test.yaml	Single		1,2
tcga_luad_coad_multi_dnn_trainer_task_cross_validation_4_bootstrap_tpm_clinical_overall.yaml	Multi	Leave BRCA out	3
tcga_multi_dnn_trainer_cross_validation_4_bootstrap_test.yaml	Multi	Original	2,3
tcga_multi_dnn_trainer_task_cross_validation_4_bootstrap_tpm_clinical_overall.yaml	Multi	Without task description	3
tcga_multi_dnn_trainer_task_cross_validation_4_bootstrap_unordered_tpm_clinical_overall.yaml	Multi	Without Ordered RNA-Seq Data	3
tcga_multi_dnn_trainer_task_cross_validation_4_bootstrap_unweighed_tpm_clinical_overall.yaml	Multi	Without Weighted Random Sampler	3
tcga_multi_dnn_trainer_unique_task_cross_validation_4_bootstrap_tpm_clinical_overall.yaml	Multi	Unique RNA-Seq Feature Extractor	3

Pytorch Lighning Framework

After the reviewing process, we created a Pytorch Lighning version of our model. For running this code, you need to run:

light.py -c config/light/[CONFIG NAME FILE]

for the TCGA-only experiments. And:

light_external.py -c config/light/[CONFIG NAME FILE]

for the configs that include "external" or "SCLC" in their filename.

Config Filename	Task	Notes	Table
STL_BRCA.yaml	Single		2,3
STL_LUAD.yaml	Single		2,3
STL_COAD.yaml	Single		2,3
STL_BRCA_external.yaml	Single	External validation	4
STL_LUAD_external.yaml	Single	External validation	4
STL_COAD_external.yaml	Single	External validation	4
MTL_TCGA.yaml	Multi	Only TCGA	2,4
MTL_train_SCLC_test.yaml	Multi	External validation	4

Example

main.py -c config/tcga_multi_dnn_trainer_task_cross_validation_4_bootstrap_unweighed_tpm_clinical_overall.yaml

Running the above results will yield the results for all three cancers in 'Without weighted random sampler' category in Table 3.

Expected output

Single task

[INFO]	2023-08-23 12:12:51,572 - 1 Fold for TCGA-BRCA...
[INFO]	2023-08-23 12:12:53,221 - epoch               : 10
[INFO]	2023-08-23 12:12:53,221 - train_auroc         : 0.73994 ±0.00000
[INFO]	2023-08-23 12:12:53,221 - train_auprc         : 0.27325 ±0.00000
[INFO]	2023-08-23 12:12:53,221 - train_c_index       : 0.71927 ±0.00000
[INFO]	2023-08-23 12:12:53,221 - train_recall        : 0.78689 ±0.00000
[INFO]	2023-08-23 12:12:53,221 - train_precision     : 0.16327 ±0.00000
[INFO]	2023-08-23 12:12:53,221 - train_loss          : 0.33689 ±0.03093
[INFO]	2023-08-23 12:12:53,221 - valid_auroc         : 0.60477 ±0.00000
[INFO]	2023-08-23 12:12:53,222 - valid_auprc         : 0.18471 ±0.00000
[INFO]	2023-08-23 12:12:53,222 - valid_c_index       : 0.59909 ±0.00000
[INFO]	2023-08-23 12:12:53,222 - valid_recall        : 0.30000 ±0.00000
[INFO]	2023-08-23 12:12:53,222 - valid_precision     : 0.27273 ±0.00000
[INFO]	2023-08-23 12:12:53,222 - valid_loss          : 0.34016 ±0.00701

The model will first train on each cancer dataset separately and then calculate the bootstraped results separately as well. Once the bootstrap for one cancer ends, the model will train on the data for a different cancer and then proceed to calculate the bootstrapped results.

[INFO]	2023-08-23 12:16:59,480 - bootstrap_auprc     : 0.35305 ±0.09661
[INFO]	2023-08-23 12:16:59,480 - bootstrap_c_index   : 0.55428 ±0.07729
[INFO]	2023-08-23 12:16:59,480 - bootstrap_recall    : 0.38975 ±0.20531
[INFO]	2023-08-23 12:16:59,480 - bootstrap_precision : 0.35371 ±0.17675

Each run takes between 5 to 7 seven minutes.

Multi task

Training:

[INFO]  1 Fold for TCGA_BLC...
[INFO]  epoch               : 10
[INFO]  train_auroc         : 0.74027 ±0.00000
[INFO]  train_auprc         : 0.44451 ±0.00000
[INFO]  train_c_index       : 0.70998 ±0.00000
[INFO]  train_recall        : 0.86752 ±0.00000
[INFO]  train_precision     : 0.27395 ±0.00000
[INFO]  train_loss          : 0.42498 ±0.04557
[INFO]  valid_auroc         : 0.73718 ±0.00000
[INFO]  valid_auprc         : 0.47157 ±0.00000
[INFO]  valid_c_index       : 0.71856 ±0.00000
[INFO]  valid_recall        : 0.69863 ±0.00000
[INFO]  valid_precision     : 0.30000 ±0.00000
[INFO]  valid_loss          : 0.41974 ±0.04541

Bootstrapping:

[INFO]  bootstrap_0_auroc                       : 0.83943 ±0.04359
[INFO]  bootstrap_1_auroc                       : 0.64477 ±0.05999
[INFO]  bootstrap_2_auroc                       : 0.71172 ±0.07293
[INFO]  bootstrap_0_auprc                       : 0.34872 ±0.09029
[INFO]  bootstrap_1_auprc                       : 0.50874 ±0.08170
[INFO]  bootstrap_2_auprc                       : 0.49778 ±0.10215
[INFO]  bootstrap_0_c_index                     : 0.82355 ±0.04298
[INFO]  bootstrap_1_c_index                     : 0.58646 ±0.04947
[INFO]  bootstrap_2_c_index                     : 0.69602 ±0.06749
[INFO]  bootstrap_0_recall                      : 0.77102 ±0.10841
[INFO]  bootstrap_1_recall                      : 0.50554 ±0.12414
[INFO]  bootstrap_2_recall                      : 0.63841 ±0.13585
[INFO]  bootstrap_0_precision                   : 0.27854 ±0.09747
[INFO]  bootstrap_1_precision                   : 0.55147 ±0.10589
[INFO]  bootstrap_2_precision                   : 0.38913 ±0.11199

Where the indices next to bootstrap_ correspond to each cancer.

Cancer type codes:

For the config files that use three datasets and the outputs of the Bootstrapped results, the following indices correspond to the datasets in this order:

1. BRCA: 0
2. LUAD: 1
3. COAD: 2

Data

Make sure to include the data for all three cancers in a folder called Data with subfolders Data/TCGA-BRCA, Data/TCGA-COAD, Data/TCGA-LUAD.