Deep Learning Benchmarks on L1000 Gene Expression Data.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Gene expression data offers physiological insights beyond genomic sequences.
• Developing high-capacity machine learning (ML) models for gene expression analysis requires robust benchmarks and baseline comparisons.
• Current ML model development is hindered by a lack of standardized tasks and characterized baseline performance.

Purpose of the Study:

• To establish benchmark tasks and baseline performance metrics for ML models analyzing gene expression data.
• To enable direct comparisons for future methodological work in deep learning for gene expression analysis.
• To provide curated datasets and benchmark tasks for the research community.

Main Methods:

• Profiling various classifiers, including linear models, random forests, decision trees, K-nearest neighbors (KNN), and feed-forward artificial neural networks (FF-ANNs).
• Testing graph convolutional neural networks (GCNNs), a novel method incorporating prior biological domain knowledge.
• Utilizing two curated views of the public LINCS corpus and one private dataset for biologically motivated tasks.

Main Results:

• Graph convolutional neural networks (GCNNs) demonstrated high performance, particularly on large datasets.
• Feed-forward artificial neural networks (FF-ANNs) consistently achieved strong performance across tasks.
• Among non-neural network classifiers, linear models and KNN classifiers were the most effective.

Conclusions:

• The established benchmarks and curated datasets facilitate direct comparisons for novel ML methods.
• GCNNs show significant potential for gene expression data analysis by integrating biological knowledge.
• This work provides a foundation for advancing deep learning methodologies in the field of genomics.