Engineering Nature's Most Advanced Nanomachines

Proteins are the most advanced nanotechnology built by nature. They are the molecular machines that run almost everything in all of nature: copying an entire genome with breathtaking fidelity, moving every muscle, powering every cell, holding every thought. The paths to curing disease, editing genes, digesting plastic, making crops climate resilient, and building new materials all run through proteins.

And yet proteins have been brutally hard to engineer. The most valuable applications such as Keytruda, a single engineered antibody, now earn over $30B a year treating cancer. Enzymes are already everywhere, from detergent to industrial manufacturing. And yet a single drug still takes 10-15 years and over $2B to reach approval. The reason is simple: protein engineering has always been a game of trial and error.

For most of biotech, breakthroughs began as discoveries. A strange bacterial immune system was developed as a gene editing system. A naturally occurring antibody became a drug. An enzyme found in some corner of nature became an industrial catalyst. Academic labs uncovered mechanisms, organisms, proteins, and pathways that happened to work, and companies spent years adapting them to specific use cases.

That era produced extraordinary science. But it also meant that biology's most important problems were often attacked with imperfect starting points. We took what nature gave us, modified it as best we could, and hoped the result was good enough.

That is discovery, and while we need more of them, it is not engineering.

Engineering runs the other way. You begin with the outcome you want and design every part of the system to reach it - nothing kept because it happened to be lying around, every part there for a reason. It is how we build bridges and chips and aircraft. It is not yet how we build biology.

The obvious hope is that AI will close the gap with generative protein models and will simply scale their way into the engineering era. And the models are real - we have ones that generate entirely new proteins from scratch: sequence based models like ProGen, autoregressive in the way ChatGPT is, and structure based ones like RFdiffusion, generative in the way image diffusion models are. On the narrow objectives we point them at - bind this target, fold into this shape - these designs can already beat what nature happened to evolve.

But blind scaling alone won't get us there, and it's worth being honest about why.

These models break on out-of-distribution problems, which is exactly where the valuable applications live, and they need enormous amounts of wet-lab data to improve. Brute-force scaling isn't the answer in biology the way it was for language, as much as we want it to be. We're working with a fraction of internet-scale data, and simply adding parameters hasn't produced the breakthroughs we saw with LLMs.

More fundamentally: the loss function nature optimised for is not the loss function we care about. Nature tuned proteins for an organism's survival in its own environment and not to be a durable therapeutic, a high-yield industrial enzyme, or a highly efficient compact editor. The labeled data for the objectives we actually want doesn't exist, and no amount of scraping nature will conjure it. This is why just scaling raw protein models hasn't worked and even lower parameter protein models perform at par if not better than bigger ones.

So just hoping that more genomic data and bigger models will solve everything isn't the right strategy as the data we would actually need is simply too expensive to be generated at the magnitude required.

But what if there were another way? Instead of only scaling raw data, what if we used these newfound capabilities to look deeper into the proteins themselves? To work out what actually makes a specific protein better, and to find generalizable signals that hold across protein classes, across structures and sequences?

We still generate data and run experiments in volume - biology demands it. What changes is the aim: every run is pointed not just at a design that works, but also why it works - and that understanding is what we build the next design on.

For the first time, that is possible. We can run open-ended, long-range experiments around the clock. We can analyze tens of thousands of designs in silico in a day. Work that used to take years of research now takes months - and AI will only compress that further. The bottleneck is no longer how fast we can try things. It is whether we are trying to find the right signals.

That is exactly why I am building Mandrake Bio. To turn protein design from a discovery problem into an engineering discipline.

At Mandrake, we spend as much energy on why something worked as on the fact that it worked. We want to work out what actually makes a protein better, then feed that understanding directly back into how the next one is generated. To do that, we're building custom assays and tight feedback loops, designed to pull as much signal as possible from every experiment - signals that generalise across whole families of editors, not just a single protein. This also gets us a really diverse data-set measuring Protein Characteristics X Function along with traditional Protein X Function datasets.

We are also assembling one of the most richly connected DNA-protein database - collating data from the treasure trove of publicly available, but not easily accessible datasets. Not to mine it brute-force, but as the substrate for that understanding. We've already processed ~21.9 Tbp of genomic sequence across 6.72M genomes, spanning viruses, bacteria, archaea, and every eukaryotic kingdom.

And we have early signs that our approach works. In a GEM X Adaptyv global competition to design a binder to RBX1 - a target long considered nearly impossible - coupling deeper biological understanding with frontier AI, let us design the only strong binder in the competition, against some of the leading labs and companies in the field. While our in-house results haven't been published yet, it validates our core thesis: you can brute-force biology up to a point, but for the problems that matter most, you have to understand it deeply.

We are starting with in-vivo gene editing - a transformative technology still held back by the same imperfect starting points. Crop improvement takes 7-10 years; a single gene therapy can cost millions of dollars per patient. We think crop improvement should take a fraction of that, and a gene therapy should cost what a surgery does. The next generation of editing systems won't be mined by accident; they will be designed from the ground up for the biological problems we actually care about.

That world is not here yet. Biology is still hard. We are still learning which signals to trust, which assays to run, and which loops actually compound. But that is the wild part: working out what makes nature tick, then turning that understanding into engineering will unlock an era where we can build genuine cures, take on ageing, and design the most efficient materials we've ever made.

We are hiring across research and engineering. If solving one of the hardest frontiers in biology genuinely excites you, write to us.

Let's make life programmable.