BioSequence2Vec: Efficient Embedding Generation for Biological Sequences

Representation learning is an important step in the machine pipeline. Given current biological sequencing data volume, explicit representation prohibitive due to dimensionality of resulting feature vectors. Kernel-based methods, e.g., SVM, are a proven efficient and useful alternative for several (ML) tasks such as sequence classification. Three challenges with kernel methods (i) computation time, (ii) memory usage (storing $$n\times n$$ matrix), (iii) matrices limited kernel-based ML (difficult generalize on non-kernel classifiers). While can be solved using approximate challenge remains typical methods. Similarly, although non-kernel-based applied by extracting principal components (kernel PCA), it may result information loss, while being computationally expensive. In this paper, we propose general-purpose approach that embodies methods’ qualities avoiding computation, memory, generalizability challenges. This involves computing low-dimensional embedding each sequence, random projections its k-mer frequency vectors, significantly reducing needed compute dot product store representation. Our proposed fast alignment-free method used input any distance (e.g., k nearest neighbors) non-distance decision tree) based classification clustering tasks. Using different forms sequences input, perform variety real-world tasks, SARS-CoV-2 lineage gene family classification, outperforming state-of-the-art predictive performance.