The integration of data from multiple modalities generated by single-cell omics technologies is crucial for accurately identifying cell states. One challenge in comprehending multi-omics data resides in mosaic integration, in which different data modalities are profiled in different subsets of cells, as it requires simultaneous batch effect removal and modality alignment. Here, we develop Multi-omics Mosaic Auto-scaling Attention Variational Inference (mmAAVI), a scalable deep generative model for single-cell mosaic integration. Leveraging auto-scaling self-attention mechanisms, mmAAVI can map arbitrary combinations of omics to the common embedding space. If existing well-annotated cell states, the model can perform semisupervised learning to utilize existing these annotations. We validated the performance of mmAAVI and five other commonly used methods on four benchmark datasets, which vary in cell numbers, omics types, and missing patterns. mmAAVI consistently demonstrated its superiority. We also validated mmAAVI's ability for cell state knowledge transfer, achieving balanced accuracies of 0.82 and 0.97 with less 1% labeled cells between batches with completely different omics. The full package is available at https://github.com/luyiyun/mmAAVI.

Streptococcus pneumoniae is the most common human respiratory pathogen, and β-lactam antibiotics have been employed to treat infections caused by S. pneumoniae for decades. β-lactam resistance is steadily increasing in pneumococci and is mainly associated with the alteration in penicillin-binding proteins (PBPs) that reduce binding affinity of antibiotics to PBPs. However, the high variability of PBPs in clinical isolates and their mosaic gene structure hamper the predication of resistance level according to the PBP gene sequences. In this study, we developed a systematic strategy for applying supervised machine learning to predict S. pneumoniae antimicrobial susceptibility to β-lactam antibiotics. We combined published PBP sequences with minimum inhibitory concentration (MIC) values as labelled data and the sequences from NCBI database without MIC values as unlabelled data to develop an approach, using only a fragment from pbp2x (750 bp) and a fragment from pbp2b (750 bp) to predicate the cefuroxime and amoxicillin resistance. We further validated the performance of the supervised learning model by constructing mutants containing the randomly selected pbps and testing more clinical strains isolated from Chinese hospital. In addition, we established the association between resistance phenotypes and serotypes and sequence type of S. pneumoniae using our approach, which facilitate the understanding of the worldwide epidemiology of S. pneumonia.