CPEM: Accurate cancer type classification based on somatic alterations using an ensemble of a random

이 논문에서는 돌연변이 프로필, 돌연변이율, 돌연변이 스펙트럼, 체세포 복제 수 등과 같은 다양한 입력 특징의 기여도를 평가하고, 이를 정확한 암 유형 분류에 활용하는 CPEM(Cancer Predictor using an Ensemble Model) 모델을 제안한다.

keywords: cancer, classification, ensemble type: 논문 리뷰 기수: 1기 기술 분류: 신약개발 날짜: 2020년 10월 18일 스터디 그룹: 대회참가(신약개발) 작성자: Chanran Kim

Abstract

• DNA sequencing technologies → 대규모 genomic data의 빠른 수집 → commonplace

• the classification of cancer type

  • based on somatic alterations(체세포 변화) detected from sequencing analyses

• In this study,

  • evaluate the contributions of various input features

    • such as,

      • mutation profiles

      • mutation rates

      • mutation spectra

      • signatures

      • somatic copy number alternations (can be derived from genomic data)

  • utilize them for accurate cancer type classification

  • CPEM (Cancer Predictor using an Ensemble Model)

  • tested on 7,002 samples representing over 32 different cancer types (TCGA DB)

  • feature selection - relatively high importance

  • 84% accuracy (nested 10-fold CV) → 94% accuracy (narrow down the target cancers to the six most common types)

Introduction

Workflow

1. feature vector 만들기

   • 22,421 gene-level mutation profiles

   • 2 mutation rates

   • 6 mutation spectra

   • 13,040 gene-level copy number alterations

   • 96 mutations signatures

2. dimension reduction

   1. LSVC(Linear Support Vector Classifier) based feature selection

3. train DNN, RandomForest → predict by Ensemble (Average)

Results

Experimental setup

• DNN

  • MLP

    • multinomial class cross-entropy w/ softmax loss

    • ReLU activation function

    • 3 layers (hidden_size: 2048)

    • Adam

    • lr: 1e-5

    • dropout: 40%

• 입력 feature의 optimal한 수와 feature selection 방법을 inner 10-fold CV에 대한 accuracy를 최대화하는 값으로 선택한다. Feature selection이 확정되면, 독립적인 테스트셋을 사용하여 CPEM 성능을 테스트

Efficacy of Various Mutation Features

• random forest classifier → initial prediction model

• measure changes in classification performance as each feature was added using 10-fold CV

  • gene-level mutation profiles (46.9%) + mutation rates (51.2%) + mutation spectra (58.5%) + gene-level SCNAs (61.0%) + mutation signatures (72.7%)

    • 특별히 무슨 관계가 있다기 보다는 결국 다 넣는게 젤 좋은거 아닌가요...? 어떤 점을 특별히 발견한 것인지 잘 모르겠습니다.

• select the top ten features based on the importance score from the classifier

  • 2 x mutation rates (2/2)

  • 1 x mutation signature (1/96)

  • 2 x mutation spectra (2/6)

  • 5 x mutated genes (5/22,421?)

• This result is consistent with previous studies that identified distinct mutational landscapes from many different types of cancer genomes

• This result confirms that genetic features play an important role in cancer initiation and progression, and also contribute to improving the accuracy of cancer classification

Optimal Feature Selection

• Feature Selection으로 10%만 선택 (크기를 90% 줄임)

• extra tree-based → random forest classifier

• LASSO, LSVM → DNN

Discussion

• kNN은 성능이 심각

  • 이런건 섞지 않는게 나음

Methods

Feature Selection

• Tree-based feature selection

  • feature importance calculated during the training of the decision tree classifier

• Lasso feature selection

  • w: regression coefficient vector

  • w는 입력 특성 xi를 필터링하기 위한 importance값으로 사용

  • w가 sparse하면 적은 features가 선택되고, 반대의 경우도 마찬가지.

• LSVC feature selection

  • squared hinge loss w/ L1-norm sparsity term

  • minimization 후에는 각 클래스에 대한 w가 sparse하므로, 0이 아닌 w 요소를 수집해서 feature select