The Application of Deep Learning for Cancer Subtype Identification (UCT Prague)

Abstract:

Human cancer is a heterogeneous disease initiated by random somatic mutations and driven by multiple genomic alterations. Cancer subtype classification has the potential to significantly improve disease prognosis and move closer toward personalized cancer treatment and prevention. The assumption behind molecular subtyping is that patients of similar gene expression patterns are likely to have similar responses to therapies and clinical outcomes due to the tumor microenvironment. Thus, molecular subtyping can reveal information valuable for a range of cancer studies from cancer initiation and tumor biology to prognosis and personalized medicine. Molecular subtyping was widely described in breast and colorectal cancer, however, it was not broadly explored in other indications. In this work, molecular subtypes are explored via deep learning models, specifically the standard and variational autoencoder latent space. The analysis is performed on a range of publicly available datasets such as Gene Expression Omnibus (GEO), The Cancer Genome Atlas (TCGA) and the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC).

Goals:

• Prepare the data for modeling
• Quality control of the data with classification framework DeepCC
• Implementation of the standard and variational autoencoder and baseline models
• Comparision of models

Data analysis overview:

• Data preparation steps:
  • To unify gene names and gene IDs, Affymetrix / Illumina / Agilent probeset IDs were converted to Entrez ID
  • Removal of unlabelled samples and unification of subtype in case of breast and colorectal cancer datasets
  • Removal of non-tumor tissue samples
  • Replacing missing expression values with zero constant
  • Removal of technical variation with conversion of genes to functional spectra
  • Principal component analysis (PCA) to reduce the dimensionality of the data
  • Oversampling of data with adaptive synthetic sampling approach (ADASYN) to solve imbalancing problem
• Used models:
  • Standard Autoencoder
  • Variational Autoencoder
  • kernel PCA with poly and sigmoid kernels
  • PCA
  • Data without compression

Data:

• Input data is represented by a matrix, where columns are sample names and rows are ENTREZ ID; in some cases the data were directly downloaded from GEO database with GEOparse package, otherwise data were loaded from cBioPortals or Bioconductor packages (breastCancerNKI, breastCancerUNT, breastCancerTRANSBIG, and breastCancerUPP); colorectal cancer datasets were loaded from Synapse
• Data collection was the most time consuming step as majority of GEO gene expression datasets have around 100 samples
• A list of datasets used for the project is shown below

Cancer	ID	Technology	Tissue type	Normalization	Number of samples	Source
CRC	GSE33113	Affymetrix HG133plus2	Fresh frozen	fRMA	80	Synapse
CRC	GSE39582	Affymetrix HG133plus2	Fresh frozen	fRMA	466	Synapse
CRC	KFSYSCC	Affymetrix HG133plus2	Fresh frozen	fRMA	229	Synapse
CRC	GSE35896	Affymetrix HG133plus2	Fresh frozen	fRMA	51	Synapse
CRC	PETACC3	Almac's Affymetrix Array	FFPE	fRMA	526	Synapse
CRC	GSE13067	Affymetrix HG133plus2	Fresh frozen	fRMA	56	Synapse
CRC	GSE20916	Affymetrix HG133plus2	Fresh frozen	fRMA	45	Synapse
CRC	GSE23878	Affymetrix HG133plus2	Fresh frozen	fRMA	24	Synapse
CRC	GSE14333	Affymetrix HG133plus2	Fresh frozen	fRMA	129	Synapse
CRC	GSE2109	Affymetrix HG133plus2	Fresh frozen	RMA	244	Synapse
CRC	GSE17536	Affymetrix HG133plus2	Fresh frozen	fRMA	147	Synapse
CRC	GSE13294	Affymetrix HG133plus2	Fresh frozen	fRMA	124	Synapse
CRC	GSE37892	Affymetrix HG133plus2	Fresh frozen	fRMA	107	Synapse
CRC	TCGA	RNA sequencing	Primary Tumor	RSEM	207	Synapse
BRCA	GSE96058	RNA sequencing	Primary Tumor	FPKM	3273	GEO
BRCA	GSE65194	Affymetrix HG133plus2	Primary Tumor	GC-RMA log2	98	GEO
BRCA	GSE81538	RNA sequencing	Primary Tumor	FPKM	405	GEO
BRCA	ROCK	Affymetrix HG133plus2	Primary Tumor	log2 andRMA	1468	GEO
BRCA	METABRIC	Illumina Array	Primary Tumor	quantile normalization	1898	cBioPortal
BRCA	TCGA	RNA sequencing	Primary Tumor	RSEM	981	cBioPortal
BRCA	CPTAC	RNA sequencing	Primary Tumor	TPM	122	cBioPortal
BRCA	SMC	RNA sequencing	Primary Tumor	TPM	168	cBioPortal
BRCA	NKI	Agilent Rosetta Array	Primary Tumor	N/A	337	Bioconductor
BRCA	TRANSBIG	Affymetrix HG U133A Array	Primary Tumor	MAS5 normalized	198	Bioconductor
BRCA	UPP	Affymetrix HG U133A Array	Primary Tumor	log transformed and scaled by adjusting the mean intensity	190	Bioconductor
BRCA	UNT	Affymetrix HG U133A Array	Primary Tumor	RMA	92	Bioconductor
BRCA	GSE60789	Illumina Array	Primary Tumor	Illumina BeadStudio software	55	GEO
BRCA	GSE59595	Illumina Array	Primary Tumor	RSN	133	GEO
BRCA	GSE53031	Affymetrix HG 219 Array	Primary Tumor	RMA	167	GEO
BRCA	GSE56493	Rosetta/Merck HG2plus2	FNA	RMA	120	GEO
BRCA	GSE78958	Affymetrix HG133plus2	Primary Tumor	Partek Genomics Suite software	424	GEO
BRCA	GSE20711	Affymetrix HG133plus2	Primary Tumor	RMA	88	GEO
BRCA	GSE45827	Affymetrix HG133plus2	Primary Tumor	REML	130	GEO
BRCA	GSE21653	Affymetrix HG133plus2	Primary Tumor	RMA	242	GEO
BRCA	GSE25066	Affymetrix HG U133A Array	Primary Tumor	MAS5 normalized	508	GEO
BRCA	GSE48390	Affymetrix HG133plus2	Primary Tumor	RMA	74	GEO
BRCA	GSE135298	RNA sequencing	Primary Tumor	FPKM	93	GEO
Total	-	-	-	-	13700	-

Tissue types:

• FFPE - a formalin-fixed paraffin-embedded tissue
• FNA - a fine needle aspiration biopsy.

Normalization:

• FPKM - Fragments Per Kilobase per Million mapped fragments
• REML - Restricted Maximum Likelihood
• (f)RMA - (frozen) Robust Multi-array Average
• RSEM - RNA-Seq by Expectation-Maximization
• RSN - Robust Spline Normalization

Environment

Data processing, modeling and evaluation were implemented using Python 3.8.10 and Google Colab 0.0.1a2. The following data science and machine learning packages were used: imbalanced-learn 0.8.1, biomart 0.9.2, keras 2.6.0, numpy 1.22.4, pandas 1.3.5, scikit-learn 1.2.1, scipy 1.10.1, keras 2.11.0, rpy2 3.5.5, umap 0.5.3, and seaborn 0.11.2, matplotlib 3.5.1 and plotly 5.5.0 were used for plotting.

• Preprocessing of the data were conducted on the “KOIOS” high performance computing (HPC) system
  • OS: Gentoo 2.9
  • Kernel: x86_64 Linux 5.15.26-gentoo-x86_64
  • Shell: bash 5.1.16
  • CPU: Intel Xeon Gold 6238 @ 88x 3.7GHz
  • GPU: ASPEED Technology, Inc. ASPEED Graphics Family (rev 41)
  • RAM: 1081 GB
• Modeling was performed on Google Colab machine:
  • GPU: NVIDIA A100-SXM4-40GB
  • RAM: 25.5 GB
  • CPU: Intel(R) Xeon(R) CPU @ 4x, 2.00GHz