From ideas to experiments.
SI at the lab bench.

Give a promising idea a turn at the bench.

A molecule imagined on a computer becomes much more interesting when an experiment shows what it can do. On October 16, Google Research described three projects bringing those two worlds closer together: laboratory automation, DNA molecules that bind proteins, and the design of new protein binders.

The constructive advance is a tighter connection between Super Intelligence (SI) tools and physical experiments. Instead of treating a computer-generated candidate as the finish line, researchers can feed measurements back into the next design round. Google presents these projects as building blocks for human-directed discovery, with scientists authorizing and reviewing the work.

A shared picture of the workbench.

One project, VERID—VERified Infrastructure for Discovery Laboratories, gives software agents a continuously updated digital representation of a laboratory workcell. It tracks details such as equipment geometry, reagent identities, liquid volumes and available pipette tips.

That picture helps translate a scientific plan into something a robot can actually carry out. The framework can simulate a protocol before execution and check commands against the current laboratory state. Synchronized video and action records give researchers a trace they can inspect afterward.

Google reports a useful demonstration: persistent laboratory state let an agent resume after a plate-reader interruption without repeating the liquid-handling steps. Our reading: remembering what has already happened is a practical step toward making automated experiments more useful to scientists.

Let the results improve the next idea.

A second project used DNA aptamers—short DNA molecules selected for their ability to bind a target. Across three design-and-test rounds involving two protein targets, agents proposed sequences, received experimental measurements and revised their designs. Google reports improved confirmed on-target hit rates by the final round, with the agents learning from different experimental effects for each target.

The important distinction is that feedback came from physical tests, not just another computer prediction. Measurements helped guide which molecular features to pursue next.

In a separate protein-design project, researchers studied binders for fibroblast activation protein alpha, a protein associated with solid tumors. Google reports that leading designs achieved tight binding in laboratory tests. That is an early research result, not a treatment announcement; its significance here is progress in designing molecules for a difficult target.

Build toward a more capable research partner.

The projects are not yet one fully integrated autonomous laboratory. Google identifies combining traceable automation with closed-loop protein discovery as a next step. Human oversight remains part of the reported workflows.

For readers following American SI leadership, this is a concrete example of a U.S. company moving powerful tools into hands-on science. Our view: the opportunity is a more productive partnership—scientists set the direction, software helps explore possibilities, and experiments supply the evidence for the next decision. Discovery gets a better feedback loop.