Hongyang449/proteome_prediction

Transfer Learning Improves The State of The Art for Protein Abundance Prediction in Cancers

This is the package of our winning algorithm in the NCI-CPTAC DREAM Proteogenomics Challenge.

background: Proteogenomics Challenge

see also: Hongyang Li and Yuanfang Guan's 1st Place Solution

Please contact (hyangl@umich.edu or gyuanfan@umich.edu) if you have any questions or suggestions.

---

Installation

Git clone a copy of code:

git clone https://github.com/GuanLab/proteome_prediction.git

Required dependencies

• R (3.4.3)
• python (3.6.5)
• numpy (1.13.3). It comes pre-packaged in Anaconda.
• scikit-learn (0.19.0) A popular machine learning package. It can be installed by:

pip install -U scikit-learn

Dataset

All the omic data are 2D matrices, where columns are cancer samples and rows are genes/proteins. The CNV and RNA-seq data originally came from TCGA. The proteomic data originally came from CPTAC.

We directly downloaded the data from the challenge website and more details can be found at: https://www.synapse.org/#!Synapse:syn8228304/wiki/448372

To run the code, download the following omic data from here and put them into the directory data/raw/

• retrospective_breast_CNA_sort_common_gene_16884.txt
• retrospective_breast_proteome_sort_common_gene_10005.txt
• retrospective_breast_RNA_sort_common_gene_15107.txt
• retrospective_ova_CNA_sort_common_gene_11859.txt
• retrospective_ova_JHU_proteome_sort_common_gene_7061.txt
• retrospective_ova_PNNL_proteome_sort_common_gene_7061.txt
• retrospective_ova_rna_seq_sort_common_gene_15121.txt

Then preprocess the data and generate 5-fold cross validatation using code in

• data/trimmed_set
• data/normalization
• data/cv_set

Four models

We have two sets of code in parallel

• prediction/breast
• prediction/ova

1. generic model

This model directly approximates the protein level based on the corresponding mRNA level.

prediction/breast/rna/

2. gene-specific model

This model considers the gene-gene interactions in regulating protein abundance, in which mRNA levels of all genes were used as features to make predictions. The base learner is random forest with maximum depth of 3 and 100 trees.

prediction/breast/individual/

3. trans-tissue model

Similar to the gene-specific model, this model uses combined samples from breast and ovarian cancer samples.

prediction/breast/individual_transplant/

4. ensemble model

Our final results are the ensemble of the 1-3 models mentioned above.

prediction/breast/final/

5. result analysis and figure preparation

analysis/