[Paper Review] Learning Embeddings from Cancer Mutation Sets for Classification Tasks

Geoffroy Dubourg-Felonneau, Yasmeen Kussad+4 more|arXiv (Cornell University)|Nov 20, 2019

Machine Learning in Bioinformatics8 references4 citations

TL;DR

This paper proposes Flatsomatic, a variational autoencoder (VAE) that learns low-dimensional, informative embeddings from somatic mutation sets in cancer genomics. By compressing mutation position data into a latent space using a custom F1-score-based loss and beta-VAE regularization, the method achieves superior clustering performance over PCA and matches raw data in classification tasks, enabling better use of sparse, high-dimensional mutation profiles in precision oncology.

ABSTRACT

Analysis of somatic mutation profiles from cancer patients is essential in the development of cancer research. However, the low frequency of most mutations and the varying rates of mutations across patients makes the data extremely challenging to statistically analyze as well as difficult to use in classification problems, for clustering, visualization or for learning useful information. Thus, the creation of low dimensional representations of somatic mutation profiles that hold useful information about the DNA of cancer cells will facilitate the use of such data in applications that will progress precision medicine. In this paper, we talk about the open problem of learning from somatic mutations, and present Flatsomatic: a solution that utilizes variational autoencoders (VAEs) to create latent representations of somatic profiles. The work done in this paper shows great potential for this method, with the VAE embeddings performing better than PCA for a clustering task, and performing equally well to the raw high dimensional data for a classification task. We believe the methods presented herein can be of great value in future research and in bringing data-driven models into precision oncology.

Motivation & Objective

• To address the challenge of high dimensionality, sparsity, and variable set sizes in somatic mutation data for machine learning applications.
• To develop a robust, low-dimensional representation of somatic mutation profiles that preserves biologically relevant signals.
• To improve downstream tasks such as clustering and classification by learning unbiased, informative embeddings from mutation data.
• To explore alternative data representations that retain non-positional features and avoid sub-optimal pre-processing like binary mutation occurrence matrices.

Proposed method

• The method uses a variational autoencoder (VAE) to compress somatic mutation profiles into a low-dimensional latent space, with a focus on reconstruction fidelity and disentangled representations.
• Mutation data is preprocessed into binary occurrence vectors based on chromosomal position, with low-frequency mutations (frequency < 5) removed to reduce sparsity.
• The VAE employs a two-layer MLP encoder and decoder with batch normalization, Leaky ReLU, ReLU, and sigmoid activations, along with dropout and L1 regularization.
• A novel F1-score-based reconstruction loss is introduced and compared to binary cross-entropy, improving reconstruction performance on sparse mutation data.
• The regularization term is optimized using a warm-up schedule and beta-VAE to enhance disentanglement and generalization in the latent space.
• Performance is evaluated via clustering (NMI) and classification (F1-score), with comparisons to PCA and raw data.

Experimental results

Research questions

• RQ1 Can a variational autoencoder learn more informative low-dimensional representations of somatic mutation profiles than traditional methods like PCA?
• RQ2 Does a custom F1-score-based loss function improve reconstruction of sparse, high-dimensional mutation data compared to standard binary cross-entropy loss?
• RQ3 How does the size of the latent space affect the quality of the learned embeddings in terms of reconstruction and downstream task performance?
• RQ4 Can VAE-learned embeddings match the performance of raw mutation data in a classification task such as drug response prediction?
• RQ5 To what extent does the pre-processing step of creating a binary mutation occurrence matrix limit the model’s ability to capture biologically relevant features?

Key findings

• The MLP-based VAE with F1-score-based loss achieved a validation F1-score of 20.88% with a latent space size of 32, outperforming binary cross-entropy loss.
• The optimal latent space size for reconstruction was found to be greater than 8, with peak performance at 32 dimensions.
• The VAE embeddings achieved a Normalized Mutual Information (NMI) score of 21% in clustering, significantly outperforming PCA’s 11% NMI.
• In a drug response prediction task, Flatsomatic embeddings (64-dimensional) achieved an F1-score of 0.667, matching the raw mutation occurrence data (8298-dimensional).
• The use of a warm-up schedule for the beta-VAE regularization term improved model performance, indicating its importance in learning disentangled representations.
• The bidirectional LSTM-based VAE underperformed the MLP-VAE, achieving only 17.1% F1-score, suggesting that sequence modeling is less effective for unordered mutation sets.