Efficient Identification of Phylogenetically Informative Alignment Sites via Sparse Learning

Identifying phylogenetically informative sites in multiple sequence alignments is critical for accurate tree reconstruction and efficient data curation in phylogenomics. Existing approaches often rely on predefined topologies or heuristic criteria, limiting their generality and interpretability. Here, we introduce a novel, topology-agnostic framework for quantifying site-wise phylogenetic information using sparse learning via Lasso (Least Absolute Shrinkage and Selection Operator) regression. By modeling site log-likelihoods as predictors of tree likelihood across a large ensemble of random topologies, our approach isolates the minimal subset of sites that meaningfully contribute to phylogenetic signal. We validate the method using both simulated and empirical mammalian datasets, demonstrating that Lasso-selected sites yield topologies nearly identical to those inferred from full alignments. Furthermore, we show that a simple entropy-based proxy (Shannon H ≥ 0.5) approximates Lasso results with high fidelity, enabling rapid site-level assessments. Compared to commonly used filtering tools, our method yields more concise alignments with equivalent or superior phylogenetic resolution. These findings establish sparse learning as a principled, scalable, and practical approach for evaluating and optimizing phylogenetic data.