Convolutional neural network models for cancer type prediction based on gene expression

Milad Mostavi; Yu-Chiao Chiu; Yufei Huang; Yidong Chen

Source-linked AI summary

Convolutional neural network models for cancer type prediction based on gene expression

Milad Mostavi, Yu-Chiao Chiu, Yufei Huang, Yidong Chen

arXiv:1906.07794v1Jun 18, 2019q-bio.GNcs.LGq-bio.QM

TL;DR

Cancer-type prediction and marker identification can be affected by tissue of origin. This paper presents three CNN architectures for gene-expression classification and reports 95.7% prediction accuracy, alongside model interpretation for elucidating cancer markers.

Problem
The study addresses tissue-of-origin effects in cancer classification and their relevance to interpreting cancer marker genes.
Method
The authors present three CNN architectures that classify pan-cancer RNA-Seq gene-expression data.
Results
95.7% prediction accuracy was achieved for cancer-type classification, comparable to earlier published studies.
Takeaways & Limitations
The model interpretation scheme helps elucidate biologically relevant cancer marker genes after accounting for tissue of origin.
Takeaways & Limitations
The authors report low confidence when classifying cancer types with few marker genes.

Abstract

from arXiv · show

Background Precise prediction of cancer types is vital for cancer diagnosis and therapy. Important cancer marker genes can be inferred through predictive model. Several studies have attempted to build machine learning models for this task however none has taken into consideration the effects of tissue of origin that can potentially bias the identification of cancer markers. Results In this paper, we introduced several Convolutional Neural Network (CNN) models that take unstructured gene expression inputs to classify tumor and non-tumor samples into their designated cancer types or as normal. Based on different designs of gene embeddings and convolution schemes, we implemented three CNN models: 1D-CNN, 2D-Vanilla-CNN, and 2D-Hybrid-CNN. The models were trained and tested on combined 10,340 samples of 33 cancer types and 731 matched normal tissues of The Cancer Genome Atlas (TCGA). Our models achieved excellent prediction accuracies (93.9-95.0%) among 34 classes (33 cancers and normal). Furthermore, we interpreted one of the models, known as 1D-CNN model, with a guided saliency technique and identified a total of 2,090 cancer markers (108 per class). The concordance of differential expression of these markers between the cancer type they represent and others is confirmed. In breast cancer, for instance, our model identified well-known markers, such as GATA3 and ESR1. Finally, we extended the 1D-CNN model for prediction of breast cancer subtypes and achieved an average accuracy of 88.42% among 5 subtypes. The codes can be found at https://github.com/chenlabgccri/CancerTypePrediction.

Background

The proposed model is designed to elucidate biologically relevant cancer marker genes after eliminating tissue-of-origin effects. Its light hyperparameter requirements may facilitate future cancer diagnosis.

The model aims to elucidate biologically relevant cancer marker genes after eliminating tissue-of-origin effects.
The model has light hyperparameters for training.
The model can be easily adapted to facilitate cancer diagnosis in the future.

Background

Cancer diagnosis and early detection motivate machine-learning approaches that can classify tumor types and identify biomarkers. Existing methods overlook tissue-of-origin effects, prompting systematic comparison of CNN input formats and interpretation methods for extracting diagnostic markers.

Motivation: Cancer diagnosis and early detection require methods that classify tumor types and identify cancer markers for clinical use.Novel machine-learning methods are motivated by the need to discover biomarkers, while detection screens have reported sensitivity around 70-80%.
Problem: These approaches ignore tissue of origin within cancer types, making interpretation unlikely to distinguish tissue-specific from cancer-type-specific genes.Consequently, functional analysis and biomarker selection for cancer detection cannot be reliably performed from such models.
Contribution: The study systematically compares CNN models built for different input data formats and interrogates the capacity of their convolution kernels.It evaluates tumor-type prediction before and after removing tissue-specific gene-expression influence using TCGA cancer and normal samples.
Contribution: A model-interpretation scheme examines genes involved in deep-learning predictions to assess whether CNNs can yield diagnostic markers for future validation.The approach is presented as demonstrating distinctive CNN characteristics and the feasibility of extracting diagnostic markers.

Methods

The study used TCGA pan-cancer RNA-seq data, filtered and standardized gene-expression inputs, and developed three shallow CNN architectures with distinct input and convolution designs. Guided gradient saliency was then used to define cancer marker genes from model-derived gene effects.

Data and preprocessing: 10,340 cancer samples spanning 33 cancer types and 731 samples from 23 normal tissues were downloaded from TCGA using TCGAbiolinks in December 2018.The study also collected PAM50 subtypes for 864 breast cancer samples and added Gaussian noise with standard deviations of 0–500% of each gene’s average to test robustness.
Data and preprocessing: Gene expression was represented as log2(FPKM + 1), low-information genes were removed using mean and standard-deviation thresholds, and 7,091 genes remained.Genes with mean < 0.5 or st. dev. < 0.8 across all TCGA samples were filtered to reduce non-informative or noise-sensitive features.
CNN architectures: The CNN designs were constrained to one convolutional layer, favoring shallow models to reduce overfitting and training-resource demands when samples are limited relative to parameters.The 2D-Hybrid-CNN was designed to combine 2D inputs with simple 1D convolution operations and capture global unstructured features.

Results

The CNN models converged rapidly without obvious overfitting, but normal-sample evaluation revealed tissue-of-origin effects. Adding a normal output class supported robust cancer classification and enabled marker-gene interpretation, including differential-expression evidence and breast-cancer markers.

Tissue-of-origin effects: 19 of 23 normal classes were classified into their corresponding cancer types, indicating that the model partially recognized tissues of origin.Kidney, liver, lung, and digestive-system normal samples were clearly grouped together.
Tissue-of-origin correction: A 34th prediction node labeled “Normal” was introduced to remove tissue-of-origin traces from cancer samples and improve robust cancer-type prediction.The retrained models used 33 tumor-class nodes plus one normal-sample node.
Marker-gene interpretation: 2,090 to 91 unique marker genes were identified as the saliency threshold increased from score > 0.5 to > 0.9, averaging ~108 markers per cancer type.Marker counts ranged from 4 to 346 across the 34 classes; DLBC, BRCA, and PRAD had the most markers.
Breast-cancer subtype prediction: 88.3% average precision was achieved for the evaluated breast-cancer subtype prediction task.Further details were reported in Table 2.

Discussion

The discussion compares three CNN architectures, emphasizing the 1D-CNN’s simplicity, comparable accuracy, robustness to noise, and suitability for subsequent interpretation. It also reports tissue-origin control, high-order gene interactions, and avenues for future improvement.

Architecture comparison: 1D-CNN and 2D-Hybrid-CNN achieved comparable accuracy (95.7%), slightly improving the result (95.6%) in.The three proposed CNN architectures addressed prediction and interpretation of unstructured gene expressions.
Architecture comparison: The 1D-CNN reduced hyperparameters from 26 millions to ~200 thousand through a much-simplified design.It requires no particular input order and has only one convolutional layer, which is desirable when genomic data are difficult and costly to collect.
Robustness: 75% noise ratio marked the point until which both models remained extremely robust; above it, 1D-CNN outperformed 2D-Vanilla-CNN.The performance of both models gradually dropped after this threshold, indicating more stable 1D-CNN performance under unwanted noise.
Tissue-origin control: 92.5% precision was achieved for normal-tissue prediction, while only 2 of 128 normal kidney samples were classified into cancer groups.Combining tumor and normal samples with 34 prediction nodes was intended to eliminate tissue-of-origin influence, although kidney cancers remained difficult to distinguish.
Future directions: Future work may incorporate GTEx transcriptomic data and additional genome-wide profiles, including DNA mutation, copy number variation, and DNA methylation.These additions were proposed to investigate tissue-origin issues and enrich model-training inputs to improve classification accuracy.
Interpretation: The CNN classified tumors into correct subtypes despite modest gene-effect differences, suggesting that simple expression high-versus-low rules do not explain its predictions.The model’s benefit was capturing high-order interactions among genes to make accurate predictions.

Conclusions

The study presented three CNN architectures for cancer type prediction from high-dimensional gene-expression inputs while considering tissue of origin. The model achieved 95.7% prediction accuracy and enabled interpretation of cancer markers for each cancer type, with future refinement aimed at earlier-cancer markers.

Conclusions: Three CNN architectures predicted cancer types from high-dimensional gene-expression inputs while considering tissue of origin.The conclusion describes these architectures as unique.
Conclusions: 95.7% prediction accuracy was equivalent to earlier published studies.The conclusion pairs this accuracy with comparison to earlier studies.
Conclusions: The CNN construction was drastically simplified and achieved a significant reduction from tissue of origin.The conclusion presents these as advantages relative to earlier studies.
Conclusions: The interpretable CNN model elucidated cancer markers for each cancer type, with future refinement intended to identify markers for earlier cancer.The conclusion links model interpretation to cancer-marker discovery and identifies earlier-cancer markers as a future goal.

Declarations

The article provides its supporting dataset and reports no competing interests. It acknowledges partial research and publication-cost funding, states that funders had no role in the study, and describes author contributions.

Data availability: The dataset supporting the article’s conclusions is included within the article.
Competing interests: The authors declare that they have no competing interests.
Funding: Research and publication costs were partially supported by NCI Cancer Center Shared Resources, NIH, CPRIT, and SALSI awards.
Funding: The funding sources had no role in study design, data collection, analysis, interpretation, or manuscript writing.
Author contributions: All authors conceived the study, interpreted the data, wrote the manuscript, and approved its final version; MM and YCC designed the model and performed data analysis.

Tables

The tables document hyperparameter selection for the 1D-CNN and 2D-Vanilla-CNN models, breast cancer subtype classification, and the hyperparameters and training times of four CNN models. All models used early stopping, with most using patience = 4, and training included 5-cv evaluation on a specified Linux server.

Hyperparameter selection: Table 1A compares 1D-CNN hyperparameter settings using trained and tested statistical measures, with final parameters highlighted.The table identifies the selected configuration among the tested settings.
Hyperparameter selection: Table 1B compares 2D-Vanilla-CNN hyperparameter settings using trained and tested statistical measures, with final parameters highlighted.The table identifies the selected configuration among the tested settings.
Breast cancer subtype classification: Table 2 reports breast cancer subtype classification results using the 1D-CNN model.The table focuses specifically on subtype classification rather than broad cancer-type classification.
Model training: Table 3 lists hyperparameters and training times for four CNN models.The table enables comparison of model configurations and training duration.
Implementation: Early stopping was used for all models with patience = 4 except the last model, results included 5-cv, and training ran on a Linux server with a Xeon 8176 CPU @2.1GHz and 4x28 cores.The implementation notes specify the stopping criterion, validation procedure, and hardware environment.

Figures

The figures depict three CNN architectures for gene-expression classification, contrasting 1D and 2D input/convolution designs. They also summarize model learning and prediction performance, interpretability analyses of 1D-CNN markers, and robustness testing under additive Gaussian noise.

Model architectures: The three CNN designs use a 7100-gene vector, a 100x71 matrix with one convolution layer, or parallel vertical and horizontal convolution layers.The 1D-CNN uses vector-formatted input; the 2D-Vanilla-CNN reformats input as a 100x71 matrix; and the 2D-Hybrid-CNN uses two parallel convolution layers.
Performance evaluation: Figures compare learning curves, micro-averaged accuracy, precision, recall, and confusion matrices across tumor-only and combined tumor-normal training settings.The comparisons include three models, 33 tumor types, and a 1D-CNN trained with 33 tumor types plus normal samples.
Model interpretation: Interpretation figures examine gene-effect score distributions, marker-based t-SNE plots, class-specific markers, differential expression, BRCA marker effects, and associated functions.Differential expression is expressed as the absolute difference between a class and all other samples in log2(FPKM+1), with marker identification shown at gene-effect score > 0.5.
Robustness testing: Noise-testing figures report classification accuracy when models trained on original TCGA data are tested on TCGA data with additive Gaussian noise.The tests vary the amount of additive Gaussian noise in the inputs and compare both classifiers under noisy conditions.