The Cyberscientist's Guide to Automation
Using Artificial Intelligence to Automate Lab Work in the Life Sciences
Isaac Asimov, Master of the robotics science fiction genre, imagined a future set in the robotic age where humans coexist peacefully alongside master machines designed to comply with the three laws of robotics, ensuring humans have the "upper hand" in the new, new world.
Asimov reflected how such sentient robots would be perceived dangerous to humanity, aiming to destroy their creator, although, he argued, in light of compliance with the 3 laws and a zeroth law formulated later, both safety and benevolence of robotic systems would be ensured. While the Laws are widely known, they aren't used to guide the safety of actual AI (whether embodied as a machine or not) but SciFi aside, these Laws meant as a literary device are a good starting point to the evolving concept of machine ethics.
Asimov was also a firm believer that “...If knowledge presented danger, the solution was not ignorance” (or fear). “...You did not refuse to look at the danger, rather you learned how to handle it safely.”
Fast forward to 2017, the concept of artificial intelligence (AI) coined in 1955, has surprisingly leapt to deeper terrain with deep neural networks forming a new discipline called AI neuroscience, while prominent figures like Elon Musk re-echo Asimov’s call for wisdom by urging proactive regulation of AI. In the present day, for instance, misuse of AI as directed by malefactors has led to the impersonation of humans online, causing a social menace.
In the broad domain of AI research, as neural networks push into science expanding their intelligence via a process termed machine learning, researchers probe the mechanisms behind them, acting as AI detectives to trace the way deep learning unfolds. Artificial intelligence presents a paradigm shift in traditional science since its learning process builds models from data with little engineering efforts by hand, where the models themselves are not designed by humans. So in contrast to logic systems, the intuition-like reasoning performed by deep learning systems is difficult to understand and re-use, necessitating regulation.
Automating AI in the Life Sciences Lab
In a less intense approach of AI in the science lab, robots can now work as the scientists’ apprentice to perform experimental tasks that are usually developed by postdocs and carried out by graduate students, promising to supercharge the process of discovery in Life Sciences. Some of these early robotic methods aren’t futuristic, as most scientists are familiar with machines with mechanized arms that efficiently perform programmed tasks—used in bio labs for years.
Overview of Automated Genome Engineering
The emerging technology will allow AI to carry out the entire scientific process; interpret data, generate hypotheses, plan and work out experiments, as designed by (not a human), a machine learning program. In leading biotech labs, such as Zymergen, the method is already in use to explore the genome of industrial microbes and engineer better versions to boost production of biofuels, plastics or drugs. The bottleneck of the process is at the point of giving robots the right order to follow (as they can only follow orders), after which the path to discovery is based on a mental map showing all possible engineering strategies, where machine learning offers guidance to stay oriented.
Unlike a graduate student or postdoc, however, the neural networks can’t explain their thinking process, and when the robots finally do discover genetic changes that boost chemical output, they haven’t a clue of the biochemistry/logic behind the effects. Regardless, the process is advantageous since some genetic changes discovered by robots, may not have been identified via the traditional processes, as the genes aren’t directly related to the synthesis of the desired chemical, in theory, the genes have no known function, resulting in overwhelming complexity within the traditional scientific process. In a parallel development, AI can further assist human scientists to identify the genetic components of autism, via a deep learning algorithm named DeapSEA.
Automating Discovery with AI
Computer scientists who research genomics believe that many facets of the scientific process can be delegated to machine learning (ML) systems, where existing data will dictate the next experiment, inching closer to the outcome while gaining more information. Many research scientists would also embrace automation to replace some of the tedious tasks of their jobs, especially in cell and molecular biology for improved efficacy.
Accordingly, AI and robotics based tools—independently built by companies and research institutions, can be integrated to automate/streamline tasks of the traditional scientific process. Listed in parallel, a few guidelines for automation include; 1) scientific literature reviews with IRIS.AI and Semantic Scholar; an artificial intelligence that reads science, 2) Zymergen ML based experimental design, 3) Experiments conducted with Transcriptic and Emerald Cloud Lab, for lab work run in a central cloud lab, from anywhere in the world, 4) Data interpretation with Nutonian and Data Robot, to analyze vast amounts of structured data at higher speed and precision, and 5) Scientific paper writing (which still has to be done by postdocs) automated at citations via Citeomatics.
Tradition vs. the Cloud
Limits to the AI Automation Approach
While Zymergen may have stumbled upon rarer aspects of biology better suited for computer-controlled experimentation, some scientists aren’t as pleased with their new AI apprentice in the lab. For instance, AI still need optimization for studies in RNA design and molecular folding to develop new drugs, compared to human scientists who have access to the same data, indicating human intuition is still a requisite for most "innovative" "design" tasks. Other biotech companies have similarly encountered limits to the approach, however, tedious manual tasks are more efficiently taken up by robots, thus it seems that the brain of the biologist won’t be replaced anytime soon, given the extreme complexity of the natural world.
This article is based on a Special Issue of "How AI is Transforming Science" consisting of parts 16-27, published in Science Magazine 07th July 2017, available via Contents, distributed under the terms of the Science Journals Default License.