A negative selection heuristic to predict new transcriptional targets

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Supervised machine learning (ML) advances gene regulatory network inference using transcriptomic and proteomic data.
• Traditional methods often struggle with the scarcity of negative examples in gene regulatory inference.
• This limitation can hinder supervised classifier performance, especially with limited training data.

Purpose of the Study:

• To enhance supervised identification of transcriptional targets by improving training data.
• To address the challenge of limited negative examples in gene regulatory inference.
• To develop a method for selecting reliable negative examples from unlabeled data.

Main Methods:

• Introduced a heuristic approach leveraging known transcriptional network topology.
• Selected reliable counter-negative examples from unlabeled datasets.
• Empirically evaluated the heuristic using Escherichia coli datasets.

Main Results:

• The heuristic restores a conventional positive/negative training condition, significantly improving classification performance.
• Achieved 60% precision in predicting BCL6 direct core targets in human B cells.
• Demonstrated the effectiveness of the negative example selection strategy.

Conclusions:

• The scarcity of positive examples negatively impacts supervised classifier performance in learning transcriptional relationships.
• Selecting reliable negative examples, inspired by text mining, boosts classifier performance.
• This approach opens new avenues for identifying novel transcriptional targets.