Integrating single-cell datasets produced by multiple omics technologies is essential for defining cellular heterogeneity. Mosaic integration, in which different datasets share only some of the measured modalities, poses major challenges, particularly regarding modality alignment and batch effect removal. Here, we present a deep probabilistic framework for the mosaic integration and knowledge transfer (MIDAS) of single-cell multimodal data. MIDAS simultaneously achieves dimensionality reduction, imputation and batch correction of mosaic data by using self-supervised modality alignment and information-theoretic latent disentanglement. We demonstrate its superiority to 19 other methods and reliability by evaluating its performance in trimodal and mosaic integration tasks. We also constructed a single-cell trimodal atlas of human peripheral blood mononuclear cells and tailored transfer learning and reciprocal reference mapping schemes to enable flexible and accurate knowledge transfer from the atlas to new data. Applications in mosaic integration, pseudotime analysis and cross-tissue knowledge transfer on bone marrow mosaic datasets demonstrate the versatility and superiority of MIDAS. MIDAS is available at https://github.com/labomics/midas.

Background: The construction of genetic maps based on molecular markers is a crucial step in rice genetic and genomic studies. Pure lines derived from multiple parents provide more abundant genetic variation than those from bi-parent populations. Two four-parent pure-line populations (4PL1 and 4PL2) and one eight-parent pure-line population (8PL) were developed from eight homozygousindicavarieties of rice by the International Rice Research Institute (IRRI). To the best of our knowledge, there have been no reports on linkage map construction and their integration in multi-parent populations of rice.Results: We constructed linkage maps for the three multi-parent populations and conducted quantitative trait locus (QTL) mapping for heading date (HD) and plant height (PH) based on the three maps by inclusive composite interval mapping (ICIM). An integrated map was built from the three individual maps and used for QTL projection and meta-analysis. QTL mapping of the three populations was also conducted based on the integrated map, and the mapping results were compared with those from meta-analysis. The three linkage maps developed for 8PL, 4PL1 and 4PL2 had 5905, 4354 and 5464 bins and were 1290.16, 1720.01 and 1560.30 cM in length, respectively. The integrated map was 3022.08 cM in length and contained 10,033 bins. Based on the three linkage maps, 3, 7 and 9 QTLs were detected for HD while 6, 9 and 10 QTLs were detected for PH in 8PL, 4PL1 and 4PL2, respectively. In contrast, 19 and 25 QTLs were identified for HD and PH by meta-analysis using the integrated map, respectively. Based on the integrated map, 5, 9, and 10 QTLs were detected for HD while 3, 10, and 12 QTLs were detected for PH in 8PL, 4PL1 and 4PL2, respectively. Eleven of these 49 QTLs coincided with those from the meta-analysis.Conclusions: In this study, we reported the first rice linkage map constructed from one eight-parent recombinant inbred line (RIL) population and the first integrated map from three multi-parent populations, which provide essential information for QTL linkage mapping, meta-analysis, and map-based cloning in rice genetics and breeding.

Multiparental advanced generation intercross (MAGIC) populations provide abundant genetic variation for use in plant genetics and breeding. In this study, we developed a method for quantitative trait locus (QTL) detection in pure-line populations derived from 8-way crosses, based on the principles of inclusive composite interval mapping (ICIM). We considered 8 parents carrying different alleles with different effects. To estimate the 8 genotypic effects, 1-locus genetic model was first built. Then, an orthogonal linear model of phenotypes against marker variables was established to explain genetic effects of the locus. The linear model was estimated by stepwise regression and finally used for phenotype adjustment and background genetic variation control in QTL mapping. Simulation studies using 3 genetic models demonstrated that the proposed method had higher detection power, lower false discovery rate (FDR), and unbiased estimation of QTL locations compared with other methods. Marginal bias was observed in the estimation of QTL effects. An 8-parental recombinant inbred line (RIL) population previously reported in cowpea and analyzed by interval mapping (IM) was reanalyzed by ICIM and genome-wide association mapping implemented in software FarmCPU. The results indicated that ICIM identified more QTLs explaining more phenotypic variation than did IM; ICIM provided more information on the detected QTL than did FarmCPU; and most QTLs identified by IM and FarmCPU were also detected by ICIM.

Recent breakthroughs in single-cell multi-omics technologies have enabled simultaneous measurement of diverse cellular modalities, offering unprecedented biological insights. However, integrating such multimodal data faces dual challenges: Small-scale paired-modality studies (hundreds of cells) risk overfitting, while large-scale reference atlases often struggle to generalize effectively to new datasets. To overcome these challenges, we present multimodal integration with self-supervised learning (MINERVA), a unified deep learning framework employing self-supervised strategies for single-cell multimodal integration. MINERVA outperforms six state-of-the-art methods in dimensionality reduction, missing feature imputation, and batch effect correction, even with limited training cells. For large-scale applications, MINERVA constructs scalable multi-tissue references that support zero-shot knowledge transfer to unseen datasets, instant cell type annotation, novel cell states identification, and comprehensive downstream analyses, all without model retraining. Uniquely bridging small-scale precision with atlas-level generalization, MINERVA serves as a versatile tool for both de novo integration and cost-effective atlas reuse in single-cell research.

Single-cell sequencing has transformed our understanding of cellular heterogeneity, enabling the construction of multi-omics atlases through data integration. However, conventional atlas updates require full reintegration of all datasets, creating scalability challenges that limit the timeliness and adaptability of biomedical research. Here we present multimodal integration with continual learning (MIRACLE), an online learning framework for scalable multimodal integration. Using dynamic architecture adaptation and data rehearsal, MIRACLE continually integrates diverse datasets while preserving biological fidelity. Across evaluations, MIRACLE achieves accurate online integration with substantially improved efficiency, refining and expanding atlases with new cross-modal, cross-tissue and cross-disease data. Applied to respiratory infections, it reveals both shared and pathogen-specific immune mechanisms in coronavirus disease 2019, influenza A and tuberculosis. Overall, MIRACLE provides an efficient and collaborative solution for the continual integration, sharing and exploration of biological knowledge.