Automation is coming to life science
We are moving toward a new paradigm of reproducible, scalable research.
I recently announced the 2022 Bioautomation Challenge, a program I designed to kickstart academic use of cloud labs to promote reproducibility and big data collection. It has generated a lot of discussion about what the future of research science looks like. As a person who has worked both as a software engineer (Dropbox), a bench scientist (MIT Biological Engineering), and extensively with laboratory automation (like PRANCE and PyHamilton), let me provide some perspective on where we’re headed.
In-person research will never go away.
The first thing to know about automation in life science research is that in-person experiments aren’t going away. Even researchers who succeed in incorporating automation into their research workflows will still do plenty of science in-person, by-hand. Claiming otherwise causes pro-automation groups to lose credibility with bench scientists.
In general, discovery-based research benefits from scientists using all their senses to physically interact with — and think about — what they’re building. In science, often the serendipitous observations are the ones that matter most. In engineering, developing new devices inherently requires that your prototype is in front of you so that you can eyeball it/smack it/duct tape it. Many people think with their hands, and that’s OK. No matter how advanced automation becomes, some tasks will always benefit from being done by the researcher. The goal should be to automate everything else.
Owning robots is not the way to automate life science research.
The joke goes that you can identify the liquid handling robot in an academic lab because it’s the box that has the most dust on it. It’s funny because it’s true: liquid handling robots in academic labs are notoriously underutilized. Robots are expensive to buy (say, $300k for an acoustic liquid handler), expensive to service ($30k/year service contract), tricky to learn to use (today, every robot has a different custom GUI, with no universal programmable API), and annoying to maintain (especially when most users are novices, who cause a lot of wear and tear). Turnover in academic labs tends to be too fast to make in-house robots useful: by the time a student has learned to effectively use the robot, they graduate.
In contrast, core facilities, biofabs, and “cloud labs” sidestep many of these problems. Timesharing expensive equipment is financially a good deal for everyone, and allows academic labs to access a much wider wide array of instruments than they would ever be able to purchase. Fixed technical staff operates and services robotics, avoiding turnover issues. These facilities are capable of revealing functionality through a single integrated programming interface, providing a huge advantage over using a bespoke GUI for each instrument. They also avoid the traditional difficulties with outsourcing science to CROs by collecting metadata about every step of the process so that the researcher can understand exactly what occurred in their absence. For all these reasons, embracing work done remotely, specified using standardized instructions, would be a fantastic move for research science.
Hybrid workflows are already here.
Successfully incorporating automation into life science requires that we acknowledge the importance of the hybrid approach: some workflows are done in-person, other workflows are remotely automated.
This is not a new idea. In fact, hybrid workflows are already here. For example, DNA synthesis is a process so routine that nobody does it themselves anymore. Instead, DNA synthesis is done off-site and shipped next-day air to the researcher. Similarly, standard Sanger and NGS DNA sequencing is often also outsourced to a sequencing provider, involving a second shipping step at the end of many workflows. Offloading these routine protocols saves everyone’s time and money because routine processes can be done more cost-effectively and more robustly by specialized providers.
Today, most experiments done in life science involve a substantial component that is done remotely, to everyone’s benefit.
What does the future of automated life science research look like? It follows a similar hybrid model, where an increasing fraction of the hands-on-experiments are performed remotely and systematically. Put the samples in front of the researcher at whichever step actually benefits from this, and seek to replace other steps with robust, repeatable protocols that are executed remotely.
Value to academia is more than just the balance sheet.
Cloud lab providers often balk at the hybrid approach, pointing out that it “will never make financial sense” to work only partially in the cloud. This is a reasonable argument when you consider for-profit startups as the primary customer. In academia, the story is very different.
Life science research fundamentally moves slowly because of the reproducibility crisis. It is a perpetual struggle to transfer protocols between labs and get them to ‘work in your hands.’ Automated protocols are valuable in part because they can be frictionlessly shared between groups. The ability to simply run an optimized protocol from another lab on your sample tomorrow, side-by-side with the original controls, would be transformative to life science research. The value of that agility would be immense. Similarly, automation offers the opportunity to capture optimized protocols from students before they graduate, ensuring that future lab members never need to reinvent the wheel. Writing standardized, sharable protocols is a mechanism for truly ‘standing on the shoulders of giants.’
Overcoming the status quo with the 2022 Bioautomation Challenge
Despite overwhelming benefits of embracing automation in research science, there are two stumbling blocks that prevent a change of the status quo. First: grant money. Access to a facility that enables fully programmable science costs $300k/year, plus reagents. This is within the reach of well-resourced labs, but there are currently no grants specifically dedicated to providing academics with automation access, and it seems unlikely that conservative funders like the NIH would look favorably on this as a line item in a budget. Second: risk. Embracing automation represents a substantial paradigm shift that is sure to be bumpy. Can important methods be automated gracefully? Will automation be as valuable as imagined? Together, very few academic labs have chosen to experiment with remote automation.
To alter the status quo, Schmidt Futures has funded the 2022 Bioautomation Challenge, a program we designed to remove barriers to academic use. Groups can propose a particular method they want to automate. Funded proposals will receive 3 months of development time in Emerald Cloud Lab, a generous reagent budget, and continued cloud lab access for 9 months. We will focus in particular on proposals in the protein engineering space, an area that benefits most from the ability to collect large, high-fidelity datasets and integrate them with machine learning methods. Although we are especially interested in academic users, the challenge is also open to groups in industry. Proposals are due March 16th, 2022.
The 2022 Bioautomation Challenge is itself an experiment
Where do we get started when thinking about incorporating automation into research? What sorts of workflows benefit from being done with automation? The Bioautomation Challenge is itself a crowdsourced experiment in which we challenge labs worldwide to experiment with new ideas about how automation can accelerate their research in order to see what works and what doesn’t.
Whatever else happens, we can be sure that there will be some success stories. It will be fascinating to discover what they are!
Acknowledgements: Thanks to Brian Trippe, Ben Thuronyi, Devon Stork, and Alex Lenail for their feedback on this essay.