The DREAM cloud lab that runs protein-engineering experiments automatically — NSF AI award $4.41M (Northwestern)

1Why protein engineering suits automated experimentation

A protein is a chain of 20 kinds of amino acid, and the order determines what it does. The number of possible orders is astronomical even for short sequences: a chain of just 100 gives 20 to the 100th power of possibilities, far beyond the number of atoms in the universe. Searching that space for something with a desired function by intuition and experience alone is not realistic.

The response has been to iterate: design, build, test, learn. Predict and make it, measure and learn from what missed, feed that into the next prediction. Because results depend on how many times that loop can be turned, the speed of one turn and the number of candidates per turn matter directly. That structure is why automation is in such demand in this field.

2Machine-readable failure as an asset

The other effect of automation is that conditions which did not work are recorded in the same form as those that did. In hand-run experiments, trials that fell short tend to leave thin records; in machine-run experiments, success and failure become data at the same granularity. For a predictive model, those combinations that did not work are precisely the information that narrows the search.

Rapid and Automation in the title are therefore not only about speed — they also mean raising the density of usable training data.

3$4.41M obligated, $20M estimated

The obligated amount is $4,411,752 and the estimated total is $20,000,000. Among the ten PCL test-bed nodes the obligated figure sits toward the lower end, while the estimated total is close to the top of the range. The period runs four years, from 2026-08-01 to 2030-07-31.

4Automation changes more than the speed

Once a machine runs the experiments, the record changes shape. Done by hand, a trial that did not give the expected result is sometimes never written down. Run by machine, successes and failures survive in the same form.

Rapid and Automation, in the title, point at the right column. What raises the accuracy of a predictive model is measured data in the end, and raising its density is what automation is actually for.