Phylogenetic inference from large-scale multiple sequence alignments (MSAs) is essential for understanding evolutionary relationships, tracking pathogen outbreak dynamics, and quantifying biodiversity patterns across the tree of life. Traditional distance-based and likelihood-based methods struggle to scale efficiently while capturing the complex hierarchical relationships inherent in evolutionary data. Thus, low dimensional representations of MSAs are needed to infer phylogenies from large MSAs. We present a self-supervised contrastive learning framework that learns compact, phylogenetically informative representations of MSA patches without requiring explicit phylogenetic labels.
Our architecture combines a multi-scale convolutional encoder with biologically motivated data augmentations tailored to MSA characteristics. The encoder processes local alignment patches and projects them into a normalized latent space optimized via an NT-Xent loss. Crucially, we designed domain-specific augmentations that reflect biological invariances: row shuffling captures sequence order independence, column masking simulates alignment uncertainty and missing data, row subsampling models varying taxonomic sampling, and reverse complementation accounts for strand ambiguity. This augmentation strategy enforces robust feature learning while maintaining phylogenetically relevant sequence information.
We evaluated our framework assessing multiple aspects of representation quality: hierarchical clustering fidelity through cophenetic correlation, discrete evolutionary group recovery via k-nearest neighbor classification, continuous sequence similarity preservation through correlation analysis, and latent space geometry metrics including uniformity and intrinsic dimensionality. Our analysis revealed fundamental trade-offs between learning discrete phylogenetic groupings versus preserving fine-grained evolutionary distances.
This work establishes contrastive learning as a viable approach for phylogenetic representation learning, opening avenues for scalable phylogenetic placement, distance matrix approximation, and integration with downstream evolutionary analyses.

Co-authors: Denise Kühnert