Max Jaderberg on Moving Drug Design From Wet Labs to Code

Key Takeaways

• Biopharma's core delay is physical: synthesizing real atoms and culturing biological cells creates an expensive, slow trial-and-error cycle.
• Isomorphic Labs targets a clear engineering threshold: drawing a literal dotted line on internal model leaderboards that matches real-world experimental accuracy.
• AlphaFold established the precedent: deep learning systems can generalize biological physics rather than merely memorizing training datasets.
• Physical wet labs should serve only as high-probability validation checkpoints, not iterative hypothesis discovery sandboxes.
• Predictive models paired with synthetic human tissues offer a path to phase out slow animal testing for complex diseases and human aging.

The Physical Bottleneck in Molecular Design

Drug discovery moves slowly because moving physical matter moves slowly. Teams design a candidate molecule, synthesize physical atoms, test them on living cells, and inspect the wreckage when the experiment fails. Biopharmaceutical productivity has stalled for decades under this empirical trial-and-error cycle. Moving drug discovery in silico means moving candidate design out of petri dishes and into silicon. Jaderberg emphasized this shift directly: “How can we bring all of that drug discovery work onto a computer? So how can we get out of the real world?”▶ 0:17 (play from 0:17)

He identified the physical loop as the primary constraint on biopharma progress. As Jaderberg put it, “the big time bottleneck the big cost bottleneck is synthesizing real atoms.”▶ 15:22 (play from 15:22)

AlphaFold proved that computing could break that bottleneck. By predicting protein folding structures straight from genetic sequences, it proved neural networks can generalize biological physics on silicon. Once a model demonstrates that degree of generalization, the design cycle leaves the wet lab entirely. The goal is building predictive world models that simulate molecular mechanics directly inside software.

The Dotted Line on the Leaderboard

Most machine learning research measures progress against internal algorithmic benchmarks. Isomorphic Labs takes a stricter approach: they anchor their models against physical reality. “Internally we have a dotted line on all of our leaderboards that is experimental accuracy,”▶ 16:44 (play from 16:44) Jaderberg explained.

That dotted line changes how computational teams work. A model that improves ranking metrics by 5% is useless if it still sits below the threshold of physical measurement error. Once a predictive model crosses that dotted line, physical synthesis becomes redundant during the design phase: “At that point, you don't need to go into the real world. You can do so much more design and iteration on a computer.”▶ 16:56 (play from 16:56)

Physical chemistry does not vanish under this model; its function shifts. Instead of relying on wet labs to explore molecular options through trial and error, founders can restrict physical experiments to terminal verification. Jaderberg noted: “We still need to go into the real world to validate because we're coming up with novel scientific hypotheses and we want to test them out in people.”▶ 26:02 (play from 26:02) The wet lab becomes an audit gate for high-confidence candidates rather than a lottery wheel.

What to Do With This

Find the physical feedback loop currently slowing down your product iteration. Draw an explicit accuracy threshold where a computational or synthetic test matches empirical real-world results. Stop running physical prototypes until your simulated iterations consistently beat that target line.