Roche outlines plans to move towards autonomous AI labs

TL;DR

Roche has started building autonomous AI labs, and one program went from lead identification to clinical candidate in 18 months. I think the lab will turn out to be the easier part. Portfolio reviews in large pharma typically run quarterly, and budgets move once a year. I would judge Roche on months from hit to candidate and days from a decisive result to a go or stop call.

Roche has started building autonomous AI labs, Genentech research head Aviv Regev told investors on Monday. I think the lab will turn out to be the easier part.

On 24 September I wrote that large pharma could lose most of AI's speed gain to waiting. A platform can design and test compounds every week, and a program still only gets stopped at the next governance review. Roche's plan is a useful test of that argument.

Bench speed on record in McKinsey's analysis

The bench speed is already on record in McKinsey's analysis. The cycle-time table shows Schrödinger taking SGR-1505 from hit to preclinical candidate in 10 months. Exscientia, now part of Recursion, did EXS21546 in 9 (though it later stopped that program). The large-pharma median for the same step is 32 to 50 months.

An autonomous lab closes the loop at the bench. The AI designs the experiment, the robot runs it, and the result sets up the next round. That compresses the scientific clock. The decision clock runs separately.

Portfolio reviews in large pharma typically run quarterly, and budgets move once a year. A lab that learns every week would then feed a committee that decides every quarter.

Two numbers to judge Roche's plan

So I would judge Roche's plan on two numbers. Months from hit to candidate, where the fast end is now about 10. And days from a decisive result to a go or stop call. Roche showed investors some of both. One program went from lead identification to clinical candidate in 18 months, and a new target assessment process returns a recommendation within about two weeks. I expect other large pharma companies to copy the two-week target review first.

It costs far less than an autonomous lab, and it works on the clock where I think most of the speed gets lost.