Vision

Introducing autonomous labs

About nine in ten drugs that enter clinical trials never reach a patient, often on an assumption that was testable years earlier. An autonomous lab works a question around the clock, checks every result, and tests the weakest assumption while testing it still costs a plate.

Introducing autonomous labs

About nine in ten drugs that enter clinical trials never reach a patient.

Many of them fail on an assumption that was testable years earlier: an effect that held in one cell line and not the next, a target that mattered in the model and not in the tumor, a mouse result that never carried over to people. The warning was usually there. What was missing was someone with the time to go and look for it.

That is the quiet constraint under all of drug development. A program has a handful of people, each of them has a fixed number of hours, and those hours go to moving the lead forward. The experiment that might kill the program is the one nobody is assigned to.

An autonomous lab takes a question and works it around the clock. It forms hypotheses, analyzes the data, checks every result, designs the experiment that would settle what the data can't, and writes all of it down. For a drug program, that means the weakest assumption gets tested now, while testing it costs a plate.

The Autonomous labs view in Heureka Bench: three labs, each with its question, its latest finding, a Running or Paused status, and the number of steps it has recorded.
Three labs, three questions. Each lab has its own question, its own record, and its own line of experiments, and they all run side by side. Screens in this post show demo labs with illustrative data.

Fail in a dish, not in a trial

Each approved drug carries an estimated $2.6 billion in development cost, and most of that is the cost of the programs that failed along the way. The later a program fails, the more of that bill it has already run up. So the most valuable thing a discovery team can do is be right earlier, and the second most valuable is to be wrong earlier.

An autonomous lab is built for exactly that work. It does three things a stretched team rarely has the hours for.

  • It tests the weak link first. Given a program's thesis, the lab ranks its ideas by how much each would change the answer, and goes after the assumption most likely to break while breaking it is cheap.
  • It asks whether a result travels. Before a finding moves up from one cell line to many, or from cell line to animal, the lab checks it against measured biology across many contexts. What holds in only one context gets flagged before anyone builds on it.
  • It spends on the survivors. Fewer, better experiments early, so the expensive stages only see what has already earned its place.

The lab in the screenshots below is doing this. Its question is a target-validation question: is a gene essential only in one molecular subtype of colorectal cancer, and does that dependency carry into patient tumors? Watch what it does with it. It confirms the effect in a second batch, checks that it follows the biomarker and not the tissue, records the approach that did not work, and then asks for the one knockdown that the existing data cannot replace.

Sources: Sun et al., Acta Pharmaceutica Sinica B (2022), on the ~90% clinical failure rate; DiMasi, Grabowski & Hansen, Journal of Health Economics (2016), on the cost per approved drug.

It works the way a good lab does

An autonomous lab is organized like the labs you already know, because that structure is how science catches its own mistakes.

  • Several researchers, each with their own approach. Every question is worked more than one way: the conventional analysis, an alternative that makes different assumptions, an exploratory look at what the others might miss. A finding counts when it holds up across them. When they disagree, the lab says so, and the disagreement becomes the next thing to resolve.
  • A reviewer who checks the work. Nothing reaches you until it has been checked against the data. Numbers are re-derived from the files that produced them. What doesn't hold up is struck, and the strike stays on the record.
  • A PI who decides. At each lab meeting the lab takes stock: what is known, what is still uncertain, and what to do next.
  • A notebook that misses nothing. Every step is written down as it happens and linked to the steps it builds on.

It runs the whole loop of the scientific method, hypothesis to design to experiment to learning and around again, continuously. You give it a question, and it keeps working that question until it is answered or you tell it to stop.

One autonomous lab works like a team of ten or more researchers at full stretch, every hour of the day. It gives each of your people a team.

One short read, at your hour

You do not have to watch it work. The lab keeps a brief: a short account of where things stand, updated as it goes, and delivered when you want it, every evening, once a week, or only when it needs you.

The brief opens with a headline you could repeat to a colleague in the hallway. Below it: what the lab is working on, what changed since you last looked, and what it is still unsure about. That last part matters most. A lab you can trust with a question you care about is one that tells you where it is unsure.

A lab's page in Heureka Bench. On the left, the brief: a headline finding, what the lab is working on, and what changed, including a negative result. On the right, a message box with Catch me up and Send buttons, and two earlier messages with the lab's replies.
A lab's page. The brief on the left, in the lab's own words, with the negative result listed next to the positive ones. On the right, the conversation: one question the lab answered from its record, and one change of direction it adopted.

And when you want more than the brief, talk to it. Ask what it found since Tuesday. Hand it a paper or a new dataset. Tell it to put one question ahead of another. A direction from you is adopted and recorded like any other step, so six months later it is clear which turns were yours.

Experiments designed to decide

Some questions can't be settled from data that already exists. A good lab knows when it has reached that point, and so does this one.

When it gets there, the lab designs the experiment that would decide the question and sends you a request. Every request states:

  • The hypothesis it is testing.
  • Why data alone can't settle it.
  • The design: the model systems, the controls, the replicates, the readout.
  • What settles it: the result that would count as a yes and the result that would count as a no, written down before any data comes back.

That last line is the discipline that makes the result worth having. The lab commits to how it will read the experiment before it sees the outcome, so the analysis stays honest whichever way the result falls. In a program where everyone wants the target to be real, that is the check that is hardest to keep and most worth having.

Two experiment requests waiting on the researcher. The first proposes knocking down WRN in two MSI-high and two MSS cell lines, with the hypothesis, why data cannot settle it, the design, what settles it, and where to return the data. Each has Approve, Ask for changes and Decline buttons.
Two requests, waiting on you. Each one arrives ready to run, with the result that would settle it stated in advance. Approve it, ask for changes, or decline.

You decide on every experiment. Approve one and it runs at your own bench; add the results when you have them and the lab reads them against the plan. For teams without a bench, or without the right assay, we are building a network of partner labs that run these experiments every day and send the data straight back.

A record you can defend

A result in drug development is only as good as your ability to show how you got it: to your own team at the next stage gate, to a partner, to an investor doing diligence. An autonomous lab's record is the product as much as its findings are.

Every step is numbered and typed: a hypothesis, a result, a check, a lab meeting, an experiment request. Each one links to the steps it builds on and the steps that later built on it. You can read the record by question, with the answer and the step that settled it, or step by step, searched and filtered.

The record by question: two open questions, one waiting on an experiment, and one answered. The answered question shows its answer and the two steps behind it, a check by the sceptic and a result by an analyst, each with its outcome and links to related steps.
The record, by question. Open questions first, then the one waiting on an experiment, then the one that is answered, with the steps that answered it.

Three properties of that record are deliberate:

  • Negatives stay. An approach that did not work is recorded as plainly as one that did. Dead ends are as visible as wins, so nobody pays for the same one twice.
  • Changing its mind is a new entry, never an edit. When new data overturns an earlier step, the lab records a revisit that points back to it and says why. The original stays as it was.
  • The work is real code, in your project. The scripts behind every result live in your project next to the data, tables and figures they produced. Anyone can open them and run them again.
The record as a list of all steps, with a search box and filters by kind. Visible steps include a lab meeting, a check that re-derived a result against a shuffled control with its reasoning and input file shown, a result, an experiment request, and an attempt marked inconclusive.
Every step. A check that re-derives a result against a shuffled control, with its reasoning and the file it used. Further down, an attempt that came back inconclusive, kept on the record with everything else.

Measured biology, built in

An autonomous lab reasons with Archimedes, our model trained on real experimental measurements. It places your samples among hundreds of thousands of measured ones, so the lab reasons over the measurements themselves, and reads the literature alongside them.

That shows up in the most expensive decision in discovery, which is picking what to test next. We gave the same AI scientist the same 31 cancer cell lines, with and without Archimedes, and a budget of 30 experiments each. With Archimedes it reached the same result after 17.

That is 13 experiments nobody had to run. In a discovery campaign, where each experiment can be a validated knockout model, the experiments you never needed are months back on the timeline.

A biotech's first program, before its first hire

The shape of an early biotech has been fixed for a long time: raise money, hire a discovery team, and find out a year later whether the thesis survives contact with data. An autonomous lab changes the order.

  • For founders. Bring a question and a thesis. The lab does the work of an early discovery team: it pulls the public data, builds the case, tests the assumption most likely to sink it, and designs the experiments worth running. You learn whether the idea holds before you have built a company around it.
  • For discovery teams. Run a lab per target, or per program, side by side. The questions your scientists never get to are worked in parallel, and every one of them arrives with its record.
  • For teams without a bioinformatics group. Drop the data in and get the analysis back: quality control, statistics and figures, with the code that produced them.

Biology is one of the hardest problems there is, and most of what goes wrong in disease is still unexplained. There are around 7,000 known rare diseases, and about 95% of them have no approved treatment. Closing that gap will take more science than the people we have can do alone. Autonomous labs are how we intend to help: more questions asked, more weak ideas caught early, and more of the effort spent on the programs that deserve it.

Sources: NIH National Center for Advancing Translational Sciences; Nguengang Wakap et al., European Journal of Human Genetics (2020).

It does the science. You set the direction.

The lab runs itself, and it stays yours to steer.

  • Pause it, resume it, stop it. One click. Give it a stop condition and it pauses itself when it gets there.
  • Redirect it whenever you like. A message is enough.
  • Follow it from anywhere. From Heureka Bench, from the web, and from your iPhone with Heureka Mobile.
  • Its record stays yours, whatever you decide to do with the lab.

Labs are set up in Heureka Bench, and we set up every lab with you.

Talk to us · Download Heureka Bench · See how it works