A Data Adaptive Biological Sequence Representation for Supervised Learning.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Gene expression is crucial for organism function, and DNA microarray technology enables large-scale monitoring.
• Understanding gene regulation mechanisms requires analyzing the relationship between gene expression and nucleotide sequences.
• Identifying local DNA elements (motifs) is key to inferring these regulatory relationships.

Purpose of the Study:

• To propose a novel data-adaptive representation approach for supervised learning on biological sequences.
• To develop a method for predicting biological responses based on sequence data.
• To address challenges in high-dimensionality and missing values common in biological sequence analysis.

Main Methods:

• The study introduces the Sliding Window-Random Forest (SW-RF) method for categorical sequence representation.
• SW-RF represents sequences using overlapping subsequences and a tree-based learner to create a bag-of-words-like representation.
• A lasso logistic regression classifier is trained on the learned representation to identify important patterns.

Main Results:

• The SW-RF approach demonstrated significantly improved accuracy on synthetic and DNA promoter sequence data.
• The method efficiently handles missing values in microarray datasets, a common challenge.
• The learned representation allows for the application of various classifiers for pattern identification.

Conclusions:

• SW-RF provides an effective feature-based representation for categorical sequences, particularly in bioinformatics.
• The approach enhances the accuracy of predicting biological responses from sequence data.
• SW-RF's flexibility makes it applicable to diverse categorical sequence data beyond biological applications.