Tuesday, November 03, 2026

Can AI Decode Glycolysis Enzymes By Their Moonlight?

Serendipity can be amazing. At a great lunch recently with Jake Wintermute, Patrick Boyle, and Ash Jogalekar we were roving around AI topics and synthetic biology with great relish (though my burger did not bear that condiment), and I threw out an idea I've had that riffs of a Jef Boeke paper I'm completely derelict in covering here.  I was thinking of scribbling it up, and then via a Quora request I came across a fun review that was directly relevant to my specific idea - which I'm now going to expand into a completely technically feasible, scientifically interesting, and insanely expensive grand exploration.  What I propose to do is to try to rewrite one of the most fundamental and universal biochemical processes, glycolysis, every way I can think of in as many places as possible, and in doing so uncover unknown foundational rules of biology.

Rewriting and rewiring fundamental biological systems is a new frontier.  Last year for my 19th blogging anniversary I pitched the idea of trying to remove all tryptophan or isoleucine codons from a genome.  Soon after I was at at Wyss Institute event and saw George Church, and he commented that he thought another Church diaspora member had been actually working on the problem.  Before long, out popped a publication from Harris Wang describing removing all isoleucine from 21 E.coli ribosomal proteins, which required further protein engineering to regain full functionality.

Complementation of a genetic defect in one species with a gene from another was once a common way of cloning genes or demonstrating shared function.  Perhaps the best visual demonstration of this, pun intended, is that mammalian homologs of eyeless can restore eye formation in eya null Drosophila.  It's still a valuable approach to characterizing proteins.  For example, if a human protein can complement a Saccharomyces one, then it is far more practical to screen all single amino acid variants ("Multiplex Assessment of Variant Effect") in yeast. 

The Boeke paper, with the great title "Phylogenetic debugging of a complete human biosynthetic pathway transplanted into yeast", took the complementation idea even farther, showing that an entire pathway (adenine biosynthesis) in yeast can be fully replaced by human orthologs.  The fun twist in this is that while some function was obtained with wildtype human sequences, a bit of tuning up was required for nearly full fitness.  Only a single protein was an issue.   By an elegant strategy that involved screening more orthologs and creating chimaeras, they were able to find a set of only five substitutions that brought fitness up to nearly wildtype levels.   Turns out the human protein is more likely to be degraded by the proteasome if inhibited by high product levels.  

So a bunch of intermeshed ideas.  One is to try to fully humanize the yeast proteome; for every yeast gene with a clear human ortholog, attempt to replace it.  Sort through all the single genes to see which require tune-ups.  Once an acceptable set of singles is found, then try to build combinations until we have a maximally human yeast.  Why?  Just to see what we find along the way!  Taken even further, could you pick a different organism, say a filamentous fungus such as Aspergillus, and find a new set of orthologs to knock-in for many cases where humans just lack an ortholog?  Then go closer with say Pichia, and so on.

But now this related idea of taking any yeast gene which is essential or we can engineer essentiality. what if we screen a really large library of homologs for complementation?  If the fitness results are given to a model, can it tease apart the patterns which drive success or failure?  These might include patterns such as seen in the Boeke paper, but also thermostability, interaction with other proteins, and numerous other complexities.

But with AI, could we just redesign a key part of yeast? I was originally picking glycolysis but adenine biosynthesis works too.  There's been amazing progress on computational design of novel enzymes, usually to catalyze new-to-nature reactions.  But what if we applied these to an entire extant pathway? At a minimum designed enzymes which are not recognizably homologous but with the same fold, but could you go to entirely different folds?

Could you even redesign glycolysis itself?  That is you must have the same initial inputs and same outputs, but the actual reactions are now in play?  Which gets to the fun origin-of-life/exobiology speculation: how many different possible basic biochemistries might exist for carbon-based lifeforms across the universe.

But back to the ortholog screening.  The fun additional review I stumbled on pointed out that glycolysis enzymes are some of the most notorious moonlighting proteins.  By moonlighting, it is meant that they have roles completely different from their originally discovered roles.  I was introduced to the concept, though not under that name, as an undergraduate - the crystallins which make our eye lens transparent are repurposed metabolic enzymes.  

Many of these moonlighting functions may be very taxonomically restricted, though that would seem to not be a well-explored topic.  So what if two closely related species, say Escherichia coli and Salmonella typhi, were both screened by complementation with libraries of homologs for the various glycolytic enzymes?  How strongly correlated would the fitness scores be?  If we screen the same library through more distantly related bacteria such as Pseudomonas putida or Bacillus subtilis, how different would they be?  If we find orthologs which show different fitness in different species, are chimaeras or scanning mutagenesis libraries more informative?  Or why not both?

Once we find signals of fitness defects after complementation, how to find which are cases of moonlighting and how do we find the moonlighting role? Many moonlighting functions have been basically been discovered by accidental convergence of research paths; here I want to drive straight to them.  Gene function assignment has often been framed as "what is the function of this gene"; enumerating many functions of the same gene is a much harder and less studied problem.

Clearly this won't come cheaply; I'm suggesting a lot of gene synthesis and that's still in the $50-$100 per gene range.  As Everett Dirkson might have said, a million orthologs here and a million gene variants there soon adds up to real money.  Certainly there's plenty of additional spending implied above.  But it shouldn't require a two-person detective agency to make serious inroads on further exploring glycolytic enzyme moonlighting.  




No comments: