NBT Dai Qionghai and others developed an artificial intelligence method for fusion analysis of cell microscopic images and genetic data—MUSE

Editor-in-Chief | Xi Since the 17th century Dutch scientist Leeuwenhoek used microscopes to observe biological cells, the difference in cell morphology has been an important basis for studying the internal structure of organs and tissues, and is the mechanism characteristic of organs.
Basic means of parsing
.

In recent years, with the development of single-cell genomics technology, it has become a new research method to reveal the heterogeneity of tissues and organs through the difference of gene expression, which has provided many major research projects (such as human cell atlas, human biomolecular atlas).
technical support
.

These two methods provide tissue analysis methods from different dimensions.
If they are combined, can they break through the observational limitations of a single method and achieve a higher degree of organ- and disease-specific resolution (Figure 1)? Figure 1.
Advantages of multimodal fusion: Small versus large cell types can be distinguished based on morphology (x-axis); high, medium, and low-expressing cell types can be distinguished based on gene expression levels (y-axis); The fusion of the two modalities can further fully resolve all cell types in the tissue
.

Based on the research group's accumulation in optical microscopy (Cell 2021, Nature Photonics 2019, 2020), single-cell gene analysis technology (Nature Methods 2019) and cross-application of artificial intelligence (Nature Methods 2021, Nature Machine Intelligence 2019), 2022 On March 28, Lani F.
Wu, Steven J.
Altschuler of UCSF and the team of Academician Dai Qionghai of Tsinghua University (co-authored as Bao Feng and Yue Deng) published an article in Nature Biotechnology Integrative spatial analysis of cell morphologies and transcriptional states with MUSE, proposes an artificial intelligence method for the fusion analysis of cell microscopic images and gene data.
In the study of various biological problems including the brain and Alzheimer's disease, the fusion of images and genes can greatly improve our understanding of the spatial structure of complex organs.
Cognitive competencies related to disease development parsing
.

Image and gene are two completely different modalities, their information presentation rules are different, and the analysis methods are also very different
.

Therefore, when considering both, it is necessary to accurately identify the key information of each modality related to tissue structure and disease characteristics, balancing their respective contributions to tissue specificity, while avoiding the contamination of the modality information for the other modality.
damage to information
.

In response to this problem, the paper proposes a multi-modal structural embedding (MUSE) representation learning method, which realizes the effective fusion of two modal information through three steps
.

1) Single-modal feature learning: the input original features x, y are transformed into latent space representations hx, hy respectively; 2) Single-modal label learning: The single-modal representations hx and hy are clustered to obtain each modality.
3) Fusion feature learning: the single-modal features hx and hy are fused and transformed
.

The entire framework is optimized under the constraints of two learning objectives of self-supervision and self-reconfiguration
.

Self-reconstruction ensures that the fused feature representation (z) preserves most of the information for each modality; self-supervision ensures that differences between cell types in a single modality are preserved in the fused features
.

Using the proposed method, experiments are performed on real data generated by a variety of techniques
.

seqFISH is a micro-transcriptional sequencing technology proposed by Professor Long Cai of Caltech
.

Due to the requirement of multiple rounds of imaging, the sequencing depth is generally shallow, but fluorescence microscopy images of cells can be additionally provided
.

Based on data collected from mouse cerebral cortex by seqFISH+microspatial transcriptome (Nature, 2019), the paper tested the ability of cerebral cortex structure analysis under the condition of limited number of genes (n=500)
.

In contrast, existing genetic analysis methods identified three cortical regions; in contrast, the fusion learning method of fluorescent images and genes accurately identified all four cortical regions and provided finer details within each region.
Cell type resolution (Figure 2)
.

The analysis revealed that fluorescence microscopy images of cells can provide additional, complementary information that enhances detection of complex tissue structures
.

Figure 2.
Mouse cerebral cortex data collected based on seqFISH+ technology (Nature, 2019), cortical cell type results obtained by gene single-modality analysis (top) and image-gene fusion analysis (bottom) Spatial Transcriptomics Different spatial locations of tissues can be labeled and sequenced, which has a wide range of applications in biomedical research
.

However, the sequencing result of each region is the mixed gene expression of multiple cells in the region, resulting in a relative decline in the ability to identify cell types within the tissue
.

The article further analyzes the spatial transcriptional sequencing data of pancreatic ductal carcinoma and the corresponding H&E stained sections (Nature Biotechnology 2020)
.

Analysis of genetic data alone can simply divide the entire tissue into four regions, including one cancer tissue region (Figure 3)
.

However, by analyzing gene copy variation numbers in patient tissue, it was discovered that cancer is actually caused by two cellular variants
.

In contrast, using the analysis results of the images fused by the paper method, two different cancer tissue regions can be identified in the tissue section (Fig.
3), which provides a further in-depth exploration of the interaction between cancer heterogeneity and the spatial microenvironment.
method
.

Figure 4.
Spatial distribution of cell types obtained by single-modality analysis and combined analysis of pancreatic ductal carcinoma Alzheimer's disease is a disease that has a significant impact on the cognitive ability of the elderly, and its causes are complex
.

In general, a clear sign of disease is a marked increase in Abeta polypeptides in the patient's brain
.

Exploring how Abeta affects the normal function of neighboring nerve cells has important implications for increasing the biological understanding of disease development
.

The article further analyzes mouse Alzheimer's disease brain data containing multiple disease development time points (Cell 2020)
.

Each mouse brain contained two consecutive sections, one for spatial transcriptome sequencing and one for Abeta's location and fluorescence imaging
.

Through the joint embedding method proposed in the article, each spatial sequencing position and the corresponding Abeta fluorescence imaging region were jointly analyzed
.

The method simultaneously identified the temporal development trajectories and spatial distribution differences of the disease
.

The article further carried out analysis and verification on the data of colon tissue sections
.

The method proposed in this paper can also be extended to more biological modal analysis, and to perform more than two modal data analysis, providing a basic tool for the understanding of complex biological processes
.

Link to the original text: https:// Publisher: 11 Reprint Notice [Non-original article] The copyright of this article belongs to the author of the article, and personal sharing is welcome, and it is prohibited without permission For reprinting, the author has all legal rights, and offenders will be held accountable
.