SEQL

SEQL trains linear classifiers on the explicit space of all subsequences to classify biological sequences and identify discriminative subsequence features for tasks such as protein remote homology detection and remote fold recognition.

Key Features:

• High-Dimensional Predictor Space: Operates directly in the explicit space of all subsequences from the training set rather than using kernel-induced feature spaces.
• Gradient-Bounded Coordinate Descent Algorithm: Selects discriminative subsequences efficiently without expanding the entire predictor space.
• Versatile Loss Functions: Supports binomial log-likelihood loss for logistic regression and squared hinge loss for support vector machine objectives.
• Interpretability: Produces models consisting of lists of weighted discriminative subsequences that can be related to biological phenomena.
• Comparative Performance: Achieves performance comparable to kernel support vector machines while yielding interpretable subsequence-weighted models.

Scientific Applications:

• Protein Remote Homology Detection: Identifies evolutionary relationships between proteins by detecting discriminative subsequence patterns.
• Remote Fold Recognition: Predicts protein fold relationships from amino acid sequences using discriminative subsequence features.

Methodology:

SEQL operates in the explicit subsequence feature space, trains linear classifiers using a gradient-bounded coordinate-descent algorithm that selects discriminative subsequences without expanding the full predictor space, and supports binomial log-likelihood and squared hinge loss functions.