BCDForest: a boosting cascade deep forest model towards the classification of cancer subtypes based on gene expression data

Abstract

Background: The classification of cancer subtypes is of great importance to cancer disease diagnosis and therapy. Many supervised learning approaches have been applied to cancer subtype classification in the past few years, especially of deep learning based approaches. Recently, the deep forest model has been proposed as an alternative of deep neural networks to learn hyper-representations by using cascade ensemble decision trees. It has been proved that the deep forest model has competitive or even better performance than deep neural networks in some extent. However, the standard deep forest model may face overfitting and ensemble diversity challenges when dealing with small sample size and high-dimensional biology data.

Results: In this paper, we propose a deep learning model, so-called BCDForest, to address cancer subtype classification on small-scale biology datasets, which can be viewed as a modification of the standard deep forest model. The BCDForest distinguishes from the standard deep forest model with the following two main contributions: First, a named multi-class-grained scanning method is proposed to train multiple binary classifiers to encourage diversity of ensemble. Meanwhile, the fitting quality of each classifier is considered in representation learning. Second, we propose a boosting strategy to emphasize more important features in cascade forests, thus to propagate the benefits of discriminative features among cascade layers to improve the classification performance. Systematic comparison experiments on both microarray and RNA-Seq gene expression datasets demonstrate that our method consistently outperforms the state-of-the-art methods in application of cancer subtype classification.

Conclusions: The multi-class-grained scanning and boosting strategy in our model provide an effective solution to ease the overfitting challenge and improve the robustness of deep forest model working on small-scale data. Our model provides a useful approach to the classification of cancer subtypes by using deep learning on high-dimensional and small-scale biology data.

Keywords: Cancer subtype; Cascade forest; Classification.

Fig. 1

Illustration of cascade forest structure. Each level of the cascade consists of two random forests (black) and two completely random forests (red). Suppose there are three classes to predict; each forest outputs a three-dimensional class vector, which is then concatenated for representation of original input [23]

Fig. 2

Illustration of boosting cascade forest structure. Each level of the cascade consists of two random forests (black) and two completely random forests (red). The standard deviation of top-k important features in each forest will compose a new feature to be concatenated in the next cascade level to emphasize the discriminative features

Fig. 3

Illustration of multi-class-grained scanning. a Suppose four classes (A, B, C and D) in training dataset. For each class, we produce the positive and negative sub-datasets, and then use the sub-datasets to train a binary random forest classifier. Four different types of random forests will be produced by using different training datasets (sliding window based). The out-of-bagging (OOB) score of each forest is used to calculate a normalized quantity weight to each forest. b Based on the fit forests and their quantity weights, a 500-dim instance vector can be transformed to a concatenated 1604-dim representation

Fig. 4

Overall procedure of BCDForest. Suppose there are four classes, and the sliding windows are 100-dim and 200-dim. Two cascade layers are used to give final prediction

Fig. 5

Comparison of different methods on large-scale pan-cancers dataset. Each dot presents the performance of each corresponding method on each cancer type. 11 cancer types were included in the pan-cancers dataset

Fig. 6

Comparison of BCDForest and gcForest on three cancer type datasets (BRCA, GBM and LUNG). Each dot presents the performance of each method on each cancer subtype class

Fig. 7

Comparison of overall performance of BCDForest and gcForest on BRCA, GBM and LUNG datasets. The average precision, recall and F-1 score on all subtype classes of each dataset were evaluated

Fig. 8

Comparison of overall performance of BCDForest and gcForest on COAD datasets. The average precision, recall and F-1 score on all subtype classes of each dataset were evaluated

Fig. 9

Comparison of overall performance of BCDForest and gcForest on LIHC datasets. The average precision, recall and F-1 score on all subtype classes of each dataset were evaluated