LIBD Seminar: Spot Deconvolution

An LIBD (https://www.libd.org/) talk about spot deconvolution in the post-mortem human DLPFC

Transcript

1. Spot Deconvolution in the Post-Mortem Human DLPFC Nicholas J. Eagles Research Associate nick-eagles.github.io @Nick-Eagles 1 https://speakerdeck.com/nickeagles
2. About Me - B.S. Mathematics, 2018 (UMBC) - Programming background - Interest in neuroscience, psychiatry, pharmacology 2
3. My Work at LIBD - Research Associate (Data Science) in Leo’s group - Fifth year at LIBD - Develop computational pipelines for preprocessing of bulk RNA-seq data (SPEAQeasy) and whole-genome bisulfite sequencing data (BiocMAP) - Install and maintain software at JHPCE for LIBD and collaborators - Projects - VA PTSD WGBS (2597 samples) - LIBD SCZD-control PsychENCODE WGBS (664 samples) - LIBD PsychENCODE Prenatal WGBS (20 samples) - spatialDLPFC (30 samples) - 2 first-author manuscripts, 3 middle-author, 4 other preprints 3
4. SPEAQeasy (2021) Bulk RNA-seq preprocessing pipeline Raw sequencing reads -> Analysis-ready R objects (.fastq) (.Rdata) https://doi.org/10.1186/s12859-021-04142-3 http://research.libd.org/SPEAQeasy/ 4
5. TL;DR Raw sequencing reads (.fastq) R objects, ready for statistical analysis (.Rdata) https://github.com/LieberInstitute/BiocMAP BiocMAP (2022 preprint) 5
6. BiocMAP Manuscript Documentation 6 http://research.libd.org/BiocMAP/ https://doi.org/10.1101/2022.04.20.488947
7. Background and Previous Work 7
8. Background: Human DLPFC 8 slideshare.net Louise A Huuki-Myers @lahuuki
9. DOI: 10.1038/s41593-020-00787-0 twitter.com/lcolladotor/status/1233661576433061888 from 2020-02-29 Andrew E Jaffe @andrewejaffe Kristen R Maynard @kr_maynard Keri Martinowich @martinowk
10. DOI: 10.1038/s41593-020-00787-0 twitter.com/lcolladotor/status/1233661576433061888 from 2020-02-29 DOI 10.1093/bioinformatics/btac299 Since Feb 2020 spatialLIBD::fetch_data() provides access to SpatialExperiment R/Bioconductor objects Stephanie C Hicks @stephaniehicks Lukas M Weber @lmweber
11. DOI: 10.1038/s41593-020-00787-0 twitter.com/lcolladotor/status/1233661576433061888 from 2020-02-29 twitter.com/CrowellHL/status/1597579271945715717 DOI 10.1093/bioinformatics/btac299 Since Feb 2020 spatialLIBD::fetch_data() provides access to SpatialExperiment R/Bioconductor objects
12. The SpatialDLPFC Project 12
13. Spatial DLPFC Intro - Major study using the Visium platform to study paired snRNA-seq and spatial gene expression in the post-mortem DLPFC from 10 neurotypical donors (n = 30 samples) 13
14. 14 Study Design • Neurotypical Adults • High quality RNA (RIN > 7) • 30 Visium samples (10 donors x 3 positions) • 19 snRNA-seq samples (10 donors x 2 positions) • 4 Visium-SPG samples (separate donors) -SPG (SRT + IF)
15. Visium-SPG = Visium SRT + immunofluorescence (using identical tissue samples) Sang Ho Kwon @sanghokwon17 Visium Spatially Resolved Transcriptomics (Visium SRT)
16. Visium Spatial Proteogenomics (Visium-SPG) - Gene expression captured like ordinary Visium - Multi-channel fluorescent images captured of the same tissue - Channels measure proteins marking for specific cell types Kristen R. Maynard 16 Sang Ho Kwon Visium-SPG = Visium SRT + immunofluorescence (using identical tissue samples) Fluorescent Protein Cell Type TMEM119 Microglia Neun Neurons OLIG2 Oligodendrocytes GFAP Astrocytes
17. Integrating SRT and Single-Cell Data to Map Cell Types Spatially - Single-cell data lacks spatial information, and SRT data lacks cell-type-composition information single-cell spatial Questions we can’t answer from SRT data alone: How are cellular populations distributed spatially? Which cell types are communicating in ligand-receptor interactions associated with schizophrenia? Image from Bo Xia: https://twitter.com/BoXia7/status/1261464021322137600?s=12 17 spot deconvolution
18. Spot Deconvolution 18 Cell 1 Cell 2 … Cell N Gene 1 0 0 … 0 Gene 2 2 5 … 3 … … … … … Gene N 1 0 … 0 Spot 1 Spot 2 … Spot N Gene 1 1 0 … 3 Gene 2 0 1 … 0 … … … … … Gene N 4 2 … 2 Astro Excit … Inhib Spot 1 1 1 … 1 Spot 2 0 2 … 0 … … … … … Spot N 1 0 … 2 Single- Cell Spatial Deconvolved Results Spot 1
19. Spot Deconvolution: Visually 19 Br6522_Ant_IF
20. Sean Maden @MadenSean Sang Ho Kwon @sanghokwon17 #deconvochallenge doi.org/10.48550/arXiv.2305.06501
21. Benchmark existing spot deconvolution methods, and choose the best to estimate cell-type composition in our spatial data 21 My Goal
22. Existing Spot Deconvolution Software - Explored 3 novel software methods Software name Overall approach Input Cell Counts Output Tangram (Biancalani et al.) Mapping individual cells Every spot Integer counts Cell2location (Kleshchevnikov et al.) Matching gene-expression profile Average across spots Decimal counts SPOTlight (Elosua-Bayes et al.) Matching gene-expression profile Not used Proportions 22 Excit L5 Counts PCP4 Counts
23. Benchmarking Spot Deconvolution Software: Theory - How do we measure performance or accuracy of cell-type predictions? - Make orthogonal measurements*: image-derived counts - Leverage prior knowledge: neurons localize to gray matter - Self-consistency of results: broad vs. fine cell-type results 23 Stephanie C. Hicks
24. Visium-SPG IF Images as an Orthogonal Measurement 24
25. Using IF to Quantify Cell Types - Visium-SPG IF images mark for several proteins - Fluorescence in image channels correlates with counts of measured cell types Can measure 5 distinct cell types: 25 • Astrocyte (GFAP) • Neuron (NeuN) • Oligodendrocyte (OLIG2) • Microglia (TMEM119) • Other (low signal in all channels) samuibrowser.com Chaichontat Sriworarat Stephanie C. Hicks doi.org/10.1101/2023.01.28.525943
26. Segmenting Cells on Visium-SPG IF Images 1. Segment cells on IF image 2. Manually label example cells 3. Train cell-type classifier and apply on remaining data 26
27. Constructing Dataset of Labeled Cells 1. Segment cells on IF image 2. Manually label example cells 3. Train cell-type classifier and apply on remaining data Image Channels Classified Cell Type Cell Mask 27 Annie B. Nguyen
28. Constructing Dataset of Labeled Cells 28 1. Segment cells on IF image 2. Manually label example cells 3. Train cell-type classifier and apply on remaining data 4 sections * 5 cell types * 30 cells = 600 manually labeled cells Annie B. Nguyen
29. Addressing Bias in Cell Selection - Trained logistic regression model on 600-cell dataset - Broke cells into 4 quartiles based on model confidence - Labelled 320 more cells, evenly sampled from all 4 quartiles 29 Cell Type Probability Astro 0.2 Oligo 0.3 Micro 0.1 Neuron 0.45 Other 0.05 4 quartiles * 4 sections * 5 cell types * 4 cells = 320 new cells 600 old cells + 320 new cells = 920 total cells Cell Type Probability Astro 0.01 Oligo 0.02 Micro 0.01 Neuron 0.93 Other 0.03 Less-confident neuron More-confident neuron Confidence = 0.45 Confidence = 0.93
30. Training Cell-Type Classifier 30 Model Test Precision Test Recall Decision tree 0.86 0.87 Logistic regression 0.91 0.90 Support vector machine 0.90 0.90 Dataset # Training # Test Split Old 600 480 120 80/20 New 320 240 80 75/25 Combined 920 720 200 ~78/22 1. Segment cells on IF image 2. Manually label N cells 3. Train cell-type classifier and apply on remaining data Grid search with 5-fold CV for each model to select hyperparameters Data Model Final model chosen
31. Benchmark Results 31
32. Benchmark Results: Make Orthogonal Measurements 32 Neuron Layer Broad Decision Tree - Sum across finer cell types to compare against broader - Drop EndoMural for comparison to decision tree Software Predictions
33. Benchmark Results: Make Orthogonal Measurements 33 Micro (Br6522_Ant_IF) All points in A -> one point in B
34. Benchmark Results: Make Orthogonal Measurements 34 Broad Cell Types Layer-Level Cell Types
35. Benchmark Results: Leverage Prior Knowledge - Manually annotate spots with histological layer - Explore how cell-type predictions map to annotated layers Kristen R. Maynard Br6522_Ant_IF 35
36. 36 Max across layers Not max Benchmark Results: Leverage Prior Knowledge
37. Benchmark Results: Leverage Prior Knowledge 37
38. Benchmark Results: Self-Consistency of Results 38
39. Benchmark Summary 39 Metric Tangram Cell2location SPOTlight Metric Type Avg. cor (spot-level) 0.31 0.30 0.21 Orthogonal measurements Avg. RMSE (spot-level) 1.35 1.24 1.3 Orthogonal measurements Overall prop.: (KL Div.) 0.44 0.49 0.41 Orthogonal measurements Overall prop.: (cor.) 0.46 0.37 0.47 Orthogonal measurements Overall prop.: (RMSE) 3020 3890 3040 Orthogonal measurements Histological mapping 0.69 0.77 0.23 Leverage known biology Broad vs. layer (cor.) 1.00 0.77 -0.36 Self-consistency of results Broad vs. layer (RMSE) 102 4200 4220 Self-consistency of results
40. View Spot Deconvolution Results: Public Resources 40
41. How Spot Deconvolution Results Were Used A. Better characterize unsupervised spatial domains B. Cell-cell communication; cell-type-informed ligand-receptor interactions in the context of schizophrenia risk A 41 Boyi Guo Melissa Grant-Peters
42. Conclusions - Imaging data can be leveraged to infer cell-type composition in Visium/spatial data - Tangram and Cell2location perform better than SPOTlight, with each scoring best in different metrics - Tangram matches overall snRNA-seq cell-type proportions - Cell2location slightly more accurately maps cell types to expected layers 42 - Existing spot deconvolution algorithms have limited accuracy - Incorporating other data types might improve cell-type predictions - RNAscope? Future Directions Tangram Cell2location SPOTlight
43. Get In Touch With Me Knowledge - Bulk RNA-seq and WGBS data processing - Installing and running GPU-based software - Nextflow, computational workflows Future Interests - Spatial and stitching Visium capture areas - Machine learning for genomics - Image processing 43 https://calendly.com/nick-eagles/25-minute-data-science-guidance-session DSGS By Week (2022) GitHub Contributions (~1 Year)
44. Acknowledgements LIBD Annie B. Nguyen Leonardo Collado-Torres Kristen R. Maynard Louise Huuki-Myers Abby Spangler Kelsey D. Montgomery Sang Ho Kwon Heena R. Divecha Madhavi Tippani Matthew N. Tran Arta Seyedian Thomas M. Hyde Joel E. Kleinman Stephanie C. Page Keri Martinowich JHU Biostatistics Chaichontat Swirorarat Stephanie C. Hicks Boyi Guo JHU Biomedical Engineering Alexis Battle Prashanthi Ravichandran PsychENCODE consortium University College London Genetics and Genomic Medicine Mina Ryten Melissa Grant-Peters nick-eagles.github.io @Nick-Eagles Feel free to reach out! 44
45. Counting “Correct” Layer Matches 45 Cell Type Expected Layer Astro L1 EndoMural L1 Excit L2-6 Excit_L2_3 L2 or L3 Excit_L3 L3 Excit_L4 L4 Excit_L5 L5 Excit_L5_6 L5 or L6 Excit_L6 L6 Inhib L2-6 Micro L1 or WM Oligo WM OPC L1 or WM