SUMMARY: We introduce DNABERT-S, a tailored genome model that develops species-aware embeddings to naturally cluster and segregate DNA sequences of different species in the embedding space. Differentiating species from genomic sequences (i.e. DNA and RNA) is vital yet challenging, since many real-world species remain uncharacterized, lacking known genomes for reference. Embedding-based methods are therefore used to differentiate species in an unsupervised manner. DNABERT-S builds upon a pre-trained genome foundation model named DNABERT-2. To encourage effective embeddings to error-prone long-read DNA sequences, we introduce Manifold Instance Mixup (MI-Mix), a contrastive objective that mixes the hidden representations of DNA sequences at randomly selected layers and trains the model to recognize and differentiate these mixed proportions at the output layer. We further enhance it with the proposed Curriculum Contrastive Learning (C2LR) strategy. Empirical results on 28 diverse datasets show DNABERT-S's effectiveness, especially in realistic label-scarce scenarios. For example, it identifies twice more species from a mixture of unlabeled genomic sequences, doubles the Adjusted Rand Index (ARI) in species clustering, and outperforms the top baseline's performance in 10-shot species classification with just a 2-shot training. AVAILABILITY AND IMPLEMENTATION: Model, codes, and data are publically available at https://github.com/MAGICS-LAB/DNABERT_S.

MOTIVATION: Whole genome shotgun based next-generation transcriptomics and metagenomics studies often generate 100-1000 GB sequence data derived from tens of thousands of different genes or microbial species. Assembly of these data sets requires tradeoffs between scalability and accuracy. Current assembly methods optimized for scalability often sacrifice accuracy and vice versa. An ideal solution would both scale and produce optimal accuracy for individual genes or genomes. RESULTS: Here we describe an Apache Spark-based scalable sequence clustering application, SparkReadClust (SpaRC), that partitions reads based on their molecule of origin to enable downstream assembly optimization. SpaRC produces high clustering performance on transcriptomes and metagenomes from both short and long read sequencing technologies. It achieves near-linear scalability with input data size and number of compute nodes. SpaRC can run on both cloud computing and HPC environments without modification while delivering similar performance. Our results demonstrate that SpaRC provides a scalable solution for clustering billions of reads from next-generation sequencing experiments, and Apache Spark represents a cost-effective solution with rapid development/deployment cycles for similar large-scale sequence data analysis problems. AVAILABILITY AND IMPLEMENTATION: https://bitbucket.org/berkeleylab/jgi-sparc.