BioCoder #8

Chapter 4. Flexible Robotic Platforms That Allow Scientists to Remain Scientists

Raghu Machiraju

Biotechnology has not delivered on its promise of breakthrough discoveries of immense benefit to humankind. Laboratory automation that was meant to accelerate the pace of discovery is mostly inaccessible and unusable for most experimental biologists. Current solutions are expensive and only offer improvements in throughput of specific steps in a given experimental protocol.

In this article, we argue that progress in biotechnology will come from the use of open user interfaces and open-specification middleware to drive and operate flexible robotic platforms. Such middleware will accelerate and scale the plan-execute-analyze cycle of development in the biosciences. It will also allow one to integrate multiple kinds of robotic platforms, 3D printers, and sensors, and drive these systems through biologically meaningful user interfaces. Further, the availability of such middleware will provide reasonable-cost solutions rather than the current expensive offerings. Both tinkerers and mission-oriented researchers can then create or acquire affordable and custom robotic platforms and explore the marvels of all things biological or complete specific tasks, respectively. Most importantly, such platforms will allow scientists remain scientists and not computer technicians.

In this article, we first describe the state of affairs in biotechnology and then discuss aBioBot’s solution, which provides flexible, low-cost, ergonomic experimentation while shortening the time to discovery.

The Unfulfilled Promise of Biotechnology

Although the potential of biotechnology is often touted, the industry only recently saw a scintilla of exponential growth scales that consumer electronics and personal computing have enjoyed in decades past. The emergence of genome sequencing that blows past Moore’s law spurred some of this growth. High-throughput sequencing and phenotypic screening generated a prodigious amount of data, which is not often well interpreted. There are urgent needs to test and utilize the data produced. Although this technology and its promise led to strong sales within the biopharma industry, the technology has not yet led to the expected mass-market adoption or the creation of new industries. Consequently, there is very little room for the average biotech tinkerer or the serious molecular biologist to create novel consumer technologies or services.

Although some processes in biotechnology, such as manufacturing and instrumentation, have matured, the innovation and discovery process remains archaically manual for all but the most sophisticated and resource-rich companies.  It is the hearth of biotechnology, the wet laboratory, that needs and deserves a close second look.

Impediments to Wet-Lab Productivity

For molecular biology and medicinal chemistry, the vast majority of innovation is still a largely manual process performed at the bench in a wet laboratory. Typical metabolic engineering and systems biology problems have high degrees of freedom (e.g., many genes in a synthetic pathway), leading to heterogeneous responses to perturbations. Biological science is inherently iterative—composed of several rounds of plan-execute-analyze steps. The plan-execute-analyze cycle is the norm in science and engineering. The laboratory scientist plans by choosing a protocol, then executes the protocol in the wet laboratory, and analyzes the resulting data from experiments. These steps are repeated many times until the desired results are obtained and adequate insights are gleaned. Therefore, in the absence of automation, a significant biotech project is often measured in person-decades. There are additional costs; for many mundane and cumbersome procedures, the resulting errors lead to expensive downstream protocols. Most importantly, irreproducible research is often the result.

Crucial obstacles to the widespread adoption of biotechnology are the elimination of manual processes, commoditization of methods and machines to perform them, and an open structure of the processes and assembly across the industry. Consider this: the personal computing industry arose from the commoditization of the von Neumann architecture in the form of a usable, albeit adaptable, hardware layout and an operating interface. This early, open architecture and other contextual changes (e.g., the growth of the Internet, the use of tangible user interfaces) allowed for the growth of the personal and iComputing industry. The open community of innovators and the ecosystem it spawned also ushered in the age of personal computing. Similarly, it can be argued that widespread adoption of bioinformatics also arose from the growth of open-access tools and environments including BLAST and the R language. Further, the open community of users and data analysts who shared their trials and tribulations of processing a certain sequencing data. The need for open interfaces and architecture and thriving open community is a requirement in the biotech industry.

Let’s now examine the state of the automation in the wet lab; we will undoubtedly learn that the industry offers a closed and proprietary approach.

Automation for the Wet Lab

Technologies and processes exist that significantly multiply a bench scientist’s productivity, but they are not accessible. As they exist, lab robots produce highly reproducible work and can potentially improve productivity by several factors for encodable tasks. There are many examples where automation has positively contributed to expedient and robust treatments and procedures. For instance, through shallow sequencing and highly optimized automation, the painful amniocentesis procedure can be replaced by noninvasive prenatal tests (NIPT), molecular profiling of the mother’s blood.

Commercially available laboratory robotics platforms at best occupy a small niche in the market. Some of them deliver very high throughput for expedited execution. However, high throughput does not translate into expedited discoveries, because the other two stages of planning and analyzing are still left to manual means.

Existing solutions are often unsafe to operate, overtly expensive (some in excess of $300,000), operate in limited orthogonal Cartesian coordinate systems and are encased in rigid steel truss, thus proportionally increasing the cost of ancillary equipment such as well plates and temperature sensors. Further, they require very proprietary equipment for manufacturing, such as trademarked pipette tips. Lastly, the flexibility in a typical lab cannot be replicated on the robotic bench. A variety of tubes and well plates cannot be used on a typical robot and can be placed in any orientation. A change in protocol requires extensive re-programming of the various material handling procedures. It is not common to have automation engineers in well-heeled institutions that are dedicated to these tasks.

To reiterate, we stridently believe that this state of affairs exists because of the closed nature of existing biotech enterprise. Clever innovators cannot easily incorporate their ideas into existing robotic frameworks. There is no common and tangible interface that is accessible to one and all. Even more importantly, the apparatus has very little self-awareness and cannot be taught to be aware of its configuration and capabilities. Given the plethora of degrees of freedom in space, time, and function, the latter ability is very important for biotech to be a viable and tangible possibility. Still, is the situation really that dismal?

Emerging Solutions

New efforts have emerged to mitigate the situation. Emerald Cloud Lab, Transcriptic, SyntheGo, and Arcturus BioCloud are companies scaling laboratory automation through services in the cloud. They automated experimental design as a service, which has been well received. However, they still do not offer the flexibility and direct touch and feel of the bench as desired and needed by the typical investigator. To their credit, the workflows offered by these companies improve the planning and execution stages of the iterative scientific development cycle as it pertains to selected protocols.

Another set of companies and start-ups (Modular Science, iorobotics, and OpenTrons, to name a few) are experimenting with lab-automation paradigms that share our vision. They adopt simple, inexpensive, open-design and extensible hardware platforms. Still, they offer limited capabilities to improve the iterative plan-execute-analyze development cycle. The software infrastructure to plan the protocol and execute it efficiently on the robotic platform is not available. There are novel academic solutions that have been recently proposed. The Riedel-Kruse laboratory at Stanford University offers an open source biotec processing unit (BPU) with a tangible user interface and highways to the computing cloud.

To chart a useful path, you must understand the essential computational tasks of wet lab procedures and choose those that are encodable and can be automated to accelerate the pace of discovery.

Wet Lab Work Is Encodable

Each protocol can then be represented either with a collection of natural language statements or as a dataflow process graph. In either case, protocols are a set of computations and translators that can be built between the two representations. There are at least two sets of nodes; one set represents materials, and the other set describes operations (e.g., mixing) resulting in creation of new or intermediate materials (see Figure 4-1). The edges indicate the transport of material, and in this case, describe the material mixing for preparing the master mix for polymerase chain reaction, or PCR. The dataflow graph can be translated, in turn, into G-code or CNC-code that can drive a robotic head. However, for this to occur, various receptacles or wells have to be bound to different reagents. The dataflow process graph, when annotated with spatial locations can be examined for optimal path planning and can include constraints of containment. We now describe the visual sensing module and the role it can play.

Figure 4-1. A data flow graph for preparing master mix for PCR; each node is material and can also implicitly define an operator (mixing in this case)

Visual Sensing and Feedback

The breakthrough we seek initially is to equip the robot with visual sensing capabilities and thus achieve visual servo control of the robots. The rewards are many. Because the sensors can identify all objects on the bench, it is easy to plan robotic operations there. The binding of bench objects to material nodes in the dataflow graph is achieved through the culmination of analysis of acquired data. An additional operational benefit is that the machine can be trained to compensate for small and large distortions in the work area. Visual confirmation of the work surface will allow the robot to be safer (i.e., it will be able to sense hands and obstructions) as well as less error-prone (i.e., it will automatically be able to sense if a well has liquid in it, whether a tip is missing from the box, or whether colonies or contamination exist on a plate). Further, when it rains pipette tips, it will be useful for the operator to be informed. We firmly believe that our approach of using sensing in a middleware parlance will allow for the creation of automated workflows that will be safe, reliable, accurate, and cost effective.

Middleware Is Essential

We believe that solutions lie in the availability of middleware interfaces that will allow for the planning and completion of a large number of encodable experiments. The middleware layer will then communicate with devices (robots, sensors, etc.) and consummate precisely routine and encodable tasks that would be otherwise completed manually and erroneously.

Consider this: a user interface can serve as the command and control of the experiment. Protocols in the form of annotations will be listed and annotated for materials and processes. Typical protocols include the preparation of master mix for PCR (as described earlier), DNA extraction, ELISA assay, and serial dilutions. Using either automated or manual (through annotation) means, a dataflow process graph can be extracted from the text-like description and displayed. The video from a camera will be captured and analyzed for the presence and location of well plates and other equipment. This information will be delivered to the user interface, which will be further used to plan experiments; the user can assign various reagents to well plates and even more specifically list the various sources and destinations for each reagent of the protocol. Once all assignments have been done, the user can execute the protocol. Appropriate commands (G-code) will be generated and communicated to the printer. As the robot head completes each of the protocol, it can communicate with the executive middleware, which then logs the progress and also updates the process dataflow graph.

The plug-and-play approach fostered by the middleware will allow the precise completion of ordinary tasks at reasonably lower costs with a simple and accessible user interface. Our reliance on middleware and scripting languages also allows expansion of the platform by users and other companies and the customized generation of special control work surfaces for tablets and customized workflows. New pieces such as plate loaders, refrigeration blocks, and colony pickers/counters can be designed by third parties or users and integrated for use with the middleware. We now describe our platform.

The aBioBot Solution

At aBioBot, we wish to re-create the flexibility of the familiar wet laboratory bench on a robotic platform. Our lab assistant is a robotic platform combining state-of-the-art hardware and software:

LabBench

A web browser user interface to facilitate protocol authoring, observe the layout on the bench, monitor the progress of the experiment, and log the protocol on the cloud. The user interface of LabBench is realized in HTML5 and Javascript, and leverages toolkits, including tornado and bootstrap.

Yan = Eyes

In our platform, Yan is machine vision that watches over the experiment for you. Yan’s software module learns the layout of wells and provides surveillance of the bench for untoward accidents. Yan is implemented in Python and uses the OpenCV library.

The Bot

Our robotic platform is modified from a 3D printer. We also offer a series of adapters that allow The Bot to use standard pipettes and equipment. As of now, there exists an adapter that allows the use of a standard Eppendorf pipette. The Bot is derived from open source 3D printer hardware and rapid prototyping machinery.

Although aBioBot is built on open source hardware and software, both LabBench and Yan will be accessible through open APIs and extended as necessary. Lab Bench has two functional components accessible to users:

• Every lab staff member has a book of her favorite protocols, which have been refined and improved over time. LabSmith is a protocol-authoring tool with the capability to import lab procedures and notebooks from various repositories, including OpenWetWare and protocols.io.  Most importantly, it will also allow for protocols to be changed and adopted as required.
• Every log, video, and status report for your experiment is automatically uploaded to the cloud through LabCloud.

The Path Forward

What we have accomplished over the last three months is the construction of an efficient middleware and a functional desktop prototype. We are currently testing the platform on preparing a master mix to prepare for PCR. Also, we plan to release an open API that will allow for sensors and any 3D printer platform, which will be treated as a device. We are considering the implementation of our software on both desktop and lab bench-sized hardware platforms. We believe that this is the way forward for biotech and the creation of this middleware and robotic platform will be the first steps in creating unexpected multibillion dollar new industries that harness the power of industrialized and roboticized biotechnology.

Keep watching out for announcements at the aBioBot website.

Figure 4-2. The aBioBiot prototype repurposed from a 3D printer. The black-colored trays hold required input and output material, while the blue-colored boxes hold unused and use pipette tips. Finally, the pipette is engineered to replace the printhead.

Raghu Machiraju is a cofounder and CEO of aBioBot and a professor of computer science and engineering at Ohio State University. The mission of aBioBot is to build flexible robotic platforms that will help scientists make breakthroughs by shortening the time to discovery. Raghu was aptly helped in the writing of this article by Ron Shigeta of IndieBio and his co-founders Chaitanya Kulkarni, Kun Huang, Ken Rinaldo, and Trademark Gunderson.