Hierarchical Automated Annotation of cell types

Viktor Petukhov , Peter Kharchenko

1,2

2,3

University of Copenhagen, BRIC
Harvard Medical School, DBMI
Harvard Stem Cell Institute

https://bit.ly/2PVjasv

Problem

Manual annotation is painful!

Existing solutions

scVI
scmap
Seurat
conos
... (hundred more packages)
CellAssign
p-DCS

Based on

annotation

transfer

Based on

marker genes

Annotation transfer

Annotated cells

(e.g. published data)

Not-annotated

cells (e.g. your data)

Problems:

Transfer between datasets suffers from batch-effect
Isn't aware about cell type relations
If it doesn't work, there is nothing you can do

Marker-based assignment

*Probabilistic cell-type assignment of single-cell RNA-seq for tumor microenvironment profiling, Nature Methods 2019

Benefits

Intuitive
Verifiable
Transferable across datasets

Drawbacks

Suffers from drop-outs
Requires detailed annotation
"Binarizes" expression matrix

Workflow

Inspired by Garnett*

Pliner. HA, et. al. Supervised Classification Enables Rapid Annotation of Cell Atlases. Nat. Methods 2019

Modifications of the original idea

Cell Type scoring

Estimate TF-IDF over expression matrix
Normalize by features
Estimate positive scores as sum over positive markers
Negative score is a multiplier for positive score:

s_{neg}(t) = 1 - \frac{max_{g_i \in neg(t)}(E(g_i))}{max_{g_j \in pos(t)}(E(g_j))}

s(t) = s_{pos}(t) * s_{neg}(t)

s_{pos}(t) = \Sigma_{g_i \in pos(t)}E(g_i)

Modifications of the original idea

Label propagation

Graph diffusion (using Conos routines):

~100 times faster then Garnett GLMs
Allow to work with alignment graph

More pronounced work with hierarchies
Workflow, optimized for easy marker selection
Works on graphs, so can use joint graph from multiple samples

Modifications of the original idea

Other improvements

Benchmarks: MCA Lung

> AT2
expressed: Bex4

> AT1
expressed: Cryab

>Ciliated cells
expressed: Aldh1a1, Cyp2f2

> Interstitial macrophage
expressed: Apoe, Pf4
not expressed: Trbc2

> Alveolar macrophage
expressed: Ear1, Ear2

>T cells
expressed: Cd8b1, Trbc2

>Natural killer cells
expressed: Klra8, Nkg7
not expressed: Trbc2

> Naaa DCs
expressed: Naaa

> Mgl2 DCs
expressed: Mgl2

> Plasmacytoid DCs
expressed: Plac8

> H2-M2 DCs
expressed: Epsti1, H2-M2

>Granulocytes
expressed: Il1b, Il1r2

>Endothelial
expressed: Pecam1, Flt1, Chd5, Kdr

>Fibroblasts
expressed: Dcn, Acta2, Inmt

>B cells
expressed: Cd19, Ms4a1, Cd79a

>Monocyte progenitor cell
expressed: Ctsg, Mpo

>Basophil
expressed: Ccl3, Ccl4

Cell Types

Garnett

Accuracy: 32.3%

Unclassified: 62.7%

Average TPR: 40.1%

Average Precision: 73.2%

Our code

Accuracy: 96.6%

Average TPR: 93.9%

Average Precision: 90.3%

Benchmarks: MCA Lung

Paper

CellAnnotatoR

Garnett

Benchmarks: MCA Lung, CellAssign

We couldn't get good results with CellAssign

(and we're not alone in this: Issue #35 "About results reproducibility")

CellAssign also doesn't use info about negative markers and cell type hierarchies

Accuracy: 7.0%

Average TPR: 10.2%

Average Precision: 6.2%

Benchmarks: MERFISH

Cell Types

>Inhibitory
expressed: Gad1
not expressed: Slc17a7

>Excitatory
expressed: Slc17a6, Slc17a7, Sema3c
not expressed: Gad1

>OD Mature
expressed: Ttyh2, Mbp, Opalin
not expressed: Pdgfra

>OD Immature
expressed: Pdgfra, Mki67

>Astrocyte
expressed: Aqp4

>Microglia
expressed: Selplg

>Ependymal
expressed: Cd24a
not expressed: Gad1

>Endothelial
expressed: Fn1

>Pericytes
expressed: Myh11

>Endothelial 1
expressed: Igf1r
subtype of: Endothelial

>Endothelial 2
expressed: Bmp7, Lepr
subtype of: Endothelial

>Endothelial 3
expressed: Ace2
subtype of: Endothelial

>OD Immature 1
expressed: Traf4
subtype of: OD Immature

>OD Immature 2
expressed: Mki67
subtype of: OD Immature> AT2
expressed: Bex4

> AT1
expressed: Cryab

>Ciliated cells
expressed: Aldh1a1, Cyp2f2

> Interstitial macrophage
expressed: Apoe, Pf4
not expressed: Trbc2

> Alveolar macrophage
expressed: Ear1, Ear2

>T cells
expressed: Cd8b1, Trbc2

>Natural killer cells
expressed: Klra8, Nkg7
not expressed: Trbc2

> Naaa DCs
expressed: Naaa

> Mgl2 DCs
expressed: Mgl2

Garnett

Accuracy: 23.3%

Unclassified: 75.1%

Average TPR: 8.4%

Average Precision: 26.2%

CellAnnotatoR

Accuracy: 90.0%

Average TPR: 84.6%

Average Precision: 83.7%

Benchmarks: MERFISH

Paper

CellAnnotatoR

Garnett

Black crosses are ambiguous

Benchmarks: Human Cortex

(our data)


>Astrocytes
expressed: SLC1A3, GJB6, FGFR3
not expressed: RBFOX3, SYP

> Microglia
expressed: CX3CR1, GPR34, P2RY12, MRC1
not expressed: RBFOX3, SYP

>Oligodendrocytes
expressed: MOG, ERMN
not expressed: RBFOX3, SYP

>Oligodendrocyte Precursors
expressed: CSPG4, PDGFRA, VCAN
not expressed: RBFOX3, SYP

>Vascular
expressed: DCN, PTGDS, ATP1A2, ITIH5, FLT1
not expressed: RBFOX3, SYP

>Neurons
expressed: SYT1, SYP, SNAP25, RBFOX3
not expressed: MOG, ERMN, SLC1A3, CX3CR1, GPR34

# Neurons

>Inhibitory
expressed: GAD1, GAD2, SOX6, PVALB, SST, VIP, LHX6, NDNF, CALB2, SULF1
not expressed: SLC17A7, SATB2
subtype of: Neurons

>Excitatory
expressed: SLC17A7, SATB2, RORB, CUX2, TLE4, NR4A2, SEMA3C
not expressed: GAD1, GAD2, SOX6, PVALB
subtype of: Neurons

# Inhibitory

>Pvalb
expressed: PVALB, NOS1, SULF1, LHX6, KCNS3, CRH, PLEKHH2
not expressed: LAMP5, ID2, SST, FAM89A, RELN, SEMA6A, TAC3, DDR2, VIP
subtype of: Inhibitory

>Lamp5
expressed: ID2, LAMP5, SV2C, PDGFD, CCK, RELN
not expressed: VIP, CALB2, SST, FAM89A, DDR2, NR2F2
subtype of: Inhibitory

>Sst
expressed: SST, NOS1, SEMA6A, FAM89A, LHX6
not expressed: VIP, CALB2, CRH, CHAT, CCK, LAMP5, ID2, SV2C, PDGFD, PVALB, KCNS3
subtype of: Inhibitory


>Vip
expressed: VIP, TAC3, CALB2, NR2F2, LAMA3, COL5A2, SEMA3C, FAM19A1
not expressed: ID2, NOS1, LAMP5, PDGFD
subtype of: Inhibitory

## PVALB

>Pvalb_Nos1
expressed: NOS1
not expressed: CRH
subtype of: Pvalb

>Pvalb_Sulf1
expressed: SULF1
not expressed: NOS1, CRH
subtype of: Pvalb

>Pvalb_Crh
expressed: CRH, PLEKHH2
not expressed: NOS1, RGS5
subtype of: Pvalb

## LAMP5

>Lamp5_Nos1
expressed: NOS1, SFRP1
not expressed: LAMA3
subtype of: Lamp5

>Lamp5_Crh
expressed: CRH, SFRP1
subtype of: Lamp5

>Lamp5_Reln
expressed: RELN, LAMA3
not expressed: ID2
subtype of: Lamp5

## SST

>Sst_Tac3_Lhx6
expressed: TAC3, LHX6
not expressed: CALB1
subtype of: Sst

>Sst_Calb1
expressed: CALB1
not expressed: TAC3
subtype of: Sst

## VIP

>Vip_Crh
expressed: CRH, TAC3, IGFBP5
not expressed: SEMA3C, SEMA6A, NR2F2
subtype of: Vip

>Vip_Nr2f2
expressed: CRH, NR2F2, IGFBP5
not expressed: SEMA3C, SEMA6A, TAC3, RELN
subtype of: Vip

>Vip_Sema3
expressed: SEMA3C, SEMA6A, COL5A2
not expressed: CRH, RELN
subtype of: Vip

>Vip_Reln
expressed: RELN, DDR2
not expressed: TAC3, SEMA3C, IGFBP5
subtype of: Vip

>Vip_Cck
expressed: CCK, FAM19A1, NR2F2
not expressed: RELN, TAC3, IGFBP5, SEMA3C
subtype of: Vip

# Excitatory

>L2/3_Cux2
expressed: LAMP5, CUX2, COL5A2
not expressed: PDGFD, FAT4, PARD3, PRSS12, GABRG1, COBLL1, PXDN
subtype of: Excitatory

>L2_Lamp5
expressed: LAMP5, CUX2, PDGFD, PARD3
not expressed: RORB, GABRG1, COL5A2, PXDN
subtype of: Excitatory

>L3_Prss12
expressed: PRSS12, RORB, COBLL1, CUX2
not expressed: LAMP5, GABRG1, GRIN3A, CMTM8, PXDN, OPRK1, PDGFD, FAT4, PDZD2
subtype of: Excitatory

>L3_Plch1
expressed: PRSS12, RORB, COBLL1, PLCH1
not expressed: LAMP5, GABRG1, GRIN3A, CMTM8, PXDN, OPRK1
subtype of: Excitatory

>L4_Rorb
expressed: RORB, GABRG1, CUX2
not expressed: PRSS12, CMTM8, PXDN, OPRK1, LAMP5
subtype of: Excitatory

>L5_Grin3a
expressed: GRIN3A, TLL1, CMTM8, RORB, TOX
not expressed: HTR2C, CUX2, PXDN, OPRK1, GABRG1
subtype of: Excitatory

>L5_Htr2c
expressed: HTR2C, PARD3, NXPH2, TLE4
not expressed: CMTM8, PXDN, LGR6
subtype of: Excitatory

>L6_Nr4a2
expressed: NR4A2, POSTN, HTR2C
not expressed: PRSS12, KCNIP1, NXPH2, PXDN
subtype of: Excitatory

>L6_Syn3
expressed: PXDN, OPRK1
not expressed: CUX2, RORB, HTR2C, CMTM8
subtype of: Excitatory

>L6_Tle4
expressed: TLE4, LGR6
not expressed: CUX2, RORB, HTR2C
subtype of: Excitatory

## L5_Grin3a

> L5_Grin3a_Fstl4
expressed: FSTL4, PRKG1
not expressed: FAM19A1, NTM, RGS6, SLIT3
subtype of: L5_Grin3a

> L5_Grin3a_Tox
expressed: TOX, DCC
not expressed: FAM19A1, NTM, ROBO2, RGS6, SLIT3
subtype of: L5_Grin3a

> L5_Grin3a_Slit3
expressed: FAM19A1, NTM, ROBO2, RGS6, SLIT3
subtype of: L5_Grin3a

## L6_Tle4

> L6_Tle4_Lsamp
expressed: LSAMP, RYR2
not expressed: CDH10, CNTN4
subtype of: L6_Tle4

> L6_Tle4_Cdh10
expressed: CDH10, CNTN4
not expressed: LSAMP, RYR2
subtype of: L6_Tle4

Cell Types

Cell type hierarchy