Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session E

Poster M4 - SJD: Structured Joint Decomposition and Projection for Multi-omics Data

Saturday, October 26, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Guangyan Li, MS (she/her/hers)

Computational associate I
Broad Institute of Harvard and MIT
Cambridge, Massachusetts, United States

Co-Author(s)

HC

Huan Chen, PhD

Research Scientist
Meta, United States

Last Author(s)

Carlo Colantuoni, PhD

Assistant Professor of Neurology
Johns Hopkins university
Baltimore, Maryland, United States

Primary Investigator(s)

BC

Brian Caffo, PhD (he/him/his)

Professor, Department of Biostatistics, Bloomberg School of Public Health
Johns Hopkins University
Baltimore, Maryland, United States

Introduction: Despite all the rapidly generated high dimensional multi-omics data from both in vivo tissue and in vitro cellular systems across different species, biologists have been analyzing them on an individual basis. The idea of understanding them in a combined manner has not been fully developed and realized. However, it is crucial to understand different data together in many situations. For instance, a cancer biologist doing experiments on cancer cell lines does not know to what extent the same conclusions can be applied to an actual in vivo system. Biologists who are doing mouse experiments would also want to know how much a mouse organ simulates a human organ. Having this knowledge is essential for the translation and application of most research results and conclusions. To address these gaps, we have developed (1) Structured Joint Decomposition (SJD) , a robust computational tool that provides joint decomposition and projection methods for genomics datasets that have biological similarities or any matrices with shared row structure. It allows integrated dimensionality reduction and projection for multiple datasets, such as human and mouse datasets studying brain neuronal development, or cell line and in vivo tissue datasets for cancer cell viability research. Specifically, the decomposition and projection methods developed for NMF are the most useful to find out the biological similarities and differences in different datasets.

Materials and

Methods: SJD's decomposition and projection functions are developed using mechanisms such as PCA and ICA. In addition, JointNMF, ProjectNMF, and NewNMF use non-negative matrix factorization (NMF). JointNMF jointly decomposes multiple datasets with shared genes and produces one joint gene loading and several sample embeddings. ProjectNMF estimates the sample embeddings in a new dataset when given the gene loadings matrices of a joint decomposition of multiple datasets. As input, a novel dataset and the result of a conventional NMF decomposition or from a joint decomposition done in jointNMF are taken. Then, through a modification of the multiplicative update rule (2), which is used in NMF (non-negative) dimensional
reduction, a random matrix is initiated and updated using the gene loadings. Its values are updated until a small enough change in the reconstruction error of the novel dataset is reached. The output of this is a sample embedding matrix for the new data given the old gene loadings.
The NewNMF function, on the other hand, is a NMF decomposition function with added flexibility on user inputs. It allows users to run NMF with known patterns to fix any rows and columns of either or both of the gene loadings and sample embedding matrices to create an in silico experiment. For example, the newNMF function can perform decomposition if the user provides only a new dataset, and it can perform projection if a dataset and one or both of the embedding matrices are provided. The additional flexibility makes it a useful function in many scenarios.

Results, Conclusions, and Discussions: 345 For three mid-gestational neocortical development datasets of mouse macaque , and human,
a jointNMF decomposition with 7 patterns is performed. The UMAP with 4 out of the 7 patterns in the sample embedding is plotted (Figure 1). The figures show high expression of pattern 5 genes in radial glial cells and cycling progenitors, which are the early-stage progenitor cells (Figure1p5), high expression of pattern 4 in intermediate progenitor cells (Figure1p4), high expression of pattern 7 in nascent neurons (Figure1p7), and high expression of pattern 2 in maturing neurons (Figure1p2).
The gene loadings results from this joint decomposition are then used for two projectNMF projections with two new datasets - scRNA-seq in fetal human neocortical tissue6 (Figure 2A), and Spatial transcriptomics in the fetal mouse brain7 (Figure 2B). As a validation of the decomposition results, the 4 patterns in Figure 2A have similar expression patterns across cell types in corresponding patterns in the joint decomposition. Figure 2B also shows the process of neuronal development from lining the ventricle to the surface of the cerebral cortex.
The joint decomposition aims to find common patterns in all decomposed datasets. In this case, the decomposition of mouse, macaque, and human datasets finds common patterns in their shared mammalian traits. The projection then produces sample embeddings for new datasets that show shared mammal gene expression patterns. The result shows a nice development of neural stem cells to neurons in both projected datasets. In conclusion, joint matrix decomposition and projection is an effective way to define shared molecular dynamics for heterogeneous multi-omics datasets such as different species, bulk, single cell, etc.
The SJD package and these projection functions can not only be used for genomics data or even omics data. It can be used in many vastly different fields to explore similarities of datasets in matrix format. SJD was originally developed for genomics settings, but it has a huge potential to be a useful decomposition and projection tool for many different purposes.

Acknowledgements (Optional): Reference:
1. Chen, H., Liu, J., Sonthalia, S., Stein-O’Brien, G., Xiao, L., Caffo, B., & Colantuoni, C. (2022). Structured Joint Decomposition (SJD) identifies conserved molecular dynamics across collections of biologically related multi-omics data matrices. bioRxiv (Cold Spring Harbor Laboratory). https://doi.org/10.1101/2022.11.07.515489
2. Daniel D. Lee and H. Sebastian Seung. 2000. Algorithms for non-negative matrix factorization. In Proceedings of the 13th International Conference on Neural Information Processing Systems (NIPS'00). MIT Press, Cambridge, MA, USA, 535–541.
3. La Manno G, Siletti K, Furlan A, Gyllborg D, Vinsland E, Mossi Albiach A, Mattsson Langseth C, Khven I, Lederer AR, Dratva LM, Johnsson A, Nilsson M, Lönnerberg P, Linnarsson S. Molecular architecture of the developing mouse brain. Nature. 2021 Aug;596(7870):92-96. doi: 10.1038/s41586-021-03775-x. Epub 2021 Jul 28. PMID: 34321664.
4. Nicola Micali, Shaojie Ma, Mingfeng Li, Suel-Kee Kim, Xoel Mato-
Blanco, Suvimal Sindhu, Jon
I. Arellano, Tianliuyun Gao, Alvaro Duque, Gabriel Santpere, Nenad Sestan, Pasko Raki c. Molecular programs of regional specification and neural stem cell fate progression in developing macaque telencephalon.
bioRxiv 2022.10.18.512724; doi: https://doi.org/10.1101/2022.10.18.512724
5. Trevino AE, Müller F, Andersen J, Sundaram L, Kathiria A, Shcherbina A, Farh K, Chang HY, Pașca AM, Kundaje A, Pașca SP, Greenleaf WJ. Chromatin and gene-regulatory dynamics of the developing human cerebral cortex at single-cell resolution. Cell. 2021 Sep 16;184(19):5053-5069.e23. doi: 10.1016/j.cell.2021.07.039. Epub 2021 Aug 13. PMID: 34390642.
6. Polioudakis, D., de la Torre-Ubieta, L., Langerman, J., Elkins, A. G., Shi, X., Stein, J. L., Vuong, C. K., Nichterwitz, S., Gevorgian, M., Opland, C. K., Lu, D., Connell, W., Ruzzo, E. K., Lowe, J. K., Hadzic, T., Hinz, F. I., Sabri, S., Lowry, W. E., Gerstein, M. B., Plath, K., ... Geschwind, D. H. (2019). A Single-Cell Transcriptomic Atlas of Human Neocortical Development during Mid-gestation. Neuron, 103(5), 785–801.e8. https://doi.org/10.1016/j.neuron.2019.06.011
7. Chen, A., Sun, Y., Lei, Y., Li, C., Liao, S., Meng, J., Bai, Y., Liu, Z., Liang, Z., Zhu, Z., Yuan, N., Yang, H., Wu, Z., Lin, F., Wang, K., Li, M., Zhang, S., Yang, M., Fei, T., Zhuang, Z., ... Li, C. (2023). Single-cell spatial transcriptome reveals cell-type organization in the macaque cortex. Cell, 186(17), 3726–3743.e24. https://doi.org/10.1016/j.cell.2023.06.009