NEWSLETTER · AI-NATIVE BIOTECH · October 2026 · read online (no email list yet)

Same shortage, four different labs

Four AI-first drug discovery companies tell the same story about what is hard. Their labs say something different.

The shared claim. BigHat, Genesis, Absci and Manas AI each describe accuracy in designing molecules as the thing that is scarce. Genesis puts it plainly: “if your accuracy is not sufficient, it will therefore not be useful for the downstream things that you care about, which is both potency prediction, but also prospective design.” (Genesis, Latent Space, speaker unconfirmed.) Strip away the pitch decks and that is a shared bet: if the model's answer is good enough, the rest follows.

The difference. They do not own the same thing next to that model.

BigHat built a lab around the model. Its CEO: “We have a lab that makes these molecules and generates data about them. We have AI tech that can use that lab to make molecules and test its predictions and learn from that.” And the point of it: “We wanted a situation where we could do data science and machine learning today, and make predictions tomorrow, and know the truth of those the next day after.”
Genesis also keeps wet-lab work in-house: “there's numerous advantages to having in-house wet lab, data generation clearly, dog fooding the models, getting really direct human feedback from real med chemists”
Manas AI says its focus is the generator and partners handle the rest: “Right now our focus is mainly on making molecular generators. So we have partnerships that do these other aspects of drug development.”
Absci's own statements describe outside testing and a dated clinical check: “We'll have a 13-week interim readout on that later this year with the 26 efficacy readout, first part of 27.”

My reading. If accuracy is the shortage, how fast a company can learn that it was wrong matters as much as the model. BigHat describes days. Absci describes months and a scheduled readout. Manas describes partners. These are not better or worse in this data; they are different ways to find out. The plot puts all four on one grid.

A foil. Recursion is not in the group. Its executive says “building the best drug isn't generally the number one limitation. It's picking the right target.” That is a different bottleneck, with a lab it says runs about 2.2 million experiments a week, which is why it is shown only as a reference.

What this does not show. Two or three sources per company. Machine transcripts with no audio check. Labels are readings, not measurements; low-confidence points are hollow on the plot.

Next in the series: the same grid applied to all 41 companies we have read so far.

Who owns the lab, and how fast each can check its own claim. Hollow dot = low-confidence reading.

The full table with sources is on the resource page.

Quotes come from machine-made transcripts that have not been checked against the audio; speaker identity is unconfirmed where the source has no labels. Labels are automated readings, unreviewed by a human. Research and commentary, not investment advice.