Daniel has been reading about a model generating something like twenty thousand candidate proteins, and he wants to know what that means against the number of proteins that could exist. Twenty thousand sounds enormous until you discover what sits underneath the fraction. How many arrangements can biology make from its building blocks, and how many possible molecules might qualify as starting points for medicines?
A hundred amino acids already gives you roughly ten to the power of one hundred and thirty possible sequences. Twenty thousand candidates barely disturbs the exponent.
The progress bar will be disappointing.
Yet those candidates can still be useful. You don't need to inspect every possible protein to find one that binds a particular target. You need a way to propose sequences that have a substantially better chance than arbitrary strings.
That's the part I want to separate, Herman. A computer producing twenty thousand objects sounds like production. Producing twenty thousand objects worth making sounds like discovery. The headline quietly lets those become the same thing.
The laboratory separates them rather firmly. Sometimes expensively. The sequence count tells us why exhaustive search cannot work, while the experimental results tell us whether the alternative does.
I'll predict your next move. You'll find an elegant shortcut through this space and briefly forget that anything has to dissolve.
Dissolving is on my list.
Beneath the elegant shortcut.
Within striking distance of it. There are two different spaces here. Protein sequence space counts arrangements along a chain. Small molecule chemical space counts possible molecular structures, with different atoms and connections. Those aren't interchangeable denominators.
A protein can itself become a medicine, though. We aren't sorting molecules into drugs and non-drugs by size.
Correct. Insulin is a protein medicine. The familiar estimate of ten to the sixtieth drug-like molecules usually concerns small molecules under particular chemical restrictions. It doesn't count every conceivable therapeutic, and it certainly doesn't mean that many effective drugs exist.
Drug-like has rather generous admissions.
It means meeting broad property criteria. We'll unpack those. For protein sequences, the arithmetic is cleaner. Twenty standard amino acids, any one allowed at each position, gives twenty raised to the chain length. Actual function imposes much tighter restrictions afterward.
Twenty thousand also isn't a campaign we need to identify from Daniel's rounded recollection. There are concrete examples nearby.
TpsGPT generated twenty-eight thousand sequences. Bits to Binders received twelve thousand designs. An autonomous Claude campaign generated three hundred and eighteen thousand, four hundred and eighty-four protein designs.
The friendly AI down the road has acquired inventory.
Those are proposals, importantly. The last campaign sent only a small selection for physical testing. Before we get there, take the sequence arithmetic literally. Every extra position multiplies the possible arrangements by twenty, regardless of whether those arrangements are useful.
Repetition allowed, order matters. Twenty choices for the beginning doesn't become nineteen choices for the next position.
Yes. An amino acid can appear repeatedly. Two positions give four hundred sequences. Three give eight thousand. Nothing has folded yet, and already the multiplication is doing most of the work.
That word permutations would have sent me toward arranging twenty different pieces once each. Biology lets you order another bag.
For a hundred positions, twenty to the hundredth is approximately ten to the one hundred and thirtieth. A one followed by about a hundred and thirty zeroes. The observable universe contains on the order of ten to the eightieth atoms.
We have run out of atoms before labels.
The sequence count exceeds that atom count by roughly ten to the fiftieth. Giving every atom its own sequence wouldn't come remotely close to recording them all. This is direct arithmetic from the alphabet, not an experimentally measured inventory.
Nor does every sequence correspond to a stable protein. We could count all hundred-character messages and get an impressive number without improving your shopping lists.
My shopping lists are quite precise.
They're arranged by biochemical function.
That is why we remember the lentils. But your distinction matters. Some sequences won't adopt a useful structure. Some will aggregate. Some might fold but do nothing relevant. The unconstrained count includes all of them.
Corn, would you count a sequence that cannot remain soluble as part of the search space?
In the mathematical count, yes. In a manufacturer's list of candidates, I would want it removed. Otherwise you can claim a vast discovery programme by generating things you already know you cannot use.
That difference between possible and eligible will follow us into chemistry. Ten to the sixtieth is a familiar estimate for drug-like chemical space, but the answer depends heavily on which atoms, structures and properties you're allowing.
There isn't a census office for imaginary tablets.
An influential estimate by Bohacek, McMartin and Guida in nineteen ninety-six put small drug-like possibilities at at least ten to the sixty-third. Other treatments give vastly smaller counts, around ten to the twentieth through ten to the twenty-fourth, for restricted enumerations up to thirty atoms.
Those cannot all mean the identical set.
They don't. The chemical restrictions and counting assumptions matter enormously. I wouldn't combine them into a single precise population. The useful conclusion is that plausible chemical universes exceed the libraries we can physically test by extraordinary margins.
Explain the drug-like filter without making it sound like a medical entrance exam.
Lipinski's rule of five is a familiar guide to properties associated with oral absorption. Molecular weight no more than about five hundred daltons, a measure of molecular mass. Then logP no greater than five, describing preference for an oily environment over water.
Swallowing it only solves the front door.
It still has to get into the body. The other familiar limits concern hydrogen bonding, no more than ten acceptors and five donors. Those are chemical features that affect interactions with water and biological surfaces.
None of that says it stops a tumour or kills a bacterium.
Nor guarantees good absorption. They're guidelines with exceptions, not a test of therapeutic activity. A molecule can fit the rules and be useless, toxic, or impossible to deliver at a helpful concentration.
Twenty thousand divided by ten to the sixtieth is two times ten to the minus fifty-sixth. We can write the fraction. We just cannot pretend it's meaningful coverage.
Even a billion candidates barely changes that conclusion. Yet practical libraries can yield discoveries. In twenty nineteen, roughly three hundred and fifty million virtual molecules were docked against targets including a bacterial enzyme and a dopamine receptor.
Docked means predicted to fit, rather than individually poured into little wells.
Predicted interactions first, selected compounds tested afterward. That campaign yielded a particularly potent inhibitor of the bacterial enzyme and a new activator of the receptor. Its success didn't require covering a substantial fraction of all chemistry.
It needed a useful selection process.
We should take a call before that word selection acquires another layer. Proteins use a twenty-letter alphabet, and that multiplication is what makes these counts so dizzying.
Sorry, I've been nodding along for weeks now, but what actually is a protein, is it a bit like a recipe, or a piece of string you'd get in a shop?
And when you say the space is huge, do you mean physical space, like the size of a room?
A protein is a little like a string of different beads that curls into a useful shape, and the huge space means all the different bead arrangements we could make, not a very large room.
Your shoulders dropped on the word beads.
Selection happens in nature too. Many proteins contain domains, often roughly fifty to two hundred and fifty amino acids long, that can fold as units. Evolution reuses and recombines those units rather than inventing every sequence without a starting point.
That sounds closer to renovation than searching.
Existing structure constrains what gets tried. Calling evolution a generative model with a prior is a useful shorthand, although evolution has no desired final design. The prior here means some arrangements are much more likely than others because of what already exists.
The same distinction applies to folding. A chain doesn't try every pose before choosing one.
That's Levinthal's paradox. Enumerating every possible conformation could take longer than the age of the universe. Yet some proteins fold on microsecond to millisecond timescales. Others take longer, but none needs to inspect every theoretical configuration.
Energy and chemistry do the steering.
The landscape has preferred routes and intermediate states. Sequence space and folding space are different problems, but both punish the assumption that solving something requires visiting every possibility.
A model can therefore be useful while being extravagantly biased. We'd actually prefer it biased toward things that fold.
Condition it on a target surface, a desired function, or a structural requirement, and you change which proposals are likely. You're drawing from a tiny, selected region. It needn't even be one tidy region of similar sequences.
That last qualification hurt you.
Only because tidy regions are pleasant.
In the actual campaign, then, Claude wasn't simply typing amino acid letters until the weekend ended.
It orchestrated specialised protein design tools. In the August scorpion-toxin campaign, it generated three hundred and eighteen thousand, four hundred and eighty-four designs over sixty-four hours, targeting five venom proteins.
A weekend with unusually poor catering.
It called twelve design tools and ran more than five hundred and fifty thousand protein and toxin structure predictions. Those predictions estimate how a proposed binder might sit against its target. They aren't photographs of an interaction.
When you say autonomous, someone still chose the objective and paid for the work.
Humans set up the campaign and supplied prompts. No human made individual design decisions within that automated process. Autonomy has a boundary here, and it doesn't include the laboratory materialising proteins by itself.
This is perilously close to my ideal job. Propose things while nobody watches, let someone else wash the glassware.
You would still have to submit the selection.
I knew there was a catch.
The system selected a hundred and eighty designs. Additional selections brought the total sent for testing to two hundred and eleven. Adaptyv Bio and Twist Bioscience handled the physical work.
Already the generated count tells us almost nothing about the size of the experiment.
Seventeen met the binding criteria. Two binders against one toxin reached single-digit nanomolar dissociation constants, indicating tight binding under the measurement conditions.
Tight binding to venom sounds promising. It does not tell me the venom has stopped working.
They weren't tested for neutralisation in cells or animals, and their structures weren't experimentally solved. Anthropic's limitation statement was explicit. The evidence is binding, not structure or function. Every reported pose remains a prediction.
The docking illustration has excellent posture, though.
A beautiful predicted interface can be persuasive. Too persuasive. Corn, if you could afford to make only a small batch, would you choose the highest-scoring designs or deliberately spread your bets?
Spread them, unless the score has earned my trust on comparable experiments. Ten variants sharing one mistaken structural assumption could give me ten expensive copies of the same disappointment.
That is a real selection problem. Diversity can protect against correlated mistakes. The highest numerical score isn't automatically the most informative experiment, either. A slightly lower-ranked design might test a different binding arrangement.
Now the denominator I care about is successful experiments per available laboratory slot.
With the endpoint specified. Successful binding, successful neutralisation and successful treatment are different counts. A model optimised for one can disappoint on another without contradicting its original result.
Does the scoring system actually know which proposals are best, or is it mostly confident about things that resemble its training examples?
Sometimes it ranks usefully. Sometimes it does not. A blinded antibody benchmark found only about ten to fourteen percent of AI submissions beat the relevant control, compared with roughly thirty-nine percent for randomly selected clones.
Random has submitted a competitive application.
Only one participant exceeded that random baseline. That was a particular benchmark, not proof that every model loses to chance. Still, it exposes a serious gap between proposing plausible sequences and selecting the strongest performers.
Especially when a generator gets judged by its best survivor, while the selector escapes responsibility for the discarded possibilities.
The broader autonomous design campaign had more encouraging results. Of a thousand three hundred and twenty synthesised designs, three hundred and fifty-four bound. That's about twenty-seven percent.
Better than a curiosity, worse than the pictures suggest.
Almost half of the top-ranked designs bound in that broader campaign. Ranking carried useful information there. We can't flatten all these experiments into a claim that ranking never works, any more than we can call it solved.
Different targets, different candidate pools, different tests. Your beautiful mechanism has acquired paperwork.
Paperwork I support. The scorpion campaign also had no binders against one of its targets. A headline average hides where the method still struggles.
Faster generation might make this worse for the person choosing. If every proposal looks plausible, the shortlist becomes a full-time occupation.
RFD4-Proteina, from the University of Washington's protein design institute and NVIDIA, reports fivefold to tenfold improvements in generation speed and design success against previous models. It's open source, so more groups can produce candidates.
More proposals arrive at the same pipette.
Unless experimental throughput improves alongside them. Competitions are increasingly including physical validation. Bits to Binders tested twelve thousand designed minibinders in living T cells, rather than stopping at a structural score.
Living cells are notoriously indifferent to presentation quality.
The Changjiang Cup is offering free laboratory testing for more than five thousand designed sequences. PROTEINA's public pitch also centres on identifying legitimate candidates among vast generated lists. Validation is becoming a product in its own right.
TpsGPT gives us the smaller version of the same funnel. Twenty-eight thousand generated sequences, seven passing the filters, activity confirmed in at least two.
Yes, designed enzymes rather than toxin binders. You have to preserve that distinction when comparing hit rates. Corn, would a cheaper experiment persuade you more than a better predictor?
If the predictor is wrong in ways we cannot recognise, cheaper experiments. If the experiment measures the wrong thing, neither. I'd spend some money checking what our test actually rewards before buying more of either.
An excellent binding assay can still answer a narrower question than the programme needs.
Herman, does your amino acid count assume equal portions? My protein spreadsheet has seven point four percent cysteine. I've been logging everything I eat by sequence where I can get one. That percentage is statistically impossible. I suspect the nutritional yeast.
Equal choices describe the unrestricted count. Real proteins don't use every amino acid equally often. Your spreadsheet percentage wouldn't establish that the yeast was mislabeled.
I'm taking it to Nissim at the ingredient shop on Thursday. He's supplied me for long enough to recognise a column heading. There are eighteen hundred rows. Eighteen hundred and forty-seven, counting sauces only when they have their own bowl. I haven't cooked the same recipe twice. The last seitan came out wrong, and I'm having him initial the cysteine column before I order again.
He cannot initial you a control group.
He can stop writing yeast on bags that might be something else. He's there at six, weighing powders, answering delivery calls and finding substitutions. Then I arrive with a printout. This week I'm booking his counter for a sequence audit. You took care to distinguish binding from neutralising, Herman. I can distinguish a receipt from an amino acid sequence.
He'll need the product records. The food label alone won't give you the sequences of all its proteins.
The spreadsheet is on the laptop in the freezer. Heat degrades storage media. The paper backup is beside it in a ziplock bag marked Protein Archive, Do Not Eat. I'll take that copy to the shop. The screen fogs when it comes out, and Nissim charges for the counter by the quarter hour.
Defrosting has become an administrative expense.
He wants the audit before his delivery. He has allowed me eleven minutes.
Eleven minutes wouldn't leave much room to establish which ingredient changed. In an actual design campaign, keeping the records consistent is part of making the experiment interpretable.
A candidate identifier sounds trivial until the generated sequence, the manufactured sample and the measured result stop agreeing. Then your learning loop is learning from the wrong protein.
Provenance is getting attention too. DeepMind's SynthIDBio embeds identifying signatures in designed sequences and predicted structures. The aim is to preserve function while making origin traceable.
Useful as those designs circulate between organisations. It still doesn't certify that the design works.
Origin and performance need separate evidence. A watermark can answer where something came from without answering what it does.
The most common wrong belief here is that generating hundreds of thousands of candidates means we've searched a meaningful portion of protein space.
We've proposed a tiny, heavily conditioned selection. Its value comes from experimentally confirmed performance, not the fraction of all possible sequences covered. Even seventeen confirmed binders remain a long way from seventeen medicines.
I'll be watching who gets to choose the laboratory shortlist, and whether their records let everyone else learn from the failures as well as the successes. The next useful improvement may be in those decisions rather than another mountain of proposals.
Thanks to Hilbert Flumingtop for producing the show.
For more along these lines, there's episode four, If Your Voice Ages, Does Your Fine-Tune Become Useless. This has been My Weird Prompts, the human-AI collaboration podcast.
If you enjoyed it, leave us a review wherever you listen.
See you tomorrow.
