SP-02 · Finding the Knot

In 1993, in a new lab at Oregon Health & Science University, a physician named Brian Druker went looking for a molecule that did not yet have a name he could use. He had spent years thinking about a single broken protein in a single kind of cancer, and he was convinced that if you could jam that one protein, you could stop the disease. What he needed was a chemist who had already built the jamming compound. So he reached out to Nick Lydon, a scientist at the pharmaceutical company Ciba-Geigy, and asked, in effect, whether anyone there had made something that shut down the enzyme he was obsessed with.

Lydon’s group had. They sent Druker a set of compounds to test. The most promising of them carried the workmanlike designation CGP 57148 — later renamed STI571, then imatinib, then sold to the world as Gleevec. It would become one of the most celebrated cancer drugs ever made, the proof-of-concept that you could turn a precise molecular insight into a pill that let people with a once-fatal leukemia live for decades.

That phone call is the kind of moment science likes to remember. It is clean, it has two named people, and it has an outcome you can point to. But the call is not where the target came from. The target — the thing Druker was so sure about — had been assembling itself in the scientific literature for more than thirty years before either man picked up a phone. That gap, between the tidy story and the long messy truth underneath it, is what this article is about.

What a Target Actually Is

Stage 1 (SP-01) introduced the target as the molecular handle a disease depends on; the word does more work than that first pass let on. In drug discovery, a target is a specific molecule in the body — almost always a protein — that you believe is so central to a disease that changing its behavior will change the disease itself. Block it, boost it, degrade it, reroute it: the bet is that if you act on this one molecule, the patient gets better.

Notice the verb. Believe. A target is a hypothesis wearing a noun’s clothing. When a company “nominates a target,” it is not announcing a fact about nature; it is placing a wager that a particular causal node — one knot in the tangled wiring of a cell — is holding the disease together, and that pulling on it will make the whole snarl loosen rather than tighten somewhere else.

This matters because the language hides it. People say “we found the target” the way they’d say “we found the leak,” as if the molecule were sitting there waiting to be discovered, labeled CAUSE in the body’s plumbing. It almost never works that way. Diseases are usually the product of many molecules pushing on each other, and picking one to attack is a judgment call about which push matters most. Proteins, recall, are the cell’s machines — the workers that build, signal, cut, and carry (see F-01) — and they get built from the instructions in our genes, which are less like blueprints and more like recipes the cell follows under changing conditions (see F-03). A target is a guess about which machine, among thousands, is the one to break.

So when you read that a biotech has “a promising new target,” translate it as: they have a promising new bet about cause. The rest of drug discovery is the long, expensive process of finding out whether the bet was right.

The Knot That Took Thirty Years

Here is the better way to think about where targets come from, told through the one Druker chased.

In 1960, two researchers in Philadelphia, Peter Nowell and David Hungerford, noticed something strange under the microscope. In the cells of patients with chronic myeloid leukemia — CML, a cancer of the blood — one chromosome was conspicuously too short. They did not know why, or what it meant, or whether it caused the disease or merely came along with it. They just saw a stub where a full chromosome should be. It got named for the city: the Philadelphia chromosome. Today we know that roughly 90 to 95 percent of CML patients carry it. In 1960, it was an anomaly with no mechanism.

It took thirteen more years to learn what the anomaly was. In 1973, Janet Rowley, working with better staining techniques that let her tell chromosomes apart by their banding patterns, showed that the missing piece wasn’t missing at all. Two chromosomes, numbers 9 and 22, had broken and swapped ends — a translocation, written t(9;22). The short chromosome was the leftover of a trade.

A swap of what, though? Through the 1980s, molecular biologists worked out the answer at the level of genes. The break fused two genes that should never touch: a gene called BCR on chromosome 22 and a gene called ABL on chromosome 9. The fusion, BCR-ABL, was a chimera — half of one gene welded to half of another — and it produced a protein that did not exist in healthy cells. That protein was a tyrosine kinase, a type of molecular switch that tells cells to grow and divide. Normally such a switch flicks on only when the cell receives the right signal. The fused version was jammed permanently in the ON position — constitutively active, in the jargon — screaming “divide” with no off signal. That relentless growth signal is what drives the leukemia.

Now line up the dates. The chromosome: 1960. The translocation: 1973. The fused gene and its hyperactive kinase: through the 1980s. A drug that finally blocked that kinase in patients: the late 1990s, with imatinib reaching FDA approval in 2001. From first clue to first pill, more than forty years.

That sequence is the correction this article exists to make. We tell target stories as eureka moments — “they discovered that BCR-ABL causes CML” — collapsed into a single sentence with a single subject. But no one discovered the target in the way you discover a coin on the sidewalk. A target is a multi-decade accumulation of converging evidence, narrated as a single flash of insight only in hindsight. Nowell and Hungerford saw a short chromosome. Rowley saw the swap. The molecular biologists named the gene. Each generation handed the next a slightly sharper version of the same question. By the time Druker called Lydon in 1993, “BCR-ABL is the target” was not a hunch — it was the compressed conclusion of thirty-three years of other people’s work, each layer adding causal weight until the bet looked safe enough to stake a career on.

This is what “target identification” really produces: not a discovery, but a case — a stack of evidence pointing at one molecule, built by many hands over a long time.

The Several Doors into a Target

If targets are bets assembled from evidence, then the obvious next question is: where does the evidence come from? And here the field splits, because there is no single method called “target identification.” There are several distinct trades that all answer to that name, and they barely resemble one another.

Human genetics is the route the BCR-ABL story followed, and it remains the most trusted. The logic is simple and powerful: find a genetic change that travels with a disease across many people, and you have caught nature running the experiment for you. If a broken gene reliably produces an illness — or, just as usefully, if a broken gene reliably protects people from one — then the protein that gene encodes is a strong candidate for a target. The Philadelphia chromosome is the classic case: a specific genetic lesion, present in almost every patient, encoding a specific overactive protein. You could hardly ask for a cleaner finger pointing at a cause.

Phenotypic screening comes at it from the opposite end. Instead of starting with a molecule, you start with an effect. You take thousands or millions of chemical compounds, expose diseased cells (or whole organisms) to them, and watch for the ones that produce the change you want — the cancer cell that dies, the inflamed tissue that calms. You find a compound that works before you know why it works. The target, in this approach, is reverse-engineered afterward: having found a compound that helps, you go hunting for the molecule it acts on. The job is “figure out what this drug is hitting,” which is nearly the reverse of the geneticist’s job.

Chemical biology lives in the same neighborhood — using small molecules as tools to interrogate which protein, when perturbed, moves the disease — and shades into the screening world.

Computational and AI approaches are the newest door, and the most over-hyped. Here the work is to mine enormous piles of data — genome sequences, records of which genes switch on in which tissues, maps of which proteins touch which other proteins — and let algorithms surface molecules that look causally important. To a machine-learning engineer, “target identification” is a problem of ranking and prediction across a giant dataset. To the geneticist down the hall, it is a problem of inheritance and mechanism. They use the same phrase and mean almost different professions — which is exactly the gap explored in X-01.

It is worth being honest about that last door. Much of what gets sold as an “AI revolution” in drug discovery is better understood as a data revolution: the algorithms got useful largely because the underlying biological data got vast and cheap, not the other way around — a reading the field still argues over (see X-05). The cheapness of the data is about to become the whole story.

The choice of door also depends on what kind of medicine you eventually intend to make, because each treatment type asks a different question of the same target. A small-molecule pill and an antibody and a gene therapy do not “see” a target the same way. An antibody can only reach proteins on the outside of cells or floating between them, so its version of target ID is constrained to that surface (see M-MAB-02). A drug made of genetic material — an antisense oligonucleotide — aims not at a protein but at the RNA message that the cell would otherwise translate into that protein, which means its target is chosen one step upstream (see M-ASO-02). And the gene therapies, whether a virus delivering a working copy of a gene (M-AAV-02) or a CRISPR system editing the genome inside the body (M-CRISPR-02), reframe “target” yet again — toward the DNA itself. Same word, four different jobs. Which disease you are even working in shapes the menu before you start (see F-16).

The Firehose

For most of the BCR-ABL story, evidence was scarce and slow. Reading a single gene was a feat. This is the part that has changed beyond recognition, and it changes the whole economics of nomination.

Consider the price of reading a human genome. The Human Genome Project — the first complete read of human DNA — was projected to cost about $3 billion and was finished in April 2003 for $2.7 billion, ahead of schedule. That was one genome, the product of a thirteen-year international effort. Then the cost collapsed. Sequencing a single genome dropped to roughly $100 million in the early 2000s, to about $600 by the late 2010s, to somewhere around $200 today on the highest-throughput machines. A read that once cost as much as a major public-works project now costs less than a pair of shoes.

When the cost of generating evidence collapses by that many orders of magnitude, the consequence is a flood. Genomes, the activity of genes across tissues, catalogs of mutations in tumors — all of it pours in faster than anyone can think about it. Each dataset throws off new associations, new genes that travel with disease, new molecules that might be causal knots. Target nominations have become cheap and abundant. The firehose is on.

And here is the thing the firehose does not do: it does not tell you which nominations are right. This is the hinge of the whole article, so it is worth slowing down for. There are three different things hiding inside the casual phrase “we have a target,” and the cheap genome only helps with the first.

Nomination is not validation. Producing a plausible candidate — a molecule that statistically tracks with a disease — is now easy. Proving that acting on that molecule actually changes the disease is hard, slow, and largely immune to the cost savings. A correlation in a dataset is a suggestion, not a cause.

Nomination is not druggability. Even a genuinely causal target may be a molecule no medicine we know how to build can grip. Of the roughly 20,000 human protein-coding genes, only about 667 proteins are the targets of approved drugs, by one careful count; broader estimates put perhaps 4,500 proteins — around 22 percent of all human proteins — within reach of some conceivable drug, depending on how generously “druggable” is defined. The rest are, for now, untouchable: correctly identified as important, and impossible to act on.

You can see the squeeze in the output numbers. The FDA approves, on a rolling ten-year average, roughly 46 genuinely new drugs a year — some years as few as 37, some as many as 59. That figure has crept upward over the decades, but it has not exploded. Meanwhile the cost of generating target hypotheses has fallen by orders of magnitude. Set those two trends side by side and the gap is the whole problem of modern drug discovery: nominations have become almost free, while turning a nomination into an approved medicine has stayed brutally expensive. Cheaper data made it dramatically easier to guess. It did not make the guesses true, and it did not make the true ones reachable.

This is the right place to retire the leak-in-the-plumbing image for good. A target is not a thing you locate. It is a claim you build and then have to defend — and the genomic firehose has made the building cheap without making the defending any easier.

The Question This Hands Forward

Return to Druker for a moment. By 1993 the BCR-ABL bet looked unusually safe: a single fusion protein, present in nearly every CML patient, with a clear mechanism and thirty-three years of converging evidence behind it. That is about as validated as a target gets before anyone has actually tested a drug in a person. And even then, plenty of smart people doubted that hitting one kinase could treat a cancer — the worry was that cancer was too redundant, too many backup pathways, for a single switch to matter. They were wrong, in this case. But the BCR-ABL arc is famous precisely because it is rare. It is the nomination that panned out, told against a base rate where most do not — where the knot you were so sure about turns out to be tied to nothing, or tied to too much. (The full story of that one knot, and why this particular translocation was such an unusually clean target, is its own thread; it begins in the knot in the translocation. For a target found by an almost opposite route — not a broken gene driving a cancer, but a brake the cancer had learned to press — see the brake on the brake.)

Which leaves us exactly where Druker stood with that vial of CGP 57148 in hand, and where the next article begins. He had a target — a beautifully argued bet about cause. He had a compound that shut it down in a dish. What he did not yet have was the answer to the only question that ultimately matters:

Is the knot real? Will pulling on this molecule, in an actual human body, actually loosen the disease — or will the body route around it, or break in some way no one predicted?

Nominating a target is making the case. Finding out whether the case is true is a different job entirely, with its own methods, its own failures, and a far higher bar. That job is called target validation, and it is where we go 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-01) · How to read this course · Next: (SP-03) →