AutoGenome: An AutoML Tool for Genomic Research

Denghui Liu, Chi Xu, Wenjun He, Zhimeng Xu, Wenqi Fu, Lei Zhang, Jie Yang, Guangdun Peng, Dali Han, Xiaolong Bai, Nan QiaobioRxiv 2019 · DOI

TL;DR

This paper introduces AutoGenome, an end-to-end AutoML framework designed to automate the process of building high-performance deep learning models for genomic profiling data. It features a novel Residual Fully-Connected Neural Network (RFCN) architecture and integrates automated hyperparameter and neural architecture search to simplify complex deep learning tasks for genomic researchers.

What Problem Does This Paper Address?

Deep learning has achieved remarkable success in fields like computer vision and natural language processing, inspiring its application in genomics. However, applying deep learning to genomic data, particularly non-sequence data like RNA expression or gene mutation profiles, presents challenges. Traditional architectures like CNNs and RNNs are not well-suited for data lacking inherent spatial or temporal relationships. While MLPs and Autoencoders are used, they often struggle with training very deep networks due to vanishing gradients, leading to shallower models. Furthermore, selecting appropriate architectures and hyperparameters is a complex and time-consuming task, especially for researchers new to deep learning. AutoGenome aims to democratize deep learning for genomics by automating these complex choices and providing interpretable results.

Methods

The core of AutoGenome is its novel Residual Fully-Connected Neural Network (RFCN) architecture. Unlike traditional residual networks that use convolutional layers, RFCN employs fully-connected layers as its basic unit. This allows for more flexible connections, including direct layer-to-layer connections, skip connections, or branched connections. Two variants are proposed: RFCN-ResNet (summation of previous output and input) and RFCN-DenseNet (concatenation of previous outputs).

AutoGenome further incorporates Automated Machine Learning (AutoML) techniques, specifically:

  • Hyperparameter Search: Automatically tunes parameters like learning rate, batch size, momentum, and weight decay.
  • Neural Architecture Search (NAS): Employs the Efficient Neural Architecture Search (ENAS) algorithm to automatically discover optimal RFCN architectures. This is achieved by searching within a directed acyclic graph (DAG) representation of possible architectures.
  • Feature Importance Estimation: Integrates the SHAP (SHapley Additive exPlanations) package to provide insights into which genes are most influential for model predictions, aiding interpretability.

For unsupervised learning, AutoGenome extends RFCN to a Residual Fully-Connected Variational Autoencoder (Res-VAE), which also incorporates skip connections to improve feature propagation.

Datasets

The paper utilizes publicly available datasets for evaluation:

  • Pan-cancer classification: Gene expression profiles and somatic mutation profiles from 8,127 samples across 24 tumor types from The Cancer Genome Atlas (TCGA).
  • Mouse single-cell classification: Gene expression profiles from 10,000 mouse single cells across 10 embryonic developmental stages.
  • 10X PBMC single-cell RNA-seq: Processed expression matrix and cell labels from the 10X Genomics platform, containing 16,653 genes for 4,271 single cells across 8 predefined cell types.

Key Insights and Ideas

  • RFCN's Adaptability: The RFCN architecture, by using fully-connected layers and flexible residual connections, proves more suitable for genomic profiling data than traditional CNNs/RNNs, and offers advantages over standard MLPs and VAEs.
  • Deep Networks for Genomics: RFCN architectures can be trained to be significantly deeper (up to 11 hidden layers) than typical MLP-based models in genomics, which often have fewer than four hidden layers. This depth appears crucial for achieving better performance.
  • AutoML for Genomics is Essential: The complexity of deep learning model selection and tuning for genomic data highlights the need for AutoML tools like AutoGenome to democratize its use.
  • Interpretability Matters: Beyond prediction, AutoGenome's integration of SHAP values provides crucial insights into the biological drivers behind model predictions, a critical aspect for scientific discovery.
  • RRFCN for Novel Architectures: The randomly-wired RFCN (RRFCN) variant, enabled by NAS, can discover entirely new, problem-specific architectures, pushing the boundaries of what's possible.

Key Results

  • Superior Performance: RFCN-ResNet and RFCN-DenseNet consistently outperformed XGBoost and AutoKeras in pan-cancer classification (achieving up to 95.9% accuracy) and mouse single-cell classification (up to 96.5% accuracy).
  • Deep RFCNs: The best RFCN models achieved depths of 6-7 layers (RFCN-ResNet, RRFCN) and 11 layers (RFCN-DenseNet), demonstrating the ability to train deeper networks effectively.
  • RRFCN's Novelty: The RRFCN variant, discovered through NAS, achieved the best accuracy in pan-cancer classification (96.3% for gene expression, 68.1% for mutation data) and showed promising results in single-cell classification.
  • Res-VAE Effectiveness: The residual VAE (Res-VAE) demonstrated clearer boundaries and tighter clusters in single-cell data visualization compared to PCA, t-SNE, and standard VAE, indicating better unsupervised representation learning.
  • Ease of Use: AutoGenome allows researchers to perform complex deep learning analyses with as few as five lines of code.

Limitations and Future Work

  • Fixed RFCN Architectures: While RRFCN uses NAS to find novel architectures, the base RFCN-ResNet and RFCN-DenseNet have fixed skip-connection structures, which might not be optimal for all problems.
  • Depth vs. Accuracy: The study found that increasing depth beyond 6-7 layers for RRFCN did not necessarily improve accuracy, suggesting an optimal range for current genomic problems.
  • Future Work: The authors envision AutoGenome becoming a popular tool in genomic studies. Future work could involve expanding the range of supported genomic data types and model architectures. The open-source nature of AutoGenome encourages community contributions and further development.

Summary updated

Abstract

Abstract Deep learning have made great successes in traditional fields like computer vision (CV), natural language processing (NLP) and speech processing. Those achievements greatly inspire researchers in genomic study and make deep learning in genomics a very hot topic. Convolutional neural network (CNN) and recurrent neural network (RNN) are frequently used for genomic sequence prediction problems; multiple layer perception (MLP) and auto-encoders (AE) are frequently used for genomic profiling data like RNA expression data and gene mutation data. Here, we introduce a new neural network architecture, named residual fully-connected neural network (RFCN) and demonstrate its advantage for modeling genomic profiling data. We further incorporate AutoML algorithms and implement AutoGenome, an end-to-end automated genomic deep learning framework. By utilizing the proposed RFCN architectures, automatic hyper-parameter search and neural architecture search algorithms, AutoGenome can train high-performance deep learning models for various kinds of genomic profiling data automatically. To make researchers better understand the trained models, AutoGenome can assess the feature importance and export the most important features for supervised learning tasks, and the representative latent vectors for unsupervised learning tasks. We envision AutoGenome to become a popular tool in genomic studies.
BibTeX
@misc{liu2019autogenome,
  author = {Denghui Liu and Chi Xu and Wenjun He and Zhimeng Xu and Wenqi Fu and Lei Zhang and Jie Yang and Guangdun Peng and Dali Han and Xiaolong Bai and Nan Qiao},
  title = {{AutoGenome: An AutoML Tool for Genomic Research}},
  year = {2019},
  doi = {10.1101/842526},
  howpublished = {bioRxiv},
}
Figure