Analysis of HIPC data
As an illustrative example, we analyze here the flow cytometry data from the T-cell panel of the Human Immunology Project Consortium (HIPC) publicly available on ImmuneSpace Gottardo et al. [2014].
An HIPC data set has the following structure (split into 2 files):
xx_y_values: flow-cytometry measurementsxx_y_clust: file with the corresponding manual clustering
Above, xx denotes the center where the data analysis was performed, and y denotes the patient and the replicate of the biological sample in question.
Data load
library(CytOpT)
data("HIPC_Stanford")Here are the first few lines of the flow-cytometry measurements from patient 1228 replicate 1A:
knitr::kable(head(HIPC_Stanford_1228_1A))| CCR7 | CD4 | CD45RA | CD3 | HLADR | CD38 | CD8 |
|---|---|---|---|---|---|---|
| 717.3339 | 1146.5768 | 3094.811 | 2526.265 | 1333.1118 | 1510.164 | 3203.711 |
| 681.8582 | 1398.1466 | 3168.901 | 2394.297 | 918.4464 | 1306.356 | 3056.621 |
| 402.3024 | 920.2601 | 3440.265 | 2221.533 | 1585.4507 | 1086.020 | 2728.252 |
| 1509.6527 | 1492.2483 | 3143.388 | 2592.564 | 1116.4272 | 1577.515 | 3191.382 |
| 1365.9507 | 659.6762 | 3382.406 | 2126.847 | 1317.9880 | 1277.266 | 3079.729 |
| 1388.9562 | 1213.0846 | 3486.772 | 2450.046 | 1103.4678 | 1474.890 | 3200.576 |
The manual clustering of these data into 10 cell populations (CD8 Effector, CD8 Naive, CD8 Central Memory, CD8 Effector Memory, CD8 Activated, CD4 Effector, CD4 Naive, CD4 Central Memory, CD4 Effector Memory, CD4 Activated) can be accessed from the HIPC_Stanford_1228_1A_labels object.
We will use the manual gating from patient 1228 replicate 1A as our source proportions to infer proportions for patient 1369 replicate 1A.
Computation of the benchmark class proportions for target data
Because in this example, we know the true proportions in the target data set HIPC_Stanford_1369_1A, we can assess the gap between the estimate form CytOpt and the cellular proportions from the reference manual gating. For this purpose, we compute those manual proportions with:
gold_standard_manual_prop <- c(table(HIPC_Stanford_1369_1A_labels)/length(HIPC_Stanford_1369_1A_labels))CytOpT
Optimization
set.seed(123)
res <- CytOpT(X_s = HIPC_Stanford_1228_1A, X_t = HIPC_Stanford_1369_1A,
Lab_source = HIPC_Stanford_1228_1A_labels,
theta_true = gold_standard_manual_prop,
method='both', monitoring = TRUE)
#> Running Descent-ascent optimization...
#> Done in 1.1 mins
#> Running MinMax optimization...
#> Done in 15.2 secsResults
The results from CytOpt for both optimization algorithms are:
summary(res)
#> Estimation of cell proportions with Descent-Ascent and MinMax swapping from CytOpt:
#> Gold_standard Descent_ascent MinMax
#> CD8 Effector 0.017004001 0.052364999 0.045657219
#> CD8 Naive 0.128736173 0.086210769 0.102414966
#> CD8 Central Memory 0.048481996 0.036818260 0.038026528
#> CD8 Effector Memory 0.057484114 0.064841660 0.072192722
#> CD8 Activated 0.009090374 0.017701245 0.008766140
#> CD4 Effector 0.002324076 0.008051896 0.008304633
#> CD4 Naive 0.331460344 0.342359929 0.351420976
#> CD4 Central Memory 0.281713344 0.209007941 0.204257964
#> CD4 Effector Memory 0.102082843 0.168264983 0.163335122
#> CD4 Activated 0.021622735 0.014378317 0.005623730
#>
#> Final Kullback-Leibler divergences:
#> Descent-Ascent MinMax swapping
#> 0.07071936 0.06321548
#> Number of iterations:
#> Descent-Ascent MinMax swapping
#> 5000 10000Some visualizations are provided by the plot() method:
plot(res)
#> Plotting KL divergence for iterations 10 to 1000 while there were at least 5000 iterations performed for each method.
Performance evaluation
Concordance between the manual gating gold-standard and CytOpt estimation can be graphically diagnosed with Bland-Altman plots:
Bland_Altman(res$proportions)
The methods implemented in the CytOpt package are detailed in the following article:
Paul Freulon, Jérémie Bigot, Boris P. Hejblum. CytOpT: Optimal Transport with Domain Adaptation for Interpreting Flow Cytometry data https://arxiv.org/abs/2006.09003