On the right track

The Whitehead Institute’s Center for Genome Research (Cambridge, MA) was involved in aNational Institutes of Health pilot program from 1996 to 1999 to develop methods of high-throughput DNA sequencing for the Human Genome Project.Whitehead’s efforts relied heavily on automation, which provided the speed, precision, reliability, and ease of scale-up that this ever-accelerating project required. Institute scientists and engineers had already builta medium-throughput automated pipeline that processed 4000–5000 samples a day.

However, to feed the more than 100 capillary-based DNA sequence detectors, 50,000–100,000 samples would need to be processed each day. To comfortably accomplish this, a 96-well plate would need to be processed every minute.

The robotics components came from a limited number of sources. Using a small set of suppliers makes maintenance, findingspare parts, and upgrades and modifications easier, but it can also work against you when you are subject to a single vendor’s equipment, prices, timelines, and customer service.

A farewell to arms
We initially planned to use articulated arm (pick-and-place) workstations, whose protocols we had already proven in development and early production. But we soon decided to switch to assembly-line-style automation for several reasons. First, CCS Packard’s (Torrance, CA) version of these stations, which used assembly-line automation, had similar if not better processing times. And we found that the robotic armin the initial automation phase took longer to move all the plates in the protocol than the processing time of the slowest station. By comparison, the track-based systems take seconds to move all the samples. In addition, becausepick-and-place robots move only one plate at a time, the machines must perform two moves to empty and refill samples in a station, whereas a conveyor does it in one step. Although some components on a track system can be more expensive, movement is accomplished with one motor and limited pneumatics rather than with the more complex and expensive multiple-axis robots used in pick-and-place workstations.

Synchronicity
In designing an instrument that performs multistep protocols with a variety of stations—plate movement, pipetting, plate washing, incubation, vacuum filtration, and so on—it is necessary to ensure that the activity at each station takes the same time. This is known as impedence matching, and it may require the assignment of multiple resources to, or the use of several copies of, a given station, to ensure that the tardiness of one station (bottlenecking) does not leave other stations idle for long periods.

For example, our DNA purification protocol requires the addition of two reagents, agitation on a shaker for 4 min, incubation on a magnetic plate for 8 min, and a 3-min wash. The pipettor adds the reagents in well under 1 min, and the gantry moves a plate from the track to the shaker, from the shaker to a magnetic plate, and from the magnetic plate back to the track in less than 1 min. Because the gantry is large and expensive, we did not want more than one on a machine. Therefore, basing our system on a 1-plate/min time frame meant that we needed three plate washers, four shaker positions, and eight magnetic plates. The 12 spots dictated the size of the gantry and told us that one pipettor was enough to dispense the reagents. The tracks employed by CCS Packard eliminated the need to factor plate movement, which is nearly instantaneous, into the design, and secondary functions such as plate stacking and barcode reading are easily performed during the deck’s cycle.

Impedance matching is possible with both conveyors and pick-and-place robots, but there are tradeoffs associated with each. Conveyors eliminate the addition of plate movement and can cover large distances, which allows for easy placement of the required stations and resources. For pick-and-place robots, however, a large coverage area generally means slower processing, more sources of error, and greater cost. It may also be harder to accommodate all the necessary equipment. That being said, retooling a track system is a daunting task, whereas it is easy to add new stations and resources to a pick-and-place unit if space is available, which is useful in prototyping.

Keep it moving
The most difficult problem with certain pick-and-place systems is using racks that have individual plate locations. Unless random access to plates is necessary, these systems should be avoided. This setup usually prohibits the use of continuous feed, or at least makes that type of operation extremely awkward.In one of our medium-throughput systems, continuous feed raised the cost but also reduced the number of locations that one robot needed to be taught from about 60 to 11. Continuous feed also requires a more complex LIMS that can handle either sample transfers with random barcodes or on-the-fly label printing. Collating plates with the same sample name does not work under high-throughput conditions.

The ability to add samples to a deck at any time without having to stop, remove all the finished samples, add new samples, and restart the machine speeds the process considerably.Because continuous-feed processes remove multiple ramp-ups during the day, implementing this concept is especially important for long protocols.

For example, part of our DNA purification protocol requires a 45-min ramp-up—the time before the first completed plate emerges from the system—after which a plate is produced every 1.5 min. Running the deck in 50-plate batches means that in the morning, 45 min is spent ramping up, followed by 75 min of plate processing. On completion, time is spent exchanging old plates for new, followed by a 45-min ramp-up and another 75-min run. In a modest 10-h day, this process is repeated three more times. Wasting about 3 h of process time, the protocol runs at 75% of its potential speed.

Figure 1. Continuous food for thought. Moving from batch processing (red) to a continuous-feed operation (blue) can almost double your output over a 10-h shift. But even with continuous feed, cycle time is the primary concern. Increasing the cycle time from 1.5 min (blue) to 2 min (green) decreases production by 20%.

This process is illustratedin Figure 1, which shows that the length of a protocol is of secondary importance to the cycle time when running continuous-feed equipment. A protocol with no ramp-up and a cycle time of 2 min as opposed to 1.5 min starts losing out after just 2.5 h.

Stay flexible
It is important to ensure that end users can run the decks without constant technical support. This means that either a flexible software environment has to be developed or vendor-supplied software must be extremely carefully considered.After a system is used briefly, many ways in which the user interface, protocol, and efficiency can be improved become obvious. Trying to convince a vendor that these needs are worth implementing is unrealistic. As a rule of thumb, the easier a system is to program, the less flexible it is when alterations are desired. Unfortunately, the more powerful systems require talented people and more time to deliver a finished product.

Benefits of the CCS Packard machines were a flexible software environment and arrays of sensors that signal plate location.State-driven software was the easiest way to write continuous-feed software, in which actions on the decks are the result of the presence or absence of samples. All the robotic actions are not scripted in advance. Previously, plate locations had to be tracked inside the software, but now, sensor feedback ensures that the software and reality match. The software allows users to remove unfit plates, add plates to the machines whenever they arrive, and implement complete error recovery without adverse consequences.

Of course, flexible software is only as good as the hardware that it manages. There needs to be efficient communication between the control software, the status and errors of each hardware component, and the location and status of each sample. Without this information, the software is ignorant of what is going on, which inevitably leads to more frequent mishaps and slower process times. It is also important to remember that robots crash and that some vendors cannot reboot your robots without losing samples, time, data, or all of the above.For protocols that have many plates in play and long ramp-up times, quick and simple error-recovery functions at the end-user level are critical.

Final thoughts
So many factors affect decisions regarding speed that any absolute statements would be absurd.If you have a nicely impedance-matched system, however, one thing is clear: You must get finished samples out of a station and new samples in as quickly as possible. Assuming all stations can act in parallel—an absolute must in high-throughput—productivity is largely governed by the time samples spend at the slowest station and the time it takes to move samples in and out. This gives two main areas of focus: reducing sample movement time and a station’s processing time.

Conveyor-based systems were definitely the right decision for our center. Assembly-line-style automation has been used for years by manufacturers worldwide, and it works just as well in the laboratory.