Automation has the potential to be the fifth paradigm of science by integrating data- and model-driven approaches, eventually shifting the onus of discovery from human to machine. It is the key to improving reproducibility, productivity, and throughput of research in life sciences, as well as eliminating labor-intensive workflows. However, true automation is yet to be realized, as current commercially available robotic platforms still require laboratory personnel to set up protocols, monitor progress, and move samples between stages in the workflow. Total automation could offer researchers remote access to laboratory systems to carry out experiments, execute actions, and automatically acquire data through cloud-based applications.
The future of life science research
A collaborative research project from RIKEN, the National Institute of Advanced Industrial Science and Technology (AIST), University of Tsukuba, University of Tokyo, and Keio University (funded by JST MIRAI) has successfully developed an autonomous cell culture system that can be operated remotely using artificial intelligence (AI) and LabDroid, a versatile humanoid robot. To do this, the team effectively overcame issues with fixed robotic platforms, programming languages, and computing tacit knowledge.
Robotic crowd biology
Laboratory automation is not new, and remote access to automation systems is becoming increasingly common. However, these workflows are still restricted by the use of static and isolated machinery, limiting the ability to move seamlessly between stages in a pathway. To overcome this, the project researchers have developed a high performance, dexterous LabDroid robot nicknamed Maholo, which consists of a torso pivot and two robotic arms that
can rotate on seven axes to manipulate laboratory tools for everyday tasks, such as liquid handling for cell harvesting.1 Maholo is not restricted to a specific job, and can be used to operate different laboratory instruments for various life sciences applications. This offers the potential for life sciences to step into the realm of robotic crowd biology, with a whole team of remotely activated humanoid robots able to manipulate an entire laboratory of equipment and tools in unison.
Talking to robots
For robotic biology to become a reality, machines—including LabDroids—would need to “talk” to one another to initiate the next stage in the sequence and conduct different protocols simultaneously. However, programming platforms for life sciences is not straightforward, and very few robots are amenable to the frequent protocol changes that naturally occur in research. In addition, many systems operate using their own unique programming language that prevents them from communicating effectively with other systems or understanding commands. The JST MIRAI Project has been developing a formal experimental protocol description language, called LabCode, to address this. The system interprets a command and uses a protocol description language that automatically generates programs that can be understood by different devices from various vendors, each using separate transcription languages. This allow the machines to be connected to one network, and interpret the same command codes. LabCode has been used to successfully perform an experimental protocol for genome editing at the University of Tokyo, consisting of cell seeding followed by DNA extraction and barcoding. The workflow featured a LabDroid and a Tecan Freedom EVO® liquid handling platform, and LabCode was used to describe the environmental definitions to the compiler to delegate tasks. The compiler generates the instructions for the human operators, LabDroid and the Freedom EVO, automatically distributing the workload and establishing the optimal settings.
Applying tacit knowledge to machines
Every year, potentially tens of thousands of different protocols are published that are unclear, not quantitative, or missing key information that is required for programming. This is often “tacit knowledge” that biological researchers may not even be aware of, and are unable to express to others—particularly machines—creating an education bottleneck. Tacit knowledge is all the information that researchers unknowingly accumulate with experience. The ability to effectively differentiate between different cell states in culture is an example of this, and is an important skill that is challenging to teach robots.
To overcome this, the research team worked closely with laboratory experts to implement their cell evaluation criteria into a novel AI technique that can replicate the decision-making of experienced scientists. This machine learning technique was combined with LabDroid for the autonomous induction of clinical-grade retinal pigment epithelial cells from pluripotent stem cells, a highly skill- and experience-dependent process. Using a Bayesian optimization technique, this robotic-AI system was able to determine the optimal parameter sets for differentiation and high cell viability. From a total of 200 million possible parameter combinations, the system optimized conditions to produce yields with pigmented scores 88 percent higher than pre-optimized conditions. This entire process would have taken up to five years to complete, but with automation, the results were gathered in just a few months, showing a 10-fold acceleration in research time.3
Autonomous cell culture
The efficient production and culture of high-quality, viable cells is an important stage of all cell-based applications, such as regenerative medicine and cell therapy. Autonomous cell culture systems can help standardize the quality and reproducibility of cells by removing the risk of human error. The JST MIRAI Project research team applied this robotic-AI approach to develop an autonomous, variable scheduling cell culture system. The closed-loop system, which included LabDroid and a newly developed AI technique, continually produced subcultures of human embryonic kidney cell lines (HEK293A) to determine the optimum workflow and the best time for passage.
The cycle begins with the robot passaging the cells, then periodically observing the culture plates with a microscope, using machine learning to recognize cell states. Using this data, a model-based inference system predicts the cell growth curve, and then determines the nature and timing of the next step, and delegates the relevant tasks to the robot, completing the full cycle. This system was able to run for 10 days completely unaided,2 and this lack of human intervention proved extremely beneficial when the laboratory had to close temporarily for COVID-19 lockdown restrictions, enabling researchers to continue working remotely, and demonstrating the potential of this approach for the future of life sciences research.
Automation is the key to achieving reproducible, high-throughput workflows for a number of life sciences applications. With anthropological advances in robotics and AI—including overcoming the challenges of communication, movement, and implementing tacit knowledge—machines are able to replicate the work of scientists, placing automation at the forefront of the future of research.
1 – Yachie, N., Robotic Biology Consortium., Takahashi, K. et al. Robotic crowd biology with Maholo LabDroids. Nat Biotechnol 35, 310–312 (2017). https://doi.org/10.1038/nbt.3758
2 – Ochiai, K., Motozawa, N., Terada, M., Horinouchi, T., Masuda, T., Kudo, T., ... & Takahashi, K. (2020). A Variable Scheduling Maintenance Culture Platform for Mammalian Cells. SLAS TECHNOLOGY: Translating Life Sciences Innovation, 2472630320972109.
3 – Kanda, G. N et al. Robotic search for optimal cell culture in regenerative medicine. (2020) bioRxiv 2020.11.25.392936; doi: https://doi.org/10.1101/2020.11.25.392936. In press.