Can You Trust Your Robot?

While robotic liquid handlers have allowed increased testing capabilities in the laboratory, concerns over protocol, accuracy and suitability still remain.

iRobotLaboratories around the world have turned to liquid handling robots to automate their assays. Science fiction writers and fans of Isaac Asimov’s I, Robot have wondered for years what will happen the day robots start misbehaving—but could that day be here already?

The introduction of robotic liquid handlers into biology and chemistry laboratories in the past 20 years has arguably led to significant advances in our scientific understanding by increasing our testing capabilities. But these robots also bring a set of commonly overlooked challenges, in particular, the question of whether the robot is doing what it is “supposed” to be doing.

In Asimov’s fictional future, sentient robots are kept in check using the “Three Laws of Robotics.” However, laboratory robots have no way of determining for themselves whether their behavior is suitable or not. This problem is exacerbated as many operators do not have access to the measurement systems they would need to determine the suitability of their robots’ performance.

Despite this lack of performance monitoring ability, it is far too easy to trust that the robots working away in the lab are behaving within specification, especially if they are serviced regularly. Yet those service calls generally do not check if the robot is dispensing all target volumes accurately and precisely, nor do they attempt to optimize and reset the protocol parameters of the robot for proper performance. Far too often, the volumes delivered by a robotic liquid handler are assumed to be accurate and precise; yet, without scientific evidence, how can we be sure?

Gravimetric analysis with an analytical balance is often used to measure the weight of liquid dispensed and therefore, by inference, the liquid volume. However, environmental factors that cause evapora- tion and balance drift reduce its effectiveness when working with the low microliter and nanoliter quantities typically delivered using robotic instruments. While the average accuracy across multiple dispensing channels can be determined, it is difficult and often prohibitively time consuming to determine the accuracy and precision of each channel of the robot.

In contrast, ratiometric photometry provides a simple and effective method for determining the volume dispensed by each channel into the individual wells of a microplate.

table1Table 1: Details of the sequential parameter adjustments.

Ratiometric photometry as applied by the Artel MVS Multichannel Verification System uses a dual-dye, dual-wavelength absorbance method to determine the volume of liquid dispensed in a single measurement. In practice, a sample (containing both a red and blue dye manufactured under ultra-tight tolerances) is dispensed into each dimensionally characterized well of a microtiter plate. The wells are then filled with a solution contain- ing blue diluent dye at the same concentration as that dispensed in the sample. The volume of sample solution dispensed into each well is then calculated using an innovative application of the Beer-Lambert law, and the accuracy and precision of each of the robot’s channels can be determined for specific target volumes and/or volume ranges.

The devil is in the details

To demonstrate just how sensitive automated liquid handlers can be to the protocols they are programmed with, a liquid handler with a 96-channel high-volume head (HVH) was used to dispense a target volume of 10 µL into a 96- well microplate, and the MVS was used to measure the actual volumes dispensed. The initial protocol used for Experiment A utilized brand new, un-wetted pipette tips and the various aspiration and dispensing parameters shown in Table 1.

By varying the parameters listed in Table 1 in a sequential manner, Artel scientists optimized the dispensing protocol such that both the relative inaccuracy and %CV were minimized, as can be seen from the data in Figure 1. In this experiment, two of the most crucial factors that needed to be adjusted were the leading air gap (a volume of air in the pipette tip prior to liquid aspiration, that helps to dispense the entire volume of liquid aspirated) and the target volume itself. What is most striking here is that while the coefficient of variation was consistently less than 5%, the relative inaccuracy was as high as 30%—meaning that during Experi- ment D, just 6.8 µL rather than 10 µL was being dispensed. Adding just 6.8 µL instead of 10 µL could have a staggering effect on assay results because the actual concentration of species in the  solution could be 30% lower than expected. This high level of data inaccuracy could cause scientists and clinicians to come to erroneous conclusions about what the assay results mean. At the very least, this could cause embarrassment when results are not verifiable using other techniques or in other laboratories.

figure1Figure 1: Graphical representation of the measured volume transfer performance for each experiment, where the accuracy is variable (>30%) and CV is not (<5%).

The above results highlight the need to adjust and control each aspect of a robot’s innate “personality” if it is to behave as requested. However, even when a method has been properly calibrated and verified, there are limits to the situations it can be trusted to perform. No matter how finely adapted a protocol is for a particular liquid handler, a key question that should be asked is “over what volume range will it be applicable?” It is important to recognize that whatever the robot may be, and whatever protocol may be running, there will be a volume limit beyond which a given set of parameters no longer apply.

To emphasize this point, Artel scientists used a single-channel liquid handler that allowed very little “low- level” programming and set it to work dispensing a range of volumes from 200 to 2 µL. The robot took aim at each target volume eight times and the results were analyzed using the MVS. The results of the study show that while the test protocol performs very well from 200 to 20 µL, it fails when used for delivery at 2 µL.

Not all robots are equal

Although humans think of them-selves as individuals, we do not commonly consider liquid handling robots in the same way—a tool is a tool is a tool, no matter how complex it is. Modern laboratories use different robotic platforms in different locations and often assume that the data generated by a robot in one location is directly comparable to the results generated on a robot of the same make and model running the exact same methods in a different location with different environmental conditions.

Even when laboratories use the same sets of liquid class parameters on the same make and model of automated liquid handling instruments, the amount of liquid dispensed can vary immensely—and this can lead to inaccuracy and imprecision from test-to-test or site- to-site.

To demonstrate how large the variation between identical liquid handlers using the same methods can be, researchers used the MVS to study the volume transfer performance of three robots of the same make and model at three different forensics laboratories using three different pipetting heads. The methods employed had been developed at one site and then rolled out to the other sites for implementation.

The results of the study can be found in Table 2. Even though the methods, labware and instruments were all the same, the variation in the volume transfer performance was staggering. For instance, with pipetting head 1, the %CV values varied by as much as 18%; while with pipetting head 2, the difference in relative inaccuracy between locations 2 and 3 amounted to almost 29% (+6.15% and -22.80%).

I, Verify

While Asimov predicted in I, Robot that only the “Three Laws of Robotics” would be needed to ensure that humans would be safe from a sentient robot, this article has shown that liquid handling robots are unable to obey “laws” as they do not have the ability to monitor their own performance. In the absence of such in-built performance monitoring, external performance verification needs to be performed if the assay data the robots generate is to be trustworthy.

table2 Table 2: Site-to-site comparisons.

Dispensing protocols can be finely tuned such that for a given volume range, the robot will accurately and precisely deliver the volumes of liquids it has been programmed to dispense. However, there are limits to a robot’s suitability for a given task and it is important to recognize where those limits are so the robot can be kept in check.

What’s more, even if robotic liquid handling platforms of the same make and model are used in different locations, they will each  have their own personalities, and protocols need to be optimized for each individual robot to account for their idiosyncrasies. By optimizing each protocol individually using robust measurement technology, researchers can rest assured that their reputations will be safe in the hands of the robots they rely on to generate their data.

About the AuthorTanya Knaide

Tanya Knaide

Tanya Knaide is a scientist with over 10 years of experience in leading new product development projects, product launch campaigns and uncovering customer needs to develop innovative new products and services to satisfy them. As Product Manager at Artel, Tanya has led cross-functional and inter-organizational teams that span across R&D, engineering and marketing and ensured that development and marketing projects deliver benefits to the customer in a timely manner.