Wed Mar 13, 2019 10:04 pm
Debarcoding efficiency affected by 89Y oxide into Pd105
Dear CyTOF community,
I am interested in evaluating the effect of 89Y oxide on samples barcoded with the Fluidigm 20plex kit. Please know that I have no meaningful computer sciences or bioinformatics background (could not right or read a code to save my life). However, I do like to think of myself as someone with whom you can have an intelligible conversation about the principles behind these tools we use. To get started let me 1st explain my experimental set-up:
- Samples: PBMCs from 2 different donors (human): Ref1 and Ref2
- Panel: 29-marker (surface only) which includes CD45 tagged to 89Y
- Design: Ref1 sample was split into 2 replicates (3 x 10^6 cells each); Ref2 was split into 4 replicates (3 x 10^6 cells each). Samples stained (then fixed) individually.
- Barcoding: 1 of 2 the Ref1 replicates was barcoded (BC#5). 3 of the 4 Ref2 replicates were barcoded (BC#s 2, 3, and 4). Barcoding was done simultaneously with the Iridium intercalation step. Barcoded samples were pooled for acquisition.
- Acquisition: 3 tubes were acquired, namely, Ref1 non-barcoded, Ref 2 non-barcoded, and the pooled barcoded samples (containing 1 replicate of Ref1 and 3 replicates of Ref2)
*** PS: The barcoded sample has over 5 x 10^6 events so in order to evaluate the effect of the different variables I am using a small subset (8K events) of the same tube that was acquired for QC. Otherwise it takes several minutes to generate the plots. The following Premessa screenshots are from that 8K-event file ***
After normalization I open the barcoded sample on Premessa and found the following separation plot:
My first observation is that the plot does not have the typical plateau which helps us choose the separation threshold. This is except for perhaps population #5 (the only barcode that contains 105Pd).
I then move on to inspect the event plots for each of the 4 barcoded populations. Here you find them using the standard debarcoding input (minimum separation of 0.3, no filtering by MD, which gives a 67% debarcoding efficiency):
I debarcoded using these standard parameters (minimum separation of 0.3, no filtering by MD) and uploaded the files on Cytobank to inspect (all the 4 debarcoded populations and the unassigned file).
I then proceed to process the files using manually tailored gate up to live siglets as I would if this was not a barcoded experiment.
Below are the CD3 x CD19 plots comparing the proportion of T cells, B cells, non-B non-T cells, and T-B doublets for each PBMC donor:
Finally I turn myself to analyse the file containing events not assigned to any barcode. Of note, I found could manually gate over 1 million of what look like perfectly good live singlets from the "unassigned" file. Here is the same plot. Please note the similarity in frequency distrubution this file has with the REF 2 replicates. For comparison I also included the same plot on the files containing all barcodes before deconvolution:
What I believe is happening is that the higher background on 105Pd is preferentially excluding events from barcodes that do not contain 105Pd (BC#2, BC#3, BC#4). What ends up happening is that the CD45_89Y oxide into 105Pd ends up being read as the 4th barcode isotope in those samples and they end up being excluded because they end up have a minimum barcode distance below the set threshold. This does not happen with the samples barcoded with BC#5 because that particular barcode has 105Pd in its combination.
I am obviously concerned with the debarcoding efficiency but also (and more importantly) I am concerned that this is introducing bias in the population frequencies. That is because T and B cells normally have a higher CD45 expression (and consequently a higher background on 105Pd in the case of my experiment) than NK cells. If I am understanding this correctly they (T and B cells) would have shorter BC separations and would therefore be more likely to be left unassigned compared to an NK cell event. This to me would be a very concerning bias to introduce.
Sorry for the long post but I have been giving this some thought for a while. My questions for the community are the following:
1. Does this make sense?
2. Assuming I want to keep CD45_89Y in my panel is there a way to prevent this from happening?
3. Would trying to increase the Pd BC median intensity (by adding the BC overnight with IR intercalation as opposed to 30min on the day of acquisition) potentially help by increasing the separation distance between the true barcode and the 105Pd background?
4. If this cannot be fixed on the experimental side, would it help to be less stringent with the minimum separation parameter and try to filter out unwanted events using Mahlanobis distance?
5. Any other suggestions?
I am interested in evaluating the effect of 89Y oxide on samples barcoded with the Fluidigm 20plex kit. Please know that I have no meaningful computer sciences or bioinformatics background (could not right or read a code to save my life). However, I do like to think of myself as someone with whom you can have an intelligible conversation about the principles behind these tools we use. To get started let me 1st explain my experimental set-up:
- Samples: PBMCs from 2 different donors (human): Ref1 and Ref2
- Panel: 29-marker (surface only) which includes CD45 tagged to 89Y
- Design: Ref1 sample was split into 2 replicates (3 x 10^6 cells each); Ref2 was split into 4 replicates (3 x 10^6 cells each). Samples stained (then fixed) individually.
- Barcoding: 1 of 2 the Ref1 replicates was barcoded (BC#5). 3 of the 4 Ref2 replicates were barcoded (BC#s 2, 3, and 4). Barcoding was done simultaneously with the Iridium intercalation step. Barcoded samples were pooled for acquisition.
- Acquisition: 3 tubes were acquired, namely, Ref1 non-barcoded, Ref 2 non-barcoded, and the pooled barcoded samples (containing 1 replicate of Ref1 and 3 replicates of Ref2)
*** PS: The barcoded sample has over 5 x 10^6 events so in order to evaluate the effect of the different variables I am using a small subset (8K events) of the same tube that was acquired for QC. Otherwise it takes several minutes to generate the plots. The following Premessa screenshots are from that 8K-event file ***
After normalization I open the barcoded sample on Premessa and found the following separation plot:
- Premessa - Separation Plot
My first observation is that the plot does not have the typical plateau which helps us choose the separation threshold. This is except for perhaps population #5 (the only barcode that contains 105Pd).
I then move on to inspect the event plots for each of the 4 barcoded populations. Here you find them using the standard debarcoding input (minimum separation of 0.3, no filtering by MD, which gives a 67% debarcoding efficiency):
- Event plot - Population 2 (Ref2) - Separation 0.3 - MD 30
- Event plot - Population 3 (Ref2) - Separation 0.3 - MD 30
- Event plot - Population 4 (Ref2) - Separation 0.3 - MD 30
- Event plot - Population 5 (Ref1) - Separation 0.3 - MD 30
I debarcoded using these standard parameters (minimum separation of 0.3, no filtering by MD) and uploaded the files on Cytobank to inspect (all the 4 debarcoded populations and the unassigned file).
I then proceed to process the files using manually tailored gate up to live siglets as I would if this was not a barcoded experiment.
Below are the CD3 x CD19 plots comparing the proportion of T cells, B cells, non-B non-T cells, and T-B doublets for each PBMC donor:
- Effect of barcode on cell frequency - REF 1
- Effect of barcode on cell frequency - REF 2
Finally I turn myself to analyse the file containing events not assigned to any barcode. Of note, I found could manually gate over 1 million of what look like perfectly good live singlets from the "unassigned" file. Here is the same plot. Please note the similarity in frequency distrubution this file has with the REF 2 replicates. For comparison I also included the same plot on the files containing all barcodes before deconvolution:
- Population frequency in "unassigned" files (compared to barcoded sample prior to deconvolution)
What I believe is happening is that the higher background on 105Pd is preferentially excluding events from barcodes that do not contain 105Pd (BC#2, BC#3, BC#4). What ends up happening is that the CD45_89Y oxide into 105Pd ends up being read as the 4th barcode isotope in those samples and they end up being excluded because they end up have a minimum barcode distance below the set threshold. This does not happen with the samples barcoded with BC#5 because that particular barcode has 105Pd in its combination.
I am obviously concerned with the debarcoding efficiency but also (and more importantly) I am concerned that this is introducing bias in the population frequencies. That is because T and B cells normally have a higher CD45 expression (and consequently a higher background on 105Pd in the case of my experiment) than NK cells. If I am understanding this correctly they (T and B cells) would have shorter BC separations and would therefore be more likely to be left unassigned compared to an NK cell event. This to me would be a very concerning bias to introduce.
Sorry for the long post but I have been giving this some thought for a while. My questions for the community are the following:
1. Does this make sense?
2. Assuming I want to keep CD45_89Y in my panel is there a way to prevent this from happening?
3. Would trying to increase the Pd BC median intensity (by adding the BC overnight with IR intercalation as opposed to 30min on the day of acquisition) potentially help by increasing the separation distance between the true barcode and the 105Pd background?
4. If this cannot be fixed on the experimental side, would it help to be less stringent with the minimum separation parameter and try to filter out unwanted events using Mahlanobis distance?
5. Any other suggestions?
