Figure 2. The DNA methylome contains extensive brain cellular diversity that integrable with other modalities.

[lhqing]

Figure files

https://drive.google.com/drive/u/1/folders/17ukhBF1y_acxglZtf8wsH4Hfw329q8-9

Panel legends

• [x] a. 4673 snmC Cell cluster centroids tSNE colored by 261 cell sub-classes. Dots are sized by the number of nuclei in each cluster. Numbers correspond to annotation lists below the dots.
• [x] b. 301,626 snmC cells tSNE colored by 117 dissection regions.
• [x] c. 3D model of the 117 dissection regions grouped into nine major anatomical regions, cells from each major region is colored in the corresponding small tSNE plot; the remaining cells are grey dots in the background.
• [x] d. Stacked barplot showing the cell sub-classes (upper) and neurotransmitters (lower) composition of 117 dissection regions, grouped in the same way as c.
• [x] e. Ensemble integration tSNE plot of all the neurons (n=X) from four whole-brain datasets from BICCN. In each sub-panel, cells from the corresponding dataset are colored by matched sub-class labels (see methods). The other datasets are grey dots in the background.
• [x] f. Cluster-cluster matching table between the 4673 snmC-seq clusters (columns) and X transcriptome clusters (rows). The floating sub-panels show two example integration tSNE plots and confusion matrices used to match cluster relationships.
• [x] g. Matched cell-type-specificity between mCG fraction (left) and chromatin accessibility (right) of the cluster-specific CG-DMRs (columns) in 261 cell sub-classes (rows). Each dot represents an aggregated profile of 1000 regions, colored by the peak or valley offset and sized by the area under the curve (see methods).

Supplementary files

• [ ] Extended Data Figure 2-1: Details of iterative clustering; neurotransmitters
• [ ] Extended Data Figure 2-2: Matched iterative clustering plots for snm3C-seq cell type and clusters, details of mC-m3C integration
• [ ] Extended Data Figure 2-3: Details of mC-RNA integration
• [ ] Extended Data Figure 2-4: Details of mC-ATAC integration
• [ ] Supplementary Table 2-1: Summary of mC-m3C integration
• [ ] Supplementary Table 2-2: Summary of mC-RNA integration
• [ ] Supplementary Table 2-3: Summary of mC-ATAC integration

Methods

• [ ] a-d. preprocessing and file preparation of snmC and snm3C-seq (methylome part).
• [ ] a-d. Iterative clustering of snmC and snm3C-seq datasets.
• [ ] a-d. Iterative integration method between mC and m3C.
• [ ] e. special ensemble integration to generate the four-dataset embedding
• [ ] f. Iterative integration of mC and RNA.
• [ ] f. Generate mC cluster-matched RNA profiles.
• [ ] g. Iterative integration of mC and ATAC.
• [ ] g. Generate mC cluster-matched ATAC profiles.
• [ ] g. Plot-specific method to make the aggregated dot profile. For DMR calling details, see Fig. 3.

Reproducibility

Code to generate the figure

• [ ] a-d. iterative clustering: template code; full iterative code archive
• [ ] a-d. plotting code
• [ ] e. code for four-dataset integration and embedding
• [ ] e. plotting code
• [ ] f. mC-RNA iterative integration: template code; full iterative code archive
• [ ] f. plotting code
• [ ] g. mC-ATAC iterative integration: template code; full iterative code archive
• [ ] g. plotting code: ecker-hanqing-analysis/221122-plot-mC-ATAC-integration

Processed data

• [ ] a-g. mC, m3C, RNA, ATAC cell-by-feature datasets
• [ ] a-d. Cluster metadata and annotation zarr; 25um nrrd file
• [ ] e. Embedding coords file of four-dataset integration
• [ ] f. mC-RNA integration summary zarr, including integration embedding, co-cluster and confusion matrix; csv table;
• [ ] g. mC-ATAC integration summary zarr, including integration embedding, co-cluster and confusion matrix; csv table; DMR regions used to scan bigwig; bigwig file paths;

Internal API

Input dataset and API

• [x] mC Cell MCDS
• [x] m3C Cell MCDS
• [x] RNA Zarr
• [x] ATAC Zarr
• [x] DMR regions
• [x] Genome mCG and ATAC bigwig tracks

Output dataset and API

• [x] Cluster-cluster match: m3C to mC; RNA to mC; ATAC to mC.
• [x] Integration zarr of coords, co-clusters.
• [x] Cluster aggregated multi-omic profiles, mC cluster to all other modalities.

[lhqing]

Some observations about integration based on the current algorithm (ALLCools v1.0.14)

1. Include more genes in the beginning, e.g., use all CEF from the last round, use CEF from both ref and query data
2. Before scale, filter features by a minimum std (e.g., > 0.005) in both datasets. Don't include features with little variance
3. Features need to be scaled and mean-centered before decomposition.
4. n_pc and n_cc matter, especially when the cell cluster diversity is low in the dataset, including too many PC and CC negatively impacting integration results. (There is a scale by singular value step, so variance from small components got overestimated). How to determine n_pc and n_cc?
5. Which one is the reference? Which one is the query? Based on the analysis goal or based on which dataset is more diverse?
6. Do harmony to finalize the integration helps to improve the results in some low-diversity cases.

[lhqing]

mC - AIBS 10X Integration

gs://ecker-hanqing-analysis/221015-cemba-mc-aibs-tenx-integration

Integration strategy

1. Methylation-only clustering: We did iterative clustering and defined 4,673 cell clusters in our 301,626 methylome cells.
2. Cell-type annotation: Starting from the entire dataset, we did iterative integration and manual evaluation (dissection region, gene, embedding, etc.) to annotate our 4673 mC clusters into 261 cell types using the same name from the AIBS-10X dataset.
3. Cell-cluster annotation: Within each cell type and potential overlaps from (2), we redo the iterative integration to match 5208 RNA clusters with 4673 snmC clusters.

Cell Type Match

261 / 306 RNA cell types or 3988980 / 4065284 (98.1%) RNA cells found corresponding mC cell types. See the "Cell Type Label" column in the google sheet. Cell type annotation is one-to-one, with some potential overlapping noted in the next column.

Cell Cluster Match

4669 / 5208 RNA clusters or 3958398 / 4065284 (97.4%) RNA cells found corresponding mC cell clusters. There is a total of 2240 unique matching patterns. See the "Matched AIBS 10X RNA Clusters" column in the google sheet.

Example

• All Cortical Exc

  

• Inh and Subcortical

  

• One subtype BST Tac2 Gaba

  

• Gene body mCH and RNA match

  

[lhqing]

mC - ATAC Integration

Integration strategy

1. Methylation-only clustering: same as RNA integration
2. Iterative integration within each major region: After separating NN (integrated separately), we group all the neuronal cells into ten major regions and perform iterative integration within each major region.
3. Within each round of integration, ATAC cells are assigned to mC clusters via integration co-clusters. We found this way the dissection region distribution between mC and ATAC in the next round of integration is most correlated.

ATAC Cell to mC Cluster Match

2065820 / 2312406 (89.3%) ATAC cells assigned to at least one mC cluster.

ATAC Cluster to mC Cluster relationship

We also calculated the ATAC 600-cluster proportion of ATAC cells assigned to each mC cluster. See the "Matched CEMBA ATAC Clusters" column in the google sheet.

Example

• Major Region Mid Brain, all dissection regions:

  

• Dissection region MRN

  

• Co-cluster dissection region distribution, from left to right: mC DissectionRegion %; ATAC DissectionRegion %; Delta DissectionRegion %

  

• Regulatory elements (random selected) mCG and ATAC

  

[lhqing]

mC - m3C Integration

gs://ecker-hanqing-analysis/221022-cemba-mc-cemba-m3c-integration

1. Methylation-only clustering: same as RNA integration. For m3C, we also did the same iterative clustering separately.
2. m3C cell type Annotation: Starting from the entire dataset, we performed the iterative integration to annotate m3C cell types via the mC cell types and assign m3C cell clusters to mC cell clusters.

All m3C cell clusters are assigned to at least one mC cell cluster. All the mC cell types, except CB Purkinje Cell (c21 in mC L1 Clustering), has matched m3C clusters.

We found not doing harmony for mC-m3C receive better results.

[lhqing]

snmC - AIBS TENX Integration

• Option 1 - Currently go with this one:

  • Goal: Assign each mC cluster the most likely RNA profile
  • Reference dataset: mC
  • Integration Group Strategy: Group on mC clusters and assign RNA clusters by OS cutoff (allow RNA clusters assigned to multiple mC groups

• Option 2:

  • Goal: Assign each RNA cluster the most likely mC profile
  • Reference dataset: RNA
  • Integration Group Strategy: Group on RNA clusters and assign mC clusters by OS cutoff (allow mC clusters assigned to multiple RNA groups

• Option 3:

  • Goal: Co-integrate mC and RNA clusters
  • Reference dataset: mC (?)
  • Integration Group Strategy: Group on mC and RNA clusters, each cluster only assigned to one integration group.

• Always use the deepest cluster level to calculate the overlap score

Notes

AIBS v2 annot and v3 annot v3 L2annot 306 labels, not including "LQ"