SP-04 · Finding Something That Sticks

The compound did not look like anything. On a screen, in a notebook, it was a string of letters and a structure drawn in lines and angles — STI571, the lab called it, or sometimes CGP57148B, which is the kind of name a molecule gets before anyone loves it. What it did, when you put it on a dish of cells, was nothing dramatic. The cells that carried a particular broken gene — the fusion called BCR-ABL, the engine of chronic myeloid leukemia — stopped dividing. The cells that didn’t carry it kept going. That was the whole event. No alarm, no shudder, no visible death throes under the microscope. A graph where one line bent down and another stayed flat.

It is tempting to imagine the moment was bigger than that. We like our medical history with a clap of thunder — the night the mold killed the bacteria, the instant the vaccine took. But the documented fact is quieter and, in its way, more honest: in 1996, a paper in Nature Medicine reported that this molecule selectively inhibited the growth of BCR-ABL-positive cells in a dish, with around a micromolar dose doing the work. That molecule would eventually become imatinib, sold as Gleevec, one of the drugs that changed what a cancer diagnosis means. But on the day the graph came back clean, nobody could have known that. It was, as far as anyone in the room could prove, just a compound that stuck to the right thing and let go of the wrong things. A hit.

That word is going to do a lot of work in this article, so it is worth being careful with it from the start. In drug discovery, a hit is not a drug. It is not even close to a drug. A hit is the faintest possible signal that a molecule touches the target at all — a weak, early, easily-mistaken flicker that says maybe this one. Most hits are noise. Most hits are wrong. The clean graph that would one day be Gleevec was not a finish line. It was the starting gun for a decade of work, and the people who ran it knew that even as they ran it.

A Numbers Game Wearing a Eureka’s Clothes

We tell the story of drug discovery as a series of eureka moments because eurekas are easy to remember and easy to retell. Hit discovery is, before it is anything else, a numbers game — and the numbers are worth feeling in your gut rather than just reading.

Start with the scale. To find the molecules worth pursuing for a single target, a screening effort might test on the order of a million compounds. Out of that million, you might come away with something like a thousand that show any signal at all. Push those thousand through more careful, more skeptical assays and you might end up with ten or so that are real enough and good enough to call leads — molecules worth investing serious chemistry in. And out of those ten, on a good program, one eventually walks into a clinical trial as a candidate.

A million, to a thousand, to ten, to one. (Order-of-magnitude — the exact ratios vary widely by target and by method.)

Those numbers deserve more care than they usually get on a slide. They are order-of-magnitude figures, not constants. The funnel is a teaching shape, not a law of nature. The fraction of compounds that “hit” in a screen — the hit rate — varies enormously depending on what you’re screening against and how you’re measuring. For some targets and some assays it might be around a tenth of a percent, which is the number you’ll see quoted most often as a rough default. For others, with a friendlier target or a more permissive readout, it can run up toward two percent. The point of the funnel is not the exact ratio. The point is the shape: catastrophic, deliberate, expected attrition at every step, where the overwhelming majority of what you find turns out to be wrong, and you build an entire industrial process around the assumption that it will. Hit discovery is the part of the pipeline that has made peace with being mostly wrong on purpose.

This is the first thing to understand by the end of this stage, and it is the one most at odds with the stories we tell: the eureka, if it comes at all, comes after the numbers, not instead of them. Druker’s clean graph was a eureka in retrospect. In the moment, it was one signal among the thousands of signals that screening campaigns produce, distinguished only by being real, being selective, and being followed.

What You Are Actually Screening For

So what does it mean, mechanically, for a molecule to “hit”? What is the thing you are looking for in that ocean of compounds?

For a long time, the standard picture — the one nearly every textbook still reaches for — has been lock and key. The target — usually a protein — is the lock. The drug molecule is the key. A key that fits, turns; a key that doesn’t, doesn’t. It is a clean, satisfying image, and it has taught a great many people the basic idea that shape matters, that a molecule has to fit its target to do anything. (F-04) is where shape itself gets its proper treatment — why molecules have the three-dimensional forms they do, and why those forms are everything in this business.

But lock and key is wrong in a way that matters, and correcting it is the difference between understanding why hit discovery works and being baffled by why it’s so hard.

A real lock is rigid. Its pins sit in fixed positions; the key’s teeth are machined to match. Nothing about the lock changes when the key goes in. Proteins are not rigid. A protein in solution is a restless thing, flexing and breathing through a crowd of slightly different shapes from one microsecond to the next. When a small molecule approaches, two things can happen, and usually both do at once: the molecule can catch the protein in a shape it was already sampling and stabilize it (this is called conformational selection), and the protein can shift and remold itself around the molecule once contact is made (induced fit). The binding pocket closes around the drug like a hand closing around a held object — adjusting, settling, finding a grip.

So a better image is a hand and a glove that breathes. Not a key snapping into a fixed lock, but a glove that flexes as the hand enters, a hand that splays and curls to fill the glove, the two of them arriving at a fit that neither had on its own. (F-15) takes this all the way down into how occupancy and binding actually behave dynamically over time; for now the correction is enough. Binding is not a mechanical click. It is a negotiation.

This is not pedantry, and here is why it earns its place: if binding were lock-and-key, a hit would be a final answer. The key either fits or it doesn’t. But because binding is a breathing, adjustable affair, a hit is only a starting fit — often a clumsy, weak one, where the molecule grabs the target loosely and lets go too easily and grips three other things it shouldn’t. The entire next stage of the work, hit-to-lead, is the slow business of reshaping that molecule so the hand fits the glove better. Lock-and-key hides that work. Dynamic binding explains why it exists.

The Four Ways to Find a Hit

There is no single way to fish this ocean. Over the decades the field has built several, and they coexist because each catches things the others miss.

High-throughput screening (HTS) is the brute-force classic and the source of the funnel numbers above. You take a vast physical library — hundreds of thousands to millions of real compounds, sitting in tiny wells across thousands of plates — and you test them all against your target with a robotic assay that gives a simple readout: did the target’s activity change? (F-01) covers the cell-based and biochemical assays that make this readout possible. HTS is expensive, industrial, and indiscriminate, which is exactly its strength: it has no theory about what should work, so it finds things no theory would have predicted.

DNA-encoded libraries (DEL) cheat the numbers in a clever way. Instead of storing each compound in its own physical well, you attach a unique stretch of DNA to each molecule — a barcode — and let billions of barcoded molecules mingle in a single tube. Wash them over the target, keep whatever sticks, and read the DNA to find out what you caught. A library that would fill a warehouse as separate compounds fits in a drop of liquid as a DNA-tagged mixture. You trade some certainty for staggering scale.

Fragment-based screening goes the opposite direction: smaller, not bigger. Instead of testing complete drug-sized molecules, you test tiny fragments of molecules — too small to be drugs, too weak to do much on their own. A fragment that binds even feebly to a corner of the target is telling you something true about that corner. Then you grow the fragment, or stitch two neighboring fragments together, building a drug outward from a few footholds you know are real. It is the difference between trying ten thousand finished keys and instead studying the lock’s pins one at a time.

AI-generated hits are the newest entry, and the one most likely to be oversold to you. The promise is that instead of screening molecules that exist, you can have a model propose molecules that should bind — generating candidates computationally and synthesizing only the promising ones. So far that promise lives mostly in early pipelines rather than approved drugs, and the track record that would settle how much the AI itself contributes is still being written. The structural prediction tools that make some of this plausible are taken up in (X-02), on what AlphaFold actually does and doesn’t do for a medicinal chemist. The honest framing of the whole AI turn is in (X-05): the so-called “AI spring” in drug discovery is really a data spring — these methods are only as good as the screening data they learned from, which means the brute-force methods above didn’t get replaced, they got repurposed into training sets.

Cutting across all four is an older and deeper split: what are you screening against?

In target-based screening, you have already decided which protein you blame for the disease — say, the BCR-ABL fusion in CML — and you screen for molecules that bind or block that one target. It is rational and focused. It also assumes you picked the right culprit, and biology punishes that assumption often.

In phenotypic screening, you don’t commit to a target at all. You screen for molecules that produce an effect you want — cells stop dividing, a diseased phenotype reverses — and you figure out which protein got hit afterward, sometimes much afterward. It is humbler about how little we understand, and it has a long history of finding drugs whose mechanism nobody could explain until years later.

“A Hit” Means Four Different Things

Everything above quietly assumed the hit was a small molecule — a compact organic compound binding a protein. For most of this curriculum’s history that was the only kind of hit there was. It isn’t anymore, and one of the most useful things to carry out of this stage is that the word “hit” splinters into different meanings depending on what kind of medicine you’re making.

For a small molecule, a hit is what we’ve described: a compound, pulled from a library, that binds and modulates a target in an assay. That story lives in (M-SM-04).

For an antibody, a hit isn’t a compound you pulled off a shelf — it’s a binder your biology produced. You immunize an animal, or run a display technology that lets you sift enormous panels of antibody variants, and a hit is one that latches onto your target with the right specificity. The discovery logic is selection from a living or quasi-living repertoire, not screening a chemical catalog. See (M-MAB-04).

For an antisense oligonucleotide or siRNA, a hit barely involves shape-fitting at all in the same sense — it is fundamentally a sequence match. The molecule is a short stretch of nucleic acid designed to pair with a specific RNA message; finding a hit means finding the sequence (and the chemistry around it) that engages the right transcript without engaging the wrong ones. (M-ASO-04) takes this apart.

For a degrader — a PROTAC and its relatives — a hit is stranger still. You’re not looking for a molecule that merely binds and blocks; you’re looking for one that drags the target into the cell’s disposal machinery and gets it destroyed. A degrader hit is defined by an outcome — the target goes away — not just by binding. (M-PROTAC-04) is where that logic gets its due.

Four modalities, four definitions of the same word. And some medicines fold several of these together at once: an antibody-drug conjugate, the subject of (CS-ENHERTU-01), is in effect three discovery programs welded into one — an antibody hit, a small-molecule payload, and the chemistry that links them — each with its own version of “finding something that sticks.”

The Clean Graph Is a Beginning

Come back to the dish. Come back to STI571 and the two lines on the graph, one bending down where the leukemic cells stopped, one running flat where the healthy cells carried on.

What that graph established was the smallest possible true thing: this molecule touches this target, selectively, in a dish. Not that it would survive a body’s metabolism. Not that it could be dosed safely, or absorbed at all, or reach the cells that needed it in a living person. Not that it would work in a single patient. All of that — the entire remaining length of the pipeline, the years and the failures and the trials — was still ahead. The compound that became Gleevec is famous now partly because it survived all of that, and partly because, by the standards of the funnel, it nearly didn’t get made at all. The story of how this particular hit was found, and almost wasn’t, is (CS-GLEEVEC-02).

That is the shape of this stage, and it is worth holding onto as a corrective to every triumphant discovery story you’ll ever hear. The eureka is mostly a trick of memory, assigned after the fact to the one signal out of thousands that happened to be followed all the way home. A hit is a faint flicker against an ocean of noise, found by industrial-scale fishing and read by people who have made peace with being wrong almost every time. Binding is a glove that breathes, not a key in a lock, which is exactly why a fit that exists is not yet a fit that works. And the clean-looking graph — the one that makes the whole thing look easy in hindsight — is not the end of the search.

It is the quiet, unceremonious beginning of everything hard that comes next.

This is the free spine of The Lead Compound.

You’ve just read one stage of how a medicine actually gets made. The spine is the free through-line — the whole pipeline, start to finish. The full course goes deeper: every drug class (antibodies, mRNA, cell and gene therapy, peptides, and more) and the real, documented stories behind the medicines that defined them — Ozempic, Keytruda, Gleevec, Comirnaty, and dozens more.

← Previous (SP-03) · How to read this course · Next: (SP-05) →