Upgrade to Pro — share decks privately, control downloads, hide ads and more …



"Lessons from working on the edge of human brain transcriptomics with spatially-resolved transcriptomics and deconvolution" seminar on 2023-05-23 at The Francis Crick Institute


More Decks by Leonardo Collado-Torres

Other Decks in Science


  1. @lcolladotor lcolladotor.github.io lcolladotor.github.io/bioc_team_ds Lessons from working on the edge of

    human brain transcriptomics with spatially-resolved transcriptomics and deconvolution Leonardo Collado Torres, Investigator The Francis Crick Institute May 23 2023 Slides available at speakerdeck.com/lcolladotor
  2. Zoom in: snRNA-seq → deconvolution of bulk RNA-seq Matthew N

    Tran @mattntran Kristen R Maynard @kr_maynard Louise A Huuki-Myers @lahuuki Keri Martinowich @martinowk Stephanie C Hicks @stephaniehicks
  3. What is Deconvolution? • Inferring the composition of different cell

    types in a bulk RNA-seq data Louise A Huuki-Myers @lahuuki
  4. Reference Single Cell Data 9 deconvolution(Y, Z) = Proportion of

    Cell Types Louise A Huuki-Myers @lahuuki
  5. 10x snRNA-seq Reference Data Tran, Maynard et al., Neuron, 2021

    AMY DLPFC HPC NAc sACC Astro 1638 782 1170 1099 907 Endo 31 0 0 0 0 Macro 0 10 0 22 0 Micro 1168 388 1126 492 784 Mural 39 18 43 0 0 Oligo 6080 5455 5912 6134 4584 OPC 1459 572 838 669 911 Tcell 31 9 26 0 0 Excit 443 2388 623 0 4163 Inhib 3117 1580 366 11476 3974 @mattntran Matthew N Tran
  6. 1vAll Markers vs. Mean Ratio Markers 12 Louise A Huuki-Myers

    @lahuuki research.libd.org/DeconvoBuddies/
  7. 1vAll Markers vs. Mean Ratio Markers 13 Louise A Huuki-Myers

    @lahuuki research.libd.org/DeconvoBuddies/
  8. Method Summary Method Regression Correction for Technical Variation Other Features

    MuSiC Wang et al, Nature Communications, 2019 W-NNLS regression (Weighted - Non-negative least squares) None Tree guided deconvolution, good for closely related cell types Bisque Jew et al, Nature Communications, 2020 NNLS regresion Gene specific transformation of bulk data Leverage overlapping bulk & sc data SCDC Dong et al, Briefings in Bioinformatics, 2020 W-NNLS framework proposed by MuSiC Option for Gene specific transformation of bulk data (from Bisque) Multiple reference datasets can be used, results combined with ENSEMBL weights DWLS Tsoucas, Nature Communications, 2019 Dampened Weighted least squares None 15
  9. Method Regression Method Run Time Marker Evaluation Adjust for snRNA-seq

    vs. Bulk Tissues Tested Consider Cell Size Reference Set MuSiC W-NNLS Min. Internal Weighting No Pancreatic Islet, Rat & Mouse Kidney Yes Bisque NNLS Min. No Yes Adipose, DLPFC Recommend 3+ donors SCDC W-NNLS Min. Internal Weighting Yes Pancreatic Islet, mouse mammary Can input multiple references DWLS DWLS Hours Internal Selection No Mouse kidney, lung, liver, small intestine 16
  10. Which Method is the Most Accurate? • Benchmarking shows that

    different methods perform best on different data sets (Cobos et al, Nature Communications, 2020) • Benchmarking results from different papers on “real” data ◦ MuSiC paper: MuSiC > NNLS > BSEQ-sx > CIBERSORT ▪ Pancreatic Islet: Beta cells vs. HbA1c (Fig 2a) ◦ Bisque paper: Bisque > MuSiC > CIBERSORT ▪ DLPFC: Microglia vs. Braak stage, Neuron vs. Cognitive diagnostic category (Fig 4) ◦ SCDC paper: SCDC > MuSiC > Bisque > DWLS > CIBERSORT ▪ Pancreatic Islet: Beta cells vs. HbA1c (Fig 4b) ◦ Cobos benchmark: DWLS > MuSiC > Bisque > deconvoSeq ▪ Human PMBC flow sorted (Fig 7) 17 Louise A Huuki-Myers @lahuuki
  11. Mean Proportions By Region: Tran et al, bioRxiv, 2020 (5

    donors, 6 cell types) Louise A Huuki-Myers @lahuuki
  12. Peric = Mural + Endo Mean Proportions By Region: Tran

    et al, Neuron, 2021 (8 donors, 10 cell types) Louise A Huuki-Myers @lahuuki
  13. • Run with set of 20 & 25 marker genes

    per cell type • Bisque is more robust to changes in the marker set than MuSiC Method Sensitivity to Marker Set 25 vs. 20 Genes Louise A Huuki-Myers @lahuuki
  14. #deconvochallenge Challenges and opportunities to computationally deconvolve heterogeneous tissue with

    varying cell sizes using single cell RNA-sequencing datasets doi.org/10.48550/arXiv.2305.06501 Sean Maden @MadenSean
  15. Motivation • Improve Deconvolution algorithms by considering differences in size

    and RNA content between cell types • Use smFISH with RNAscope to establish data set of: ◦ Cellular composition ◦ Nuclei sizes of major cell types ◦ Average nuclei RNA content of major cell types How do we measure total RNA content of a cell if we can only observe a few genes at a time? Use a TREG Data-driven Identification of Total RNA Expression Genes (TREGs) for Estimation of RNA Abundance in Heterogeneous Cell Types research.libd.org/TREG/ doi.org/10.1101/2022.04.28.489923 Louise A Huuki-Myers @lahuuki #TREG
  16. What is a TREG? • Total RNA Expression Gene •

    Expression is proportional to the overall RNA expression in a nucleus • In smFISH the count of TREG puncta in a nucleus can estimate the RNA content Data-driven Identification of Total RNA Expression Genes (TREGs) for Estimation of RNA Abundance in Heterogeneous Cell Types research.libd.org/TREG/ doi.org/10.1101/2022.04.28.489923 #TREG
  17. Validate with RNAscope • DLPFC from control, sectioned at 10μm

    • 3 slides with 3 sections each ◦ TREG candidate + cell type marker genes • Images analyzed with HALO TREG Gene AKT3 ARID1B MALAT1/ POLR2A Cell Type Markers GAD1, SLC17A7, MBP GAD1, SLC17A7, MBP SLC17A7, MBP Cell Type Marker Excit SLC17A7 Inhib GAD1 Oligo MBP Kelsey D Montgomery Sang Ho Kwon research.libd.org/TREG/ doi.org/10.1101/2022.04.28.489923
  18. Patterns of Observed Puncta • TREGs were expressed in most

    cells • AKT3 tracks really well with pattern of expression seen in snRNA-seq (ARID1B is also pretty good) snRNA-seq RNAscope Gene Mean Prop. Cells with Expression Prop. non-zero in DLPFC snRNA Standardized β (95% CI) AKT3 0.948 0.92 -1.38 (-1.39,-1.37) ARID1B 0.908 0.94 -0.62 (-0.62,-0.61) MALAT1 0.910 1.00 -0.11 (-0.12,-0.11) POLR2A 0.853 0.30 -0.98 (-0.99,-0.98) snRNA-seq NA NA -1.33 (-1.35,-1.31) Remember: MALAT1’s puncta data is unreliable research.libd.org/TREG/ doi.org/10.1101/2022.04.28.489923
  19. #deconvochallenge Challenges and opportunities to computationally deconvolve heterogeneous tissue with

    varying cell sizes using single cell RNA-sequencing datasets doi.org/10.48550/arXiv.2305.06501 Sean Maden @MadenSean
  20. Zoom in: spatial omics Kristen R Maynard @kr_maynard Keri Martinowich

    @martinowk Stephanie C Hicks @stephaniehicks Andrew E Jaffe @andrewejaffe Stephanie C Page @CerceoPage
  21. Visium Platform for Spatial Gene Expression Image from 10x Genomics

    - A slide contains 4 capture areas, each full of thousands of 55um-wide “spots” (often containing 1-10 cells) - Unique barcodes in each spot bind to particular genes; after sequencing, gene expression can be tied back to exact spots, forming a spatial map Kristen R. Maynard 33
  22. bioconductor.org/packages/spatialLIBD Pardo et al, 2022 DOI 10.1186/s12864-022-08601-w Maynard, Collado-Torres, 2021

    DOI 10.1038/s41593-020-00787-0 Brenda Pardo Abby Spangler @PardoBree @abspangler Louise A. Huuki-Myers @lahuuki
  23. 2 pairs spatial adjacent replicates x subject = 12 sections

    35 Subject 1 Subject 2 Subject 3 Adjacent spatial replicates (0μm) Adjacent spatial replicates (300μm) PCP4 Maynard, Collado-Torres, et al, Nat Neuro, 2021
  24. 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
  25. BayesSpace clustering with batch correction worked best for multiple samples

    41 doi.org/10.1101/2023.02.15.528722 twitter.com/CrowellHL/status/1597579271945715717
  26. Different Resolutions of BayesSpace Clustering k = number of clusters

    • k=2: separate white vs. grey matter • k=9: best reiterated histological layers • k=16: data-driven optimal k based on fast H+ statistic 42 More Clusters = More Complexity doi.org/10.1101/2023.02.15.528722
  27. Spatial Registration Adds Anatomical Context • Validate detection of laminar

    structure • Correlate enrichment t-statistics for top marker genes of reference ◦ Cluster vs. manual annotation • Annotate with strongly associated histological layer 43 Sp k D d ~L doi.org/10.1101/2023.02.15.528722
  28. Spatial Registration of Spatial Domains • Map SpDs to Maynard

    et al. manual annotated layers • Highlight most strongly associated histological layer to add biological context 44 doi.org/10.1101/2023.02.15.528722
  29. Identify Layer Associated Neuron Populations 45 • Apply Spatial Registration

    with manual layers • 13 layer-level cell types ◦ Assign Excitatory Neurons histological layers ◦ Pool other cell type groups Kelsey D Montgomery
  30. Spot Deconvolution 47 Cell 1 Cell 2 … Cell N

    Gene 1 0 0 … 0 Gene 2 2 5 … 3 … … … … … Gene i 1 0 … 0 Spot 1 Spot 2 … Spot M Gene 1 1 0 … 3 Gene 2 0 1 … 0 … … … … … Gene j 4 2 … 2 Astro Excit … Inhib Spot 1 1 1 … 1 Spot 2 … … … … … … Spot N 1 0 … 2 Single- Nucleus Spatial Deconvolved Results Spot 1 Nicholas J Eagles @Nick-Eagles (GitHub)
  31. Existing Spot Deconvolution Software - Explored 3 novel software methods

    from the literature 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 48 Excit L5 Counts
  32. 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 49 Nicholas J Eagles @Nick-Eagles (GitHub) doi.org/10.1101/2023.02.15.528722
  33. Visium Spatial Proteogenomics (Visium-SPG) Visium-SPG = Visium SRT + immunofluorescence

    (using identical tissue samples) Sang Ho Kwon @sanghokwon17
  34. Visium-SPG for Paired Imaging and Gene Expression - Channels measure

    proteins marking for broad cell types: microglia, neurons, oligodendrocytes, and astrocytes Kristen R. Maynard doi.org/10.1101/2023.02.15.528722
  35. 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: 53 • Astrocyte (GFAP) • Neuron (NeuN) • Oligodendrocyte (OLIG2) • Microglia (TMEM119) • Other (low signal in all channels) doi.org/10.1101/2023.01.28.525943 Sriworarat, 2023. samuibrowser.com Sang Ho Kwon @sanghokwon17
  36. Segmenting Cells on Visium-SPG IF Images 1. Segment cells on

    IF image 2. Manually label N cells 3. Train cell-type classifier and apply on remaining data Sriworarat, 2023. samuibrowser.com 54 Nicholas J Eagles @Nick-Eagles (GitHub) doi.org/10.1101/2023.01.28.525943
  37. 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 55 Annie B. Nguyen
  38. Constructing Dataset of Labeled Cells 56 1. Segment cells on

    IF image 2. Manually label N cells 3. Train cell-type classifier and apply on remaining data Annie B. Nguyen 4 sections * 5 cell types * 30 cells = 600 manually labeled cells doi.org/10.1101/2023.01.28.525943 Sriworarat, 2023. samuibrowser.com
  39. 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 57 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 total 600 old cells + 320 new cells = 920 labeled cells Nicholas J Eagles @Nick-Eagles (GitHub)
  40. Training Cell-Type Classifier 58 Model Test Precision Test Recall Decision

    tree 0.86 0.87 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
  41. Layer-level: collapse Excit_L* into 1 for broad-level 60 Then combine

    Excitatory and Inhibitory into Neurons to compare vs CART-calculated counts from Visium_SPG
  42. Benchmark Results: Leverage Prior Knowledge - Manually annotate spots with

    histological layer - Explore how cell-type predictions map to annotated layers Kristen R. Maynard A B C
  43. Benchmark Results: Self-Consistency of Results Counts from software results using

    both broad and layer-level cell types were compared, by “collapsing” onto just 4 major cell types. We expect results perfectly on the diagonal! 63
  44. Viewing Spot Deconvolution Results: Samui Browser - View: - Fluorescence

    channels - Spot deconvo results - Segmented cells - Gene expression - Interactive - Quickly zoom/scroll - Full-resolution images samuibrowser.com/from?url=data2.loopybrowser.com/VisiumIF/&s=Br2720_Ant_IF&s=Br6432_Ant_IF&s=Br6522_Ant_IF&s=Br8667_Post_IF Sriworarat, 2023. 64
  45. Viewing Spot Deconvolution Results: spatialLIBD apps - View: - spot

    deconvolution results - spatial domains/ clusters - gene expression - Huge amount of aesthetic customization 65 https://libd.shinyapps.io/spatialDLPFC_Visium_SPG/
  46. 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 66 Boyi Guo Melissa Grant-Peters
  47. Visium spatial clustering works for variables with high % variance

    explained. But what about other ones? DOI: 10.1038/s41593-020-00787-0
  48. AD pathology signal is too small to detect by spatially-resolved

    gene expression alone research.libd.org/Visium_SPG_AD/
  49. sc/snRNA-seq QC metrics such as # detected genes, # UMI,

    mitochondria expression % are likely biologically related!
  50. Having more data is useful to provide context! Here 4

    new samples have low sequencing saturation (outliers) but are within range of good samples from other studies
  51. Having more data is useful to provide context! Those 4

    samples have great median UMI counts per spot ^_^
  52. Software keeps evolving and as leaders in the field we

    aim to use the best methods 78 Moses, L., Pachter, L. Museum of spatial transcriptomics. Nat Methods 19, 534–546 (2022). https://doi.org/10.1038/s41592-022-01409-2
  53. The Development Process - Making a module - New, experimental

    software can change dramatically (function and syntax) between versions - Promotes collaboration by allowing two researchers to share exact code and instantly run software without special set-up SpatialExperiment release 3.14 SpatialExperiment devel 3.15 module load tangram/1.0.2 module load cell2location/0.8a0 module load spagcn/1.2.0 https://github.com/LieberInstitute/jhpce_mod_source https://github.com/LieberInstitute/jhpce_module_config Nicholas J Eagles @Nick-Eagles (GitHub)
  54. The Development Process - Regular interaction with software authors to

    clarify functionality and report bugs - Documentation for code and author responsiveness on GitHub can be critical in successfully applying software to our data Nicholas J Eagles @Nick-Eagles (GitHub)
  55. More challenges ahead Working with multiple capture areas per tissue

    Nicholas J Eagles @Nick-Eagles (GitHub) Prashanthi Ravichandran @prashanthi-ravichandran (GH) Spot diameter error: ~1.8 → ~1.1 Another pair: ~2.8 → ~0.76
  56. lcolladotor.github.io/#projects • Every assay has caveats • We re-use tricks:

    think adding 0, multiplying by 1 • It nearly always takes a team • Data sharing accelerates science + democratizes access to it • Zooming in allows us to reduce the heterogeneity • We can learn from each other: from uniformly processing our data & re-using it → replicate / validate?
  57. www.youtube.com/@lcolladotor/playlists Videos allow us to multiply ourselves We can make

    you custom selections of videos for a specific problem on DSgs sessions
  58. @MadhaviTippani Madhavi Tippani @HeenaDivecha Heena R Divecha @lmwebr Lukas M

    Weber @stephaniehicks Stephanie C Hicks @abspangler Abby Spangler @martinowk Keri Martinowich @CerceoPage Stephanie C Page @kr_maynard Kristen R Maynard @lcolladotor Leonardo Collado-Torres @Nick-Eagles (GH) Nicholas J Eagles Kelsey D Montgomery Sang Ho Kwon Image Analysis Expression Analysis Data Generation Thomas M Hyde @lahuuki Louise A Huuki-Myers @BoyiGuo Boyi Guo @mattntran Matthew N Tran @sowmyapartybun Sowmya Parthiban Slides available at speakerdeck.com /lcolladotor + Many more LIBD, JHU, and external collaborators @mgrantpeters Melissa Grant-Peters @prashanthi-ravichandran (GH) Prashanthi Ravichandran