This is the fourth of a series of essays focused on enhancing scientific progress by rethinking how we do academic science as part of my fellowship with the Roots of Progress Institute. Please subscribe to be notified of the fifth! Thank you to Mike Riggs for assistance editing this essay and to other RPI fellows for valuable feedback.
5 am: Wake up and drive to the lab
6 am: Take samples of T cells from culture dishes in an incubator. Centrifuge them, resuspend once in a saline buffer, centrifuge once more, snap freeze the pellet on dry ice. Drive home.
7 am: Go to the gym with my wife, eat breakfast, shower. Drive back to the lab.
9am - 6:30 pm: Work, including repeating the sampling procedure at 10 am, 2 pm, and 6 pm.
7 pm: Eat dinner at home, do house chores, prepare for bed.
9:30 pm: Drive to the lab with my blue light blockers on. Take the 10 pm sample. Drive home.
10:45 pm: Go to bed.
1:30 am: Wake up to my Garmin alarm. Sneak out of bed without waking the wife. Get to the lab for the 2 am time point.
2:45 am: Get back home from the lab. Sneak into bed without waking the wife. Fall back asleep instantly (blue light blockers help here).
6:30 am: Wake up and return to my normal routine.
For the past year and a half, I’ve been cooking something up in the lab. The project takes inspiration from circadian physiology and applies it to T cell immunology. One of the key experiments I did involved taking samples of T cells, grown in a liquid medium (in a flask (in an incubator)), every 4 hours for 24 hours. I then took these samples, isolated mRNA from them, and sequenced the transcripts. This allowed me to see how gene expression changes over the 24 h period and I could then detect rhythmic transcriptomic patterns. The schedule above is how this was done.
Not too bad, but it was made worse by the fact that I had to repeat it twice more because I messed up some downstream steps (skill issue—I deserved the punishment). Do you know what would have made this experiment trivially easy? A robot.
Manual lab skills are an incredibly important criterion for success in science—and one that is often overlooked. However, this may not be the case for long. I may be in one of the last generations of scientists that even perform experiments with their own hands. Much like how kids these days don’t really need to learn cursive handwriting because tapping away at laptop keys is just better in almost every way, in the next decade or so, many of the technical skills that wet-lab scientists acquire and require will be made redundant by technological progress in the form of robots. This is terrifically exciting to me for two main reasons: scientific progress will skyrocket and the human experience of being a scientist will become much more enjoyable.
More and better science
Robots will unlock greater scientific output in three complementary ways. First and most obviously, robots do not tire as easily as humans. This means that the head of a scientific lab can have experiments running 24/7. The only downtime required would be for charging or maintenance, which can be synchronised to gaps in scientific protocols (many scientific experiments require periods of waiting to let reactions happen). The limiting factor for the number of experiments that can be done in any given week then becomes either experimental planning or money (or, more specifically, the reagents that money buys). Without having to perform the experiments themselves, the human scientist with capable robots at their disposal now has much more time to conceive of and plan more experiments. So, we not only replace the current human limit (experimental output) with a new, less restrictive human limit (experimental planning), but we also give a lift to this second limit itself. Thus, scientific output ratchets upwards.
Secondly, robots are more precise, consistent, and reliable than humans. Tubes of reagents are mixed and handled identically between replicates. Repeated pipetting dispenses identical volumes of liquid each time, with no increased risk of misplacing liquid due to fatigue. Once humanoid robots are good enough, they’ll also handle mice expertly: for example, they could perform intraperitoneal drug injections with perfect consistency across animals. Such improvements reduce technical error and decrease variability in the data. This has profound consequences for experimental design and statistical power: lower variance within treatment groups makes it easier to tell apart real biological effects with fewer replicates. For example, this would mean that a preclinical cancer model experiment can use fewer mice, which is good economically and ethically (if you, like most people, assume human-like mouse qualia). This also makes troubleshooting easier: it is easier to figure out what went wrong when random human error is taken out of the equation.
Further, replication of published results becomes much easier—especially if robot prompting/training/programming data are uploaded and stored somewhere as part of the publication process. Simpler replication helps to deter fraud. Copying robot data to other robots also makes scaling scientific output within one lab trivially easy. Training can be done once and then simply copied across—it doesn’t require months of explanation and demonstration for each new recruit, as it does with humans.
Many scientists reading this may already appreciate the profound improvement in data quality that automated systems currently enable (e.g., image analysis pipelines versus manual image segmentation and analysis, or robotic liquid handlers for sequencing library preparations versus manual calculations, pipette adjustments, and liquid transference). Humanoids in the lab would expand these benefits across experimental applications.
Thirdly, enhanced top-end performance of robots versus humans—perhaps best exemplified by robotic surgery—means that experimental designs can be much more complex and technically challenging. Scientists could conceive of a wider variety of experiments than is currently possible.
If we plug all of these improvements into my T cell RNA sequencing time course experiment, the result is markedly different:
After training and programming my robot, I begin my experiment. Instead of me losing sleep, my robot works through the night and takes samples for me. Because the robot doesn’t need to sleep, I can make my sampling regimen more aggressive: I go for every 2 hours for a full 72 hours, instead of my original 4-hour sampling over the course of just one day. This allows me to disambiguate longitudinal trends from circadian trends while increasing the temporal resolution of my experiment—I can ask more questions and get clearer answers to those questions from one experiment because a robot lifted the bar for what is possible.
Then, at each time point, the robot treats samples much more consistently than I could. At 2 pm, the robot takes each tube, holds it with identical finger pressure onto the vortex pad, and lets the tube mix for exactly 1.7 seconds—it then repeats every detail of this procedure in exactly the same way at the 2 am time point (and every other time point). Less noise from sample acquisition leads to a better signal in downstream steps, so I don’t miss any real but small effects.
Finally, the downstream RNA extraction and cDNA library preparation steps don’t need to be repeated, because there isn’t an incompetent human making pipetting errors in the protocol this time. Less reagent waste, less time waste. My real experiment had six paired replicates (i.e., six human T cell donors subjected to an experimental and a control condition) per time point. So 72 total samples. The robot-enabled version would have 432 samples. That’s a ridiculous number of tubes to handle and wells to pipette without making a mistake or slowing down … for a human. A robot breezes through it. It can also use smaller volumes (due to increased pipetting precision), which decreases reagent cost per sample. While I think the data I got from my real-world experiment are pretty cool, the data that I could get from a robot-enhanced version of the experiment would be significantly more valuable and interesting.
More rewarding and fulfilling science
The case for robots in labs from the perspective of scientific productivity is hard to argue against, but some scientists might recoil at the prospect of being replaced by clankers. Fear not, fellow meat-sack. I am here to tell you that you will probably not be replaced, and if you are, it will be a good thing both for you and for science.
Many readers will be familiar with the perils of academic hierarchies.
These exist because principal investigators (PIs) require a workforce of underpaid scientists to do the actual work in their labs. A vertical relationship emerges that some PIs use for their own gain. Young scientists are also forced to spend years at the lab bench working on someone else’s ideas before they get a sniff of independence and the freedom to pursue their own curiosities. Robots solve this issue. Robots will replace the research assistant, and that is good. For smart and ambitious young scientists, the pathway to independence will be compressed by robots swamping the supply of cheap lab labour. A scientist’s 20s will no longer be a valuable resource that PIs seek to harvest, but theirs to spend pursuing their own ideas.
Beyond RAs, the massive inflation of the number of people getting PhDs will be reversed by robots eviscerating the demand for cheap human labour in the lab. This will be glorious. There will simply be far fewer spots for PhD students, which will result in the degree becoming much more elite again—PIs will only hire PhD students if they can add real intellectual value. This will be better for both science and the scientist. For the scientists who are replaced, being deluded into thinking you are good at something you’re not is toxic for the soul—you are better off leaving science for something you have a real knack for. For the scientists who remain, being surrounded by well-selected intellectuals is exhilarating and fulfilling.
Given cheaper and more abundant lab labour and assuming constant science funding, the number of independent research positions will increase. This means the average scientist will do more of what scientists should do: ideate, create, and discover. This leads to a much more fulfilling career as a senior scientist, with less pastoral, managerial burden and more actual science.
Using robots for lab labour instead of humans can also make it easier to resist the allure of collaboration. One issue this fixes is uncertainty regarding scientific credit. The scientist is clearly due all credit, and the robot is the tool with which they perform their experiments. Currently, many PIs implicitly consider RAs and grad students to be tools to be deployed, but this is only true some of the time.
Proper attribution is one of the many issues of inter-human collaboration that can be alleviated via robots in the lab, but there are many others. Let’s not forget that collaboration is essentially a crutch reached for when scope, effort, or competence are stretched too far for our liking—if nobody in your lab can do a certain type of experiment, you go elsewhere to find someone who can. But collaborations are inefficient and dilute credit, motivation, and clarity of thought. This becomes less necessary with an army of humanoids in the lab who can perform every imaginable technique.
Less passive-aggressive vying for control over a project, less risk aversion and friction, more idiosyncratic science: the life of a scientist with robots at their command is an empowered, free, and fun one.
Is it not sad that humans will no longer need any laboratory skills in this vision of the future? I often talk about robots with my lab colleagues. Recently, one colleague suggested that some of the joy of science is in doing experiments manually. This is something that I will not deny. However, the great news is that nobody is stopping humans from dusting off the cobwebs and running a western blot if they feel like it. Robots will earn their place just like email has earned its place. People still occasionally write letters for sentimental or traditional reasons, but they are no longer forced to by a lack of choice. The same can apply to human-performed experiments if people genuinely miss doing them. The transition to robots in the labs can be very smooth and will be completely voluntary—driven only by the immense value that robots can offer both scientists and science.
Exciting times!
On PhD inflation and possible contraction, I wonder if the future limiting factor will be talented students, sufficient (soon to be autonomous) manpower, and/or money?
It will be interesting to see if funding mechanisms evolve, as science may evolve in a direction where scientists resemble one man teams or freelancers, and there is a tighter feedback loop between scientists' grant applications, performance, and future funding. Possibly creating further asymmetry across scientists, which could correlate with improved resource allocation efficiency, but may also also reduce variance and serendipitous discoveries due to there being fewer scientists overall.
Another thought is that when determining new protocols and experimental setups, manual work may still be valuable for initial design and fine tuning. (eg. this summer I was evaluating various fibroblast-within-collagen gel set ups, and I wonder how capable a robot would have handled developing a novel protocol)