The authors have declared that no competing interests exist.
Conceived and designed the experiments: JBN SJM. Performed the experiments: JBN. Analyzed the data: JBN SJM. Contributed reagents/materials/analysis tools: JBN. Wrote the paper: JBN SJM.
Although
Cells have been historically studied with population-level measurements, on the assumption that an individual cell’s phenotype is well-described by the population average. However, recent studies have shown that considerable variations in mRNA levels, protein levels, doubling time, and size exist between isogenic cells
Several approaches exist for measuring growth of microbial cells
All of these approaches are population level measurements and thus do not allow assessment of single cell parameters. A second drawback is the fact that neither morphological nor phenotypic changes over a single lineage can be assessed. Although the environment can be perturbed in these assays, transient changes are not feasible. A method capable of following thousands of continuously dividing cells over extended periods of time would be extremely useful for deriving accurate growth rate measurements, and to link single cell phenotypes with cell cycle.
Several attempts have been made to develop tools for single-cell measurements. One simple tool to measure physiological parameters of single microbial cells is the agar pad
Microfluidic devices allow for the precise handling of fluids, and thus have become popular approaches for conducting single cell studies
A different approach is to trap individual cells while daughter cells are removed, which permits long-term tracking of a few cells over many generations. Such approaches were developed for
We have developed a microfluidic system that alleviates these issues. We show here that our microfluidic platform can quantitate over 100000 divisions in a single experiment, track cells over multiple generations, and reconstruct lineages.
We designed a microfluidic device consisting of 120 microchemostat growth chambers based on a design previously applied to large-scale single cell studies in
(A) Schematic of the
(A) A flow diagram showing all analysis steps of the image processing pipeline. First, each growth chamber is divided into lanes as defined by the microfluidic ‘highway’ structures. Next, individual cells are segmented (B) and tracked using either a local (C) or a global (D) tracker. All scale bars are 10 μm.
On-chip cell growth rate was slightly slower compared to bulk culture, with the average on-chip doubling time of 128±2 minutes being 10% higher than that measured for a standard liquid culture (115±2 minutes). These differences are comparable to differences previously observed for
The primary objective for our microfluidic platform was to track thousands of cells through at least one complete cell division cycle to obtain precise growth rate measurements. A secondary objective was to track cells over several generations to construct lineage trees. Analyzing micro-colonies is the most effective way of generating complete lineage trees
(A) A five generation lineage tree reconstructed by our image processing pipeline. (B) The number of lineages that could be traced over a given number of generations during a single experiment covering 50 hours of cell growth. (C) The number of complete lineage trees 1, 2, or 3 generations in length. (D) A schematic describing our definition of percent completeness of a lineage tree and the corresponding average completeness of lineage trees of a given number of generations in length (E).
A second metric useful in describing the performance of a cell-tracking platform is the number of complete trees and the completeness of trees of a given number of generations. Obtaining complete lineage trees is challenging, even for micro-colony approaches. The primary reason is the accuracy of the segmentation and tracking algorithms. We were able to obtain 58 complete trees 3 generations in length, and 2623 complete trees encompassing 2 generations (
(A) Example of a single cell growth process and definition of the various parameters the image processing pipeline returns including: birth length, division length, fission length, elongation time, septation time, and elongation rate. Histograms of (B) single cell birth, division, and fission lengths and (C) doubling, elongation, and septation times before and after a temperature shift from 30°C to 25°C.
We used these parameters to determine the effect of changes in temperature on cell growth and morphology. Cells were initially grown at 30°C and then subjected to a temperature decrease to 25°C (
In the previous section we show the steady-state distributions of various parameters pre- and post-temperature shift. It is not known how cells that are exposed to a temperature shift during different phases of their cell cycle dynamically adjust these parameters. Since we observed a large population of unsynchronized cells, and our temperature shift occurs relatively quickly (with a transition (rise/fall) time of 40 minutes (80%) and 80 min (90%)), we could plot these parameters over time and in relation to the cell-cycle (
Each data point is a single cell division event, and the color indicates the percentage of the division time that has elapsed after the temperature shift. The generation dividing during the temperature shift and the first generation after the temperature shift are shown. Black lines indicate the moving average and the red lines the standard deviation. (A) Division lengths, (B) birth lengths, (C) number of divisions per time bin, (D) doubling times, (E) elongation times, and (F) septation times are shown.
Although cell size at birth and size at division are extremely robust as measured in steady-state cultures, we observed that these two parameters do in fact transiently vary when the temperature shift occurred at a specific time relative to the cell division cycle. Cell division length increased from the nominal length of 12.9 μm to 13.8 μm. Cells that spent up to 50% of their division cycle (the latter half of their cell-cycle) at the new, lower temperature divided at the nominal length. Cells that spent over 50% of their cell-cycle at the lower temperature divided at increasingly larger sizes, reaching a maximum of 13.8 μm if they spent 60% of their division cycle at the new temperature. If cells spent above 60% of their cell cycle at the new temperature, cell division length began to normalize until cells again divided at the nominal length if they spent their entire cell division cycle at the new temperature. This indicates that there is a temperature sensitive phase in the cell-cycle which does lead to abnormally long cells at division (around 60–70% into the cell cycle). Hence a temperature sensitive process that regulates cell-division length appears to be active during this part of the cell-cycle and possibly remains active till division. Those cells that divide at abnormally large cell-size, give rise to abnormally large daughter cells (
We could also discern time-dependent fine-structures in the doubling, elongation, and plateau times. As discussed previously these times increase proportionally upon a 5°C temperature shift. Doubling time begins to increase immediately upon temperature shift; cells that only spent a small fraction of their cell-cycle at the lower temperature already exhibit an increased doubling time. We also observed a pronounced and transient decrease in doubling time in the second generation after the temperature shift. This decrease in doubling time coincides with the slight increase in birth length also observed in the second generation after the temperature shift. This suggests that cells normalize larger birth length by decreasing the doubling time to consequently divide at the nominal length of 12.9 μm. This temporary decrease in doubling time arises primarily from a decrease in elongation time but not septation time (
Cell length is well controlled in
(A) Box plots of single cell division lengths for cells with a given birth length. (B) The average values of the box plots shown on a scatter plot, with a linear regression fit to the data ranging from 6–10 μm birth lengths. (C) The raw data shown in (A–B). The color of each data point represents the number of single cell division per point. (D–G) Data as in (C) but each data point is now colored by (D) doubling time, (E) elongation time, (F) septation time, and (G) elongation rate.
Sister pair data points are in both cases colored according to mother division length.
Cells born with a normal cell length (∼7.3 μm) divide at lengths that span the entire range of observed cell sizes (
Decreased doubling time normalizes the division length of cells. Doubling time becomes progressively shorter for cells with increased birth length (
Comparing the doubling times for sisters and division length of the mother shows that a direct link exists between mother cell length and daughter cell doubling times, as stipulated above (
We developed a microfluidic microchemostat array for the long-term culturing and interrogation of continuously growing
With the current combination of microfluidic hardware and image processing software we showed that we could analyze over 100000 division events in a single experiment, and follow thousands of cells over 7 generations (
Strain WT 972 h- was used in all experiments. Cells were grown overnight at 30°C in 3 ml YES (Difco) 2% glucose with shaking in classical batch culture prior to loading onto the chip. On the device cells were also grown in YES 2% glucose.
The chip comprises 8 rows of 15 chambers for a total of 120 chambers (
Each growth chamber is controlled by 2 valves and can be divided into three zones (
Mold fabrication was performed in the clean room facility of EPFL (CMI). Chip designs were drawn using Clewin (WieWeb). The flow wafer consists of 3 successive photoresist layers of different heights. First, a layer of 1040 SU-8 (Gersteltec) was spin coated on a silicon wafer to 1.5 ∼ μm to generate the microfluidic sieves. On this was spin coated a layer of 1050 SU-8 2.5 μm to generate the growth chambers. Each layer was exposed separately. The resulting wafer containing both exposed SU-8 layers was developed twice for 10 minutes in 2-methoxy-1-methylethyl acetate (PGMA). The medium supply channels were generated with a single layer of AZ9260 14 μm. The wafer was then processed at 180°C for 2 hours to round the AZ9260 channels. The control layer is made of a single layer of 1050 SU-8 20 μm exposed and developed using standard techniques. The control layer mask is scaled to 101.5% in order to compensate for PDMS shrinkage during the curing process.
Microfluidic devices were fabricated from polydimethylsiloxane (PDMS) (SYLGARD 184) using standard soft lithography techniques
All control lines were primed with water prior to loading the flow layer with medium, except for the chamber outlet valve, which was primed later on. The chip was loaded with medium and cells were seeded in each row.
The device was imaged on a Nikon Ti-E Eclipse and a 40X SPlan Fluor ELWD objective. The microscope was enclosed in a temperature controlled environment (Life Imaging Services). Images were acquired with a Pike F145C IRF16 (Allied Vision Technologies GmbH). All microscope automation, microfluidic device actuation, and imaging were performed with a custom written visual basic program.
All experiments were conducted with the same initial conditions. Cells were imaged for 10–12 hours at 30°C in YES + 2% glucose with 3 psi pressure driving the medium flow. After 10–12 hours we shifted to a new temperature. Temperature was continuously recorded by a thermocouple embedded in a PDMS block with the same dimensions as the microfluidic chip and placed next to the device in order to give an accurate temperature reading. The recorded temperature fluctuations were on the order of ±0.1°C.
The analysis pipeline is described in
Tracking and lineage reconstruction was performed in 4 successive phases: i) frame by frame cell matching, ii) matrix reconstruction, iii) division detection, and iv) lineage reconstruction. The first phase matched cells frame by frame. To optimize computational speed we first applied a local tracker, which, if it failed due to large cell displacement, was substituted by a global tracker.
Local alignment (
All cells from frame n were scored to cells in frame n+1. Cells that moved further than a single cell length (50 pixels or 11.52 μm) were scored as
If this local pairing left too many cells unpaired (>30), or the average displacement was bigger than an arbitrary threshold, results were cleared and a frame-by-frame adaptation of the global assignment process proposed by
For every lane in a chamber, cells close to the sieve move slower than cells near the outlet. Therefore, we impose minimal displacement constraints on cells depending on their position. Cell pairs failing to achieve this displacement are discarded by the algorithm. If too few cells are paired and the average displacement of paired cells is above 8 pixels (1.85 μm), we reapplied the tracker with a larger minimal displacement. Unpaired cells were then checked as for the local tracker.
In the second phase we constructed a matrix of all paired cells for every frame as defined by both trackers. We detected divisions in the third phase. We generate a lengthwise intensity profile for each cell and convolve it with a Mexican hat function (
In the fourth phase we reconstruct lineages using cell division information. Cell lineages contain information on individual cell profiles, doubling time, and length at birth. By detecting the time at which growth ceases, we define a plateau, which is the division length. A linear fit of the growth prior to the plateau allows us to obtain the elongation rate.
We estimated the accuracy of our automated analysis by comparing the results of an automated analysis with a manual analysis of the tracked cells from 500 frames. 621 divisions were observed manually and 614 by the software. 7 divisions were missed, among them 2 were reseeded automatically. The remaining 5 errors were either filtered downstream (2) or were lost over the tracking process later on (3). The more critical errors were, 2 divisions that were called too early (1 was filtered out, 1 was lost), 8 cells were wrongly assigned but automatically corrected while tracking and 26 wrongly assigned cells (twice the same cell) but were not automatically corrected. Out of those 26 wrong assignments, 18 were filtered out and 7 were lost in the following steps of the tracking process. In the end only one mistake was actually retained in the results. In summary, out of 614 observed divisions, we obtained 342 paired cells, among them 195 were generation 1, 110 generation 2, and 37 generation 3. After the filtering steps 268 divisions remained. Thus, our algorithm made one error out of 268 divisions, or 0.4%.
(TIF)
(TIF)
(TIF)
(TIF)
(MOV)
(MOV)