2.3.1 Magnetic Gating
Method
Magnetic gates move to accommodate the maximum number of events, by relocating gate coordinates to the location of the highest density of events. Magnetic gating has been implemented in the PC platform FlowJo.
Description from FlowJo website:
Draw any gate in the Graph window, and you can make it magnetic by clicking on the appropriate option in the Graph Tools window. Once you make a gate magnetic, FlowJo moves it until it accommodates the maximum number of events -- i.e., it relocates to where the events are. When you copy a magnetic gate from one sample to another, the gate is automatically readjusted. If a parent gate is modified, then the magnetic gate is again recalculated to accommodate the maximum number of events.
This is useful for instances where the population may move slightly from sample to sample, and you don't want to have to individually adjust each sample's gate.
When FlowJo moves a magnetic gate, it draws a black arrow in the graph. The arrow is drawn from where the gate was originally set to where the gate has been moved by FlowJo. If the gate moves only a short distance, you may have to drag aside the white Frequency of Parent label to see the arrow. Gates that move a long distance should be checked for correct positioning!
Magnetic gates are, in every other way, exactly like normal gates: you can create subsets and add statistics to populations based on magnetic gates. They are supported in both univariate (histogram) and bivariate gating tools.
Preliminary Analysis done by Jill Schoenfeld
The goal was to evaluate a subset of the GvHD Use Case Data with Magnetic Gating to find the target population of CD4+ CD8b+ T-cells with this gating structure:
- Lymphocytes (level 1 gate)
- T-cells (level 2 gate)
- CD4+ CD8b+ (level 3 gate)
- T-cells (level 2 gate)
The starting gates were from a manual analysis performed on the GvHD data.
In order to extensively test the magnetic gating algorithm, Jill looked at magnetizing each gate level individually and in combination to see how the magnetic gating functioned at any or all steps in the gating tree. Magnetic gating at each level and each combination results in 8 combinations, as shown in Table 1.
Figure 1. Graph Window images of each gating level before magnetizing the gates.
Figure 2. Graph Window images of each gating level that has been magnetized. Magnetizing the lymphocyte gate causes the histogram of CD3 to show very few CD3 positive cells, indicating that the magnetic lymph gate has not selected the lymphocytes.
This may not always be the case, depending upon sample prep and cell purification, artifacts, and viability of the sample. Magnetic gates will migrate to the closest large population, which may be a set of cells you don't want, such as dead cells or negative cells.
| Magnetized gate level (lymph,tcell,cd4/cd8) |
final pop count (CD3/CD4/CD8b) | Match Ratio compared to expert consensus |
| Lymph | 91 | 0.999166 |
| Tcell | 1 | 0.998948 |
| CD4, CD8b | 737 | 0.945550 |
| (+,+,-) | 91 | 0.999166 |
| (+,+,+) | 385 | 0.986728 |
| (-,+,+) | 737 | 0.944245 |
| (+,-,+) | 385 | 0.986728 |
(-,-,-) |
1 | 0.998948 |
Table 1 Analysis of the final target population of CD4+ CD8b+ T-cells by magnetizing all combinations of the three levels of gates in the gating structure
Even with each gate level magnetized, we still get a match ratio of >0.94 which represents an 736 extra cells in the target population compared to Mario"s original target population of 1 cell. These match ratio calculations used the consensus of all the experts, which show a wide variability. The data is shown on GvHD Preliminary Report and again here in Table 2.
| GvHD1.055 |
Match Ratio | Count | Freq of total | MFI CD4 | MFI CD8b |
| Aaron_1 | 0.99515 | 131 | 0.0095 | 41.27 | 57.14 |
| Aaron_2 | 0.99989 | 0 | 0.0000 | - - | - - |
| Aaron_3 | 0.99575 | 307 | 0.0223 | 18.24 | 48.59 |
| Brinkman | 0.97865 | 627 | 0.0434 | 16.5 | 26.91 |
| Claudio | 0.99986 | 127 | 0.0092 | 13.05 | 55.27 |
| Helene | 0.99898 | 193 | 0.0140 | 14.75 | 63.52 |
| John | 0.99991 | 5 | 0.0004 | 11.17 | 47.73 |
| Junichi | 0.99984 | 143 | 0.0104 | 12.62 | 59.48 |
| Maciej | 0.99990 | 5 | 0.0004 | 12.68 | 41.77 |
| M1 | 0.99993 | 1 | 0.0001 | 14.99 | 70.41 |
| M2 | 0.99993 | 1 | 0.0001 | 14.99 | 70.41 |
| Nick | 0.99974 | 122 | 0.0089 | 11.93 | 68.61 |
| Qianjun | 0.99723 | 296 | 0.0215 | 16.42 | 45.63 |
| Sach | 0.99982 | 51 | 0.0037 | 22.08 | 69.98 |
| Sach2 | 0.99989 | 4 | 0.0003 | 55.95 | 139.13 |
| Starting.wsp | 0.99723 | 296 | 0.0215 | 16.42 | 45.63 |
| mean | 0.99761 | 142.5 | 0.0104 | 19.54 | 60.68 |
| st dev | 0.00531 | 165.5 | 0.0120 | 12.43 | 25.08 |
| CV | 0.00532 | 1.16 | 1.16 | 0.64 | 0.41 |
Table 2 Statistics on CD4+ CD8b+ Population in the same sample for Tree Star Application Scientists
Discussion and Future Work
Preliminary results showed some clear limitations of this supervised classifier, especially the tendency to favor large clusters over small ones, thus choosing statistical significance over biological significance. The solution is to build more of the biological model into the constraints by providing bias more toward the original position (anchoring), to the position of a different gate (tethering). FJML extensions have been made to support tethering gates in the Apple Macintosh version of FlowJo. We still feel that unsupervised classifiers have potential applications in this study, but there needs to be a level of indirection so that small clusters are defined by relative position to the control samples. It appears that magnetic referential gates will be more effective for finding our target populations.