Spatial integration of multi-omics data from serial sections using the novel Multi-Omics Imaging Integration Toolset

Abstract

Background: Truly understanding the cancer biology of heterogeneous tumors in precision medicine requires capturing the complexities of multiple omics levels and the spatial heterogeneity of cancer tissue. Techniques like mass spectrometry imaging (MSI) and spatial transcriptomics (ST) achieve this by spatially detecting metabolites and RNA but are often applied to serial sections. To fully leverage the advantage of such multi-omics data, the individual measurements need to be integrated into 1 dataset.

Results: We present the Multi-Omics Imaging Integration Toolset (MIIT), a Python framework for integrating spatially resolved multi-omics data. A key component of MIIT's integration is the registration of serial sections for which we developed a nonrigid registration algorithm, GreedyFHist. We validated GreedyFHist on 244 images from fresh-frozen serial sections, achieving state-of-the-art performance. As a proof of concept, we used MIIT to integrate ST and MSI data from prostate tissue samples and assessed the correlation of a gene signature for citrate-spermine secretion derived from ST with metabolic measurements from MSI.

Conclusion: MIIT is a highly accurate, customizable, open-source framework for integrating spatial omics technologies performed on different serial sections.

Keywords: image registration; mass spectrometry imaging; spatial transcriptomics.

Figure 1:

Overview of the ProstOmics spatial multi-omics dataset on prostate tissues. Sections of each core are numbered and processed in the same order. The sections used in this study are at positions 1–3 and 6–11 with a distance between 2 adjacent sections of 10–20 µm. The following staining techniques are used: hematoxylin, erythrosine, and saffron (HES); hematoxylin and eosin (HE); Mason’s trichrome staining (MTS); and immunohistochemistry (IHC). Spatial transcriptomics is applied to section 2, MSI in positive ion mode on section 6, and MSI in negative ion mode on section 7. Histopathology was evaluated for sections 2, 6, and 7.

Figure 2:

Overview of GreedyFHist’s registration. (A) Registration between 2 images. First, images are preprocessed for affine registration (red arrows). In step 1, images are segmented from background to focus on tissue region. Then, denoising (step 2) is applied to remove noise while retaining major histological features. Grayscale conversion and downscaling (step 3) are used on the images to improve registration time. Next, affine registration is performed using Greedy (step 4). Then images are preprocessed for nonrigid registration (step 1 and step 3; blue arrows). The affine transformation matrix and preprocessed images are passed to Greedy to compute nonrigid transformation matrices (step 5). Transformation matrices are then rescaled to the original image’s resolution, composited into one transformation matrix (step 6), and applied to the moving image (step 7). (B) Groupwise registration. When registering a series of stained images, we denote one image as the fixed image and every other image in the series as moving images. First, an affine registration between each neighboring image pair is computed. Then a transformation sequence is applied to each moving image to affinely register it to the fixed image (step 1). Finally, a nonrigid registration is performed between each affinely registered image and the fixed image (step 2).

Figure 3:

Integration workflow of MIIT. (A) Preprocessing. Preprocessing of ST and MSI data. File formats are processed to reference matrices and registered to stained images, if necessary. Each number in a reference matrix is either a reference to on-tissue molecular data or 0, which denotes background. In this example, ST contains 2,077 different spot references and MSI contains 7,527 different pixel references to molecular data points. Different references are highlighted in different colors. (B) Registration. Then the ST-section is registered to the MSI-section based on stained images. (C) Fusion. (1) Reference matrix of ST is used to group MSI-data within the same spot regions and (2) grouped MSI-data are aggregated within each spot, resulting in MSI-spots. If additional annotations are provided, integrated spots can be matched against these annotations as well. (D) Export. Lastly, integrated spots are exported into the relevant file formats.

Figure 4:

Assessing registration accuracy using landmarks. Comparison of error distribution in log₁₀-scale for (A) registration of adjacent sections between GreedyFHist and HistoReg and (B) registration of distant sections between pairwise registration mode and groupwise registration mode using GreedyFHist. Median-TRE is shown at a log scale. Representative registration examples showing (C) an accurate registration (median TRE = 9.970 µm) and (D) an inaccurate registration due to tissue damage (median TRE = 593.006 µm). Landmarks of moving and warped landmarks are plotted in blue, landmarks of fixed images are green, and the distance between warped and fixed landmarks for warped images is illustrated in a red dashed line.

Figure 5:

Comparing gene signature and metabolites between gland and stroma spots. Gene scores and metabolite levels in stroma and gland spots for (A) GSCS, (B) citrate, (C) zinc, and (D) spermine. (E) Citrate, (F) zinc, and (G) spermine levels plotted against CSGS score for one sample (P28_03) for integrated spots colored according to tissue type. Linear regression lines, Spearman correlation coefficient , and P value are shown. (H) Sample-wise correlation coefficients between CSGS and citrate, zinc, and spermine. (*) denotes significance (<0.05). (I) Percentage of successfully integrated spots after tissue type matching across different section distances. (J) Distribution of gland and stroma spots in ST-section and MSI-section without tissue type matching and after tissue type matching. Spots that could not be assigned to either stroma or gland or had a different histopathology classification were discarded beforehand.

Figure 6:

Comparison of different integrated spatial multi-omics datasets. (A) Spot-wise distribution of gland and stroma, CSGS, citrate, zinc, and spermine for 4 different datasets for sample P28_3. (B–D) Spearman correlation coefficient distribution for each sample for all 4 datasets. * denotes significant correlations (P < 0.001). (E) Average for each metabolite compared to the CSGS score. Error bars represent standard deviation.