The Mathematical Frameworks for Integrative Analysis of Emerging Biological Data Workshop demonstrated the power of hackathons to both inform and develop new analysis methods to capture the complex, multi-scale nature of biological datasets from high-throughput data modalities. Notably, the hackathon studies of the workshop were specifically designed to span state-of-the-art multi-omics challenges to map the epigenetic, molecular, and cellular interaction across time and sample populations. Single-cell measurements spanning molecular modalities can inherently simplify the challenge of linking disparate biological scales, but layering new sets of molecular measurements increases the complexity of the analyses to interpret these data. The computational needs hinge on the underlying biological question being asked as well as the characteristics of the data themselves. In our workshop, different modelling considerations had to be made for multi-modal integration, as higlighted in the seqFISH and scNMT-seq challenges (matching on the same genes, or cells) and the scProteomics challenge (partially unmatched measurements). Regardless, through these hackathons we identified several common analysis themes spanning algorithmic advances, interpretation, benchmarking, and software infrastructure necessary for biological interpretation. All hackathons required methods for dealing with data quality, data loss from summarization, timing variances between and within omics layers, and batch effects. These represent the necessary challenges to overcome in the coming years, along with efficient and insightful data visualization strategies to infer regulatory relationships between different omics.
Technologies to profile biological systems at single-cell resolution and across molecular scales are advancing at an unprecedented pace. Analytically, these advances require the computational community to pursue research that can first enable robust analyses tailored to a specific biology or measurement technology, and second, that can scale and adapt to these rapid advances. Our hackathons highlighted current technologies for spatial molecular profiling. The two technologies used in this study both have limited molecular resolution. Therefore, multi-platform data combining the spatial molecular data from either seqFISH, MIBI, or imaging mass cytometry require complementary data from other single-cell technologies to provide both high spatial and molecular resolution enabled through data integration. We note that additional technologies, such as slide-seq [@doi:10.1126/science.aaw1219] and Visium from 10X Genomics produce spatially resolved molecular measurements approaching measurements of the whole transcriptome, but lack the fine spatial resolution of these alternative imaging technologies. As such, emerging technologies still require further multi-platform data integration for comprehensive analysis. The scNMT-seq challenge did not include spatially resolved data but highlighted the potential of further inference of gene regulation through concurrent profiling of RNA, methylation, and chromatin state. Technological advances for multi-omics spatial data and epigenetics data are rapidly advancing and becoming increasingly available through Nanostring, 10X Genomics, Akoya Biosciences, and others. Our workshop keynote Bernd Bodenmiller presented new research-level technological advances that enable three-dimensional spatial molecular profiling [@doi:10.1101/2020.05.24.113571]. Other technologies are currently expanding to allow for temporally resolved profiling [@doi:10.1038/s41592-020-0880-2]. Integration strategies aware of these future directions and the mathematical challenges that span technologies will be most adept at advancing biological knowledge: this was the primary aim of this workshop.
The implementation of novel analysis tools requires further robust software ecosystems, including Bioconductor [@doi:10.1038/s41592-019-0654-x], Biopython, and toolkits such as Scanpy [@doi:10.1186/s13059-017-1382-0], Seurat [@doi:10.1038/nbt.4096], or Giotto [@doi:10.1101/701680], in which users can create their analysis approaches and while anticipating stable and adaptive data structures robust for these emerging technologies. The size of these emerging datasets, particularly in the context of their application to atlas projects (e.g. the Human Tumor Atlas Network [@doi:10.1016/j.cell.2020.03.053], Human Cell Atlas [@doi:10.7554/eLife.27041, @doi:10.1038/550451a], Allen Brain Initiative, Brain Initiative Cell Census Network, or ENCODE, to cite a few) are key examples that computational efficiency and scalability of these implementations are becoming ever more critical.
In addition to new technologies, we wish to emphasize that arising multi-omics analysis methods can support the generation of new data sources to resolve the multi-scale nature of biological systems. For example, while the workshop posed the scNMT-seq data and spatial molecular datasets as distinct challenges for data integration, integration of matched datasets between these spatial and epigenetic profiling techniques could further resolve the dependence of cell-type and cellular-interactions of regulatory networks. By embedding prior biological knowledge as rules in the analysis approaches, additional sources of data can generate a new representation of a biological system. For example, curated regulatory networks from databases such as KEGG, Biocarta, GO, TRANSFAC, or MSigDB provide commonly used frameworks for this prior knowledge. These gene regulatory networks must be extended to map the impact of cellular context on transcriptional regulation that are being uncovered by emerging single-cell atlases. The regulatory networks and dynamic features captured in single-cell data also provide the potential for future techniques to predict molecular and cellular states. Our hackathons and workshop have shown that merging single-cell data with mathematical models have the potential to predict behaviors in biological systems using rules derived from only prior biological knowledge.