Recombination-Based Autonomous Protein Structure-Function Search

An interesting preprint showed up last week in the autonomous agent-driven experimentation space - an area I'm paid to be very interested in.  This didn't run on our hardware, but we certainly could reproduce the experiment on our autonomous laboratory installation Nebula if the preprint had all the critical details (I'll try not to kvetch constantly about that, but it is an annoying brake on my eagerness to reproduce some or all the results).  Anyone could run it on our Cloud Lab - I haven't priced it out but this should come in at a very reasonable number.  What I found particularly clever is the solution to the problem of keeping protein engineering by AI in a reasonable cost envelope - and this is accomplished by relying on chimeras which can be a generated from one round of DNA synthesis and then clever shuffling.

Learning protein function through autonomous experimental interaction

Coban Brooks,  View ORCID ProfilePascal Notin,  View ORCID ProfilePhilip A. Romero

https://www.biorxiv.org/content/10.64898/2026.08.14.744985v1

The preprint starts with a computational model I'll basically skim over, as the experimental execution is what grabbed me.  The scientific team trained a model to generate protein sequences - that's a hot area in which I'm not well-equipped to discuss nuances between schemes.  Then they set an agent on a course of using the model to iteratively explore protein function.

So they picked six glycosyl hydrolases from that most glorious of bacterial genera, Streptomyces. The frustrating part is the supplemental materials specify neither the protein sequences nor how they encoded them.  And that's where things would get interesting.

Their core idea is to divide the sequences into 8 different segments which can be freely recombined using Golden Gate cloning.  If you are unfamiliar, Golden Gate is a genius scheme using Type IIS restriction enzymes (which cut outside their recognition sequence) and ligase in the same pot to assemble genes in a deterministic and scarless manner.  When Golden Gate came out, I (and probably everyone else) at Codon was kicking myself, as we had come close to this but not quite there - our ligation assembly approach was far more work and far more error-prone.  That's the real beauty of Golden Gate - errors from cut pieces just reassembling are dealt with by being cut again.

So with this scheme, one can amplify the segments of interest with defined PCR primers, ligate them together at will with Golden Gate, express them with Cell Free Protein Synthesis (& I know where you can get some really sweet reagents for that!), and then assay these hydrolases with commercially available fluorogenic substrates.  The data can then be fed into an agent to update its models, generate new design for the next round of chimeras, and test those - all with no new DNA synthesis.   They used a fit-for-purpose automated workcell and achieved cycle times as low as 10 hours, though the average was a bit over a day.  

This does point to a gaping issue in synthetic biology - DNA synthesis remains too slow and too expensive.  Yes, it's far better than the Codon days two decades ago - you can get whole genes for less than we were being charged for oligos and the turnarounds are now in days not month plus, but a plate of 96 genes is still going to take a few days and cost you in the small thousands of dollars.  If you really wanted to explore sequence design space, you'd need these numbers to come down to more like a hundred dollars buys you a whole plate of genes.  Better still would be prices so low you'd worry more about the price of the plate the DNA came in or the cost of the shipping.

Anyways, back to the preprint.  An interesting twist is they didn't have one agent working on the problem, but instead three different ones, each trying to explore the specificity for a different substrate but all sharing a common scratchpad and workspace - and an equal share of the next set of recombined sequences to test.  Agent G focused on D-glucose, Agent X on D-xylose, and Agent M on D-mannose.  The choice of sugars made (or was driven by this idea?) keeping them straight easy - galactose wasn't in the mix!

There's always the question in these sorts of experiments of what sort of data to collect.  Since the reaction evolves greater fluorescence over time, each well supplied a full kinetics curve.  Each enzyme was characterized in triplicate against each of the three substrates.  

Murphy's Law can easily run rampant even in an automated lab, so there were some QC checks.  Golden Gate products were amplified by PCR upstream of Cell-Free Protein Synthesis, and those products were checked by a simple fluorescence assay - if an assembly failed it was marked as such and could be tried again in another round.  There were attempts to adjust for variation in cell-free expression, but I didn't see a check for failure to express.

Perhaps what is striking about the preprint is how little of the combinatorial space was explored - 1.7M possible enzymes but after 20 rounds only 132 chimeras were successfully assayed - an average of 6.6 per round!  The system could easily have explored more chimeras - it was designed at only 96-well scale and these reactions could almost certainly be scaled down to 384-well scale. But even within that limited space, the agent for mannose improved on the specificity of the parental enzymes - which was strongly glucose-biased - by three logs.  Agent X for xylose gained only a bit over a log.

There are so many additional directions one could take this.  For example, what if the agents designed a new set of segments to add to the pool?  Could a different set of breakpoints within the same gene selections be designed?  How dependent was progress on the combinatorial space available - what if more than 6 hydrolases were segmented?  How dependent was it on the particular set of hydrolases used - could another six have worked better?  What if these models were applied to other hexoses - would they have an advantage over the original naive model?  What if instead of the models working in synchrony, they kept the common knowledge space but ran in succession - Agent G designed rounds 1,4,7, Agent X 2,5,8, Agent M, 3,6,9, etc? If rounds were executed speculatively - one starting before a prior round finished - more rounds could be run on the same hardware in the same amount of wall-clock time?  Was this many rounds with few experiments per round truly superior to a strategy of fewer rounds with more designs assayed per round?

The Golden Gate strategy itself has interesting constraints which it could be fun to explore methods to work within them.  Overlaps are four bases long and there is abundant literature, much generated by New England Biolabs who sells a Golden Gate kit, on which overlaps work the best.  So you either need to have a fixed amino acid plus one more base either side or have an overlap using the last two bases of one codon and the first two of another.  It could be fun to design a program to go through a large multiple alignment to find sets of N enzymes that can be broken into S segments using a Golden Gate strategy, then optimize for the sets which are the most dissimilar.  

Someday the cost and slowness of gene synthesis won't be so frustrating to fast lab-in-a-loop design-build-test-learn cycles - but that isn't today. Clever approaches at recombining parts allow fast, inexpensive looping.  The era of AI agents driving autonomous laboratories is still very young and evolving quickly; I can't wait to see what happens next!