sc-OTGM: Single-Cell Perturbation Modeling by Solving Optimal Mass Transport on the Manifold of Gaussian Mixtures

Introduction

sc-OTGM provides a scalable and unified framework for:

1. Cell state classification
2. Differential gene expression analysis
3. Gene ranking for target identification through a recommender system
4. Perturbation response prediction
5. Generation of synthetic scRNA-seq data

Setup and Installation

• It is highly suggested that you install all dependencies into a separate virtual environment for easy package management.
• The dependencies are in requirements.txt. You will need to install dependencies by running in the root directory:

  $conda create -n <myenv> python=3.10
  $conda activate <myenv>  
  $pip install -e .
  $python -m ipykernel install --user --name=<myenv>

• Please check you have the same versions of these dependencies.

Code style

Perform these steps manually in the root directory:

source format_and_lint.sh
format-and-lint .

Loading benchmark datasets

CROP-seq

To evaluate the performance of sc-OTGM, we used the CROP-seq dataset from in-vitro experiments on human-induced pluripotent stem cell (iPSC)-derived neurons subjected to genetic perturbations Tianet al. (2021). These perturbations were executed via CRISPRi, enabling targeted gene knockdown to investigate its effects on neuronal survival and oxidative stress. Using the rank genes groups method from scanpy package Wolf et al. (2018) for differential expression analysis, we scrutinized the effects of knocking down 185 genes identified as potentially relevant to neuronal health and disease states. Of these, only 57 genes met our significance threshold (adjusted p-value less than 0.05), indicating a significant alteration in expression levels post-perturbation. Raw published data is available from the Gene Expression Omnibus under accession code GSE152988.

Baseline comparison on CROP-seq

In CRISPR interference (CRISPRi) experiments, not all cells receiving the CRISPR components achieve successful targeted gene knockdown. Additionally, the extent of gene suppression can vary significantly among cells due to variations in Cas9 activity, guide RNA efficiency, and individual cellular responses. Following gene knockdown, cells often activate compensatory mechanisms that alter the expression profiles of other genes. This requires robust computational models that can accurately identify genes that have been knocked down and predict the subsequent changes in the expression of other genes.

To address this, we benchmarked statistical models for their efficacy in differential gene expression analysis post-CRISPRi. The techniques include the Mann-Whitney U test, t-test, and sc-OTGM, with the results detailed in Table 1. We assessed the ability of each method to rank the true knockeddown gene within the top-k results, with performance metrics based on how lower p-values from the Mann-Whitney U test and t-test correlate with higher rankings. sc-OTGM demonstrated superior performance, particularly in Top-1 accuracy, showing its effectiveness in identifying the most likely perturbed gene. The performance advantage of using sc-OTGM decreases as the ranking threshold increases.

Table 1: Benchmarking Statistical Differential Gene Expression Analysis Techniques

Method	Top-1 Accuracy	Top-5 Accuracy	Top-10 Accuracy	Top-50 Accuracy	Top-100 Accuracy
Mann–Whitney U test	0.37	0.40	0.42	0.58	0.60
t-test	0.39	0.67	0.70	0.79	0.86
sc-OTGM	0.56	0.68	0.74	0.82	0.91

Table 2 shows the performance of sc-OTGM in identifying differentially expressed genes (DEGs) following targeted gene knockdowns in CRISPRi experiments. Based on sc-OTGM’s ranking, a cutoff of 100 is set to select the top predicted DEGs. Predicted DEGs are obtained from the top of the ranked gene list, and compared against the list of known DEGs. We conducted Fisher’s exact tests which yielded p-Values, significantly below the standard threshold of 0.05. This confirms a strong statistical correlation between known DEGs after targeted gene knockdown and those predicted by sc-OTGM.

The performance of sc-OTGM is also evaluated based on its accuracy to predict the direction of expression changes (upregulation or downregulation) in DEGs. The metrics include Accuracy (%) and F1-score, where the mean accuracy is 75.2% with a standard deviation of 15.4%, and the mean F1-score is 0.79 with a standard deviation of 0.14.

Table 2: Quantitative Analysis of Gene Perturbation Responses

Gene	p-Value	Accuracy (%)	F1-score
TUBB4A	7.37 × 10^−6	100.0	1.00
ATP1A3	3.12 × 10^−20	78.7	0.85
KIFAP3	6.39 × 10^−10	87.5	0.67
MAPT	9.01 × 10^−4	100.0	1.00
CASP3	6.93 × 10^−6	100.0	1.00
APEX1	2.61 × 10^−16	100.0	1.00
COX10	1.43 × 10^−16	83.7	0.87
NDUFS8	2.29 × 10^−36	82.7	0.83
ZNF292	4.91 × 10^−16	65.4	0.72
GSTA4	5.50 × 10^−21	89.5	0.92
STX1B	4.95 × 10^−38	72.1	0.76
OPTN	1.51 × 10^−6	100.0	1.00
SOD1	1.16 × 10^−7	88.9	0.90
NDUFV1	4.33 × 10^−30	73.9	0.79
CALB1	1.38 × 10^−4	40.0	0.46
EEF2	5.08 × 10^−29	92.2	0.92
BIN1	6.39 × 10^−8	88.9	0.89
SCFD1	1.59 × 10^−42	56.4	0.66
PON2	8.41 × 10^−55	53.6	0.60
BAX	1.27 × 10^−30	78.3	0.83
SCAPER	1.48 × 10^−31	87.2	0.89
CYB561	5.66 × 10^−33	60.4	0.66
AKAP9	3.69 × 10^−14	100.0	1.00
VPS35	9.36 × 10^−14	80.0	0.85
PRNP	3.95 × 10^−53	59.4	0.72
AP2A2	2.48 × 10^−57	75.4	0.82
SOD2	2.01 × 10^−7	90.5	0.91
BECN1	6.22 × 10^−4	73.1	0.75
SNCB	6.79 × 10^−41	87.7	0.89
CDH11	1.77 × 10^−14	66.7	0.70
ELOVL5	2.28 × 10^−14	92.3	0.93
NTRK2	4.50 × 10^−29	58.7	0.61
DAP	1.27 × 10^−45	81.7	0.86
EIF4G1	3.00 × 10^−24	76.4	0.84
TRPM7	3.66 × 10^−14	66.7	0.80
COASY	1.51 × 10^−6	83.3	0.84
TRAP1	2.11 × 10^−35	80.0	0.89
CYP46A1	4.56 × 10^−4	50.0	0.33
PARP1	5.51 × 10^−24	57.7	0.67
FOXRED1	1.49 × 10^−25	75.8	0.78
AFG3L2	1.01 × 10^−15	75.0	0.81
RAB7A	1.21 × 10^−12	83.3	0.84
PPP2R2B	3.10 × 10^−35	63.9	0.75
RGS2	5.03 × 10^−30	63.0	0.68
AMFR	4.06 × 10^−12	62.5	0.74
MRPL10	1.11 × 10^−22	43.6	0.57
ANO10	1.38 × 10^−4	40.0	0.46
DMXL1	6.72 × 10^−33	66.7	0.76
HYOU1	7.43 × 10^−35	55.3	0.65
HTT	8.81 × 10^−24	61.1	0.66
ECHS1	5.44 × 10^−9	71.4	0.79
CYCS	7.80 × 10^−3	77.8	0.80
CEP63	2.80 × 10^−14	75.0	0.77
FARP1	1.35 × 10^−22	83.3	0.88
FRMD4A	8.53 × 10^−31	83.1	0.84
RPL6	2.74 × 10^−36	67.7	0.77
PFN1	3.91 × 10^−16	80.0	0.85

• Gene Ranking Mechanism: sc-OTGM's recommendation system ranks genes across selected cell types within each dataset. This ranking approach prioritizes genes that are more likely to shift the distribution of the cell state from healthy to diseased.

Citation

Single-Cell Perturbation Modeling by Solving Optimal Mass Transport on the Manifold of Gaussian Mixtures. ICLR 2024 Workshop on Machine Learning for Genomics Explorations.

Corresponding author: andac.demir@novartis.com