AlphaFold and the Science We Can Actually Do Now
When AI skips the grunt work, researchers start asking different questions
Dorothy Crowfoot Hodgkin spent thirty years staring at insulin while the Beatles broke up, humans walked on the moon, and three generations of grad students wondered if they’d ever publish. This wasn’t failure. It was a hostage situation with physics, and we were all paying the ransom. Max Perutz spent twenty years on hemoglobin. These timescales weren’t consequences of incompetence. Proteins don’t crystallize on demand. X-rays don’t diffract themselves. The structure of a single protein could consume a career, and often did. This was how we did science: one hostage situation at a time.
What this reveals: for fifty years, we’ve been rationing scientific curiosity according to crystallization probability. The diseases that got studied, the proteins that got mapped, the questions that shaped molecular biology, all filtered through a single brutal constraint. Can we afford to wait? AlphaFold didn’t just solve protein folding. It exposed how much of science has been determined not by importance but by affordability.
You didn’t pursue a protein structure because you were curious. You pursued it because you could afford to pay the ransom. Five thousand to two hundred thousand euros per structure. Months to years. An eight percent success rate that would get you fired in any other industry. Graduate students calculated risk like Vegas bookies. Principal investigators calculated funding cycles like divorce lawyers timing their settlements. Entire research programs pivoted around a single question: will it crystallize? Meanwhile, diseases didn’t wait politely.
In November 2020, AlphaFold 2 won the CASP14 competition with a median accuracy score of 92.4 out of 100. The fifty-year-old protein folding problem wasn’t just solved. It was obliterated. By July 2022, the AlphaFold database contained two hundred million predicted structures. As of 2024, that number is two hundred fourteen million. For context, experimental structural biology produced roughly two hundred thousand structures over six decades.
The Questions We Couldn’t Afford
Matt Higgins’ Oxford team spent years chasing Pfs48/45, a malaria surface protein that should have been a vaccine goldmine. They had the sequence, fuzzy cryo-EM images, and a protein that refused to cooperate with traditional methods. The science was ready; the protein wasn’t.
This is where researchers stopped. Not because the question wasn’t interesting. Not because the vaccine wouldn’t be valuable. But because you couldn’t justify years more effort on a protein that might never crystallize.
Here’s what that reveals: malaria infects hundreds of millions annually, but vaccine development stalled at a measurement problem. Not “we can’t solve malaria” (we could). We just couldn’t afford the prerequisite step. The disease kept circulating while researchers moved to proteins that crystallized more reliably. Resource allocation determined which suffering got addressed, sorted by which measurement tools cooperated rather than which diseases mattered most.
AlphaFold said yes. “The use of AlphaFold was really transformational,” Higgins reported, “giving us a really sharp view of this malaria surface protein.” The team began testing vaccines in early 2023. The structure they couldn’t get for years appeared in minutes.
CDK20, a protein target for hepatocellular carcinoma, had no available inhibitors despite success with other members of its protein family. The likely reason: no structural information, which meant no reasonable way to design drugs against it. Using AlphaFold’s prediction, researchers achieved an initial hit compound in thirty days after synthesizing only seven molecules. Previously, this would have been prohibitively expensive, a bet few labs could afford to make.
What We Couldn’t Afford to Notice
Neglected tropical diseases have a name that reveals the problem: neglect. We named them after what we did to them, which is at least honest. These aren’t obscure conditions. Chagas disease affects millions in Central and South America, leishmaniasis circulates through tropical and subtropical regions worldwide. The WHO maintains a list of seventeen organisms causing diseases that disproportionately affect developing countries. The economics never worked, which is how we say “the suffering doesn’t occur in markets wealthy enough to generate pharmaceutical returns” without sounding quite so sociopathic. It’s almost impressive how we’ve turned suffering into a market failure rather than a moral one.
In January 2022, DeepMind and EMBL-EBI added protein structure predictions for all seventeen organisms on the WHO’s neglected disease list, plus ten more from the antimicrobial resistance list. Almost two hundred thousand new structures. The Drugs for Neglected Diseases initiative now has more than twenty new chemical entities in its portfolio and is working to enable researchers in low-income countries to play more active roles in drug discovery.
Turkish undergraduate students Alper and Taner Karagöl taught themselves structural biology during the pandemic using online AlphaFold tutorials. No formal training. No institutional resources. Fifteen research papers.
What does that reveal about what we were protecting? Traditional structural biology labs spent decades and millions fighting crystallization probabilities. These undergraduates bypassed the entire economy. Not just equipment access. Not just expertise. Decades of gatekeeping that claimed to protect intellectual rigor when it was actually rationing access based on who could afford the measurement overhead. The barrier was economic scarcity masquerading as scientific necessity (a velvet rope at the entrance to knowledge that turned out to be made of nothing but consensus).
The pattern holds: orphan GPCRs, viral proteins from non-model organisms, sustainability applications. Researchers identified novel PET hydrolases with 38.79-fold improvement in plastic degradation activity (proteins nobody would have prioritized for structure determination). Research funded by hypothesis rather than crystallization probability.
What Actually Changed
‘I alphafolded it’ now echoes through labs and thesis defenses. When nouns become verbs, we’ve normalized something before questioning it. Language betrays us first.
Give it five years. “I alphafolded it” will be what undergraduates say the way they say “I googled it.” The freshman biology major who hand-calculates protein structures will be viewed like the hipster who only listens to vinyl. There will be AlphaFold Premium (priority queue access), AlphaFold Lite (with ads). Researchers will joke about “raw-dogging crystallography” the way people joke about going off-grid.
The pattern reveals what we mean by “democratization”: not everyone gets the equipment, everyone gets the subscription. Access, not ownership. Results, not understanding. I ubered my commute. I airbnb’d my apartment. I alphafolded my protein structure. I optimized my life into a series of API calls to someone else’s platform.
This matters more than it sounds. Access without ownership means researchers can generate structures without understanding their limitations. It means dependence on platform stability, pricing models, and terms of service. It means the Turkish undergraduates can publish fifteen papers, but they’re building expertise on infrastructure they don’t control and methods they didn’t develop. When the bottleneck shifts from equipment to subscription, we’ve traded one form of gatekeeping for another. The new gate just happens to open faster and costs less per use, which makes it feel like freedom until you realize you’re still asking permission.
But this isn’t replacement. It’s restructuring. The subscription model is real, but the productivity gains are also real. Researchers using AlphaFold 2 experienced a forty percent increase in submission of novel experimental protein structures (evidence of exploration in uncharted areas). Research linked to AlphaFold 2 is twice as likely to be cited in clinical articles and patents. The database contains two hundred fourteen million predicted structures, but experimental structural biology hasn’t disappeared. Structural biologists have shifted their focus. Graduate students can now get domain boundary information for protein expression in hours rather than years. They can test structure-function hypotheses without multi-year commitments. Principal investigators can pursue risky targets without betting entire grants on crystallization success.
The bottleneck moved. It used to be: can we afford to determine this structure? Now it’s: what can we learn from this structure? That’s a different kind of science.
The Limitations That Remain
AlphaFold predicts static structures, betting on the winning horse while proteins dance, transform, and change shape when performing functions. It gives us a photograph when we need a film. A still image of a shapeshifter.
This matters profoundly for drug discovery. A MIT study tested whether AlphaFold structures could find drugs binding to bacterial proteins. They analyzed 296 essential E. coli proteins with 218 antibacterial compounds using molecular docking. The result: 0.48 AUC. Random chance.
The problem isn’t AlphaFold’s predictions. Proteins undergo substantial conformational changes upon ligand binding (the induced fit effect). Docking methods trained on bound structures struggle when predicting binding to unbound conformations. AlphaFold gives you accurate static structures, but drug binding requires modeling dynamics the tool wasn’t designed to capture.
Thirty to forty percent of the proteome comprises intrinsically disordered regions (proteins or protein segments that lack stable secondary and tertiary structures). Eighty percent of human cancer-associated proteins have long intrinsically disordered regions. These are critical for disease biology, yet fundamentally incompatible with AlphaFold’s approach. The algorithm can identify these regions through low confidence scores, but it cannot predict their functional conformations because those conformations are inherently variable. A solid background in structural biology and protein biochemistry remains essential for interpreting predictions.
What Protein Folding Reveals About Our Relationship to Scarcity
Here’s what this reveals about us: we’ve been constrained less by imagination than by logistics. The questions researchers are now pursuing (malaria vaccine targets, neglected disease proteins, plastic-degrading enzymes, orphan GPCRs) weren’t unknown or unimportant. They were just too expensive relative to available resources. We had rationed curiosity according to crystallization probability, which is a polite way of saying we let accounting departments direct scientific discovery.
What we call “scientific progress” in structural biology has been shaped less by importance than by tractability. Perutz’s hemoglobin and Hodgkin’s insulin became foundational partly because they succeeded, but equally important proteins were abandoned because they wouldn’t cooperate with x-ray diffraction. The history of molecular biology includes vast negative space: structures never pursued because the economics didn’t work.
AlphaFold doesn’t answer these questions. It just makes them visible. The database now contains structures for proteins that would never have received experimental attention. Some will prove crucial. Many won’t. But the fact that we couldn’t afford to check until now is the interesting part.
The rhetoric around AI often emphasizes replacement (automation eliminating human labor, algorithms displacing expertise). AlphaFold suggests a different dynamic. It eliminated grunt work, not thinking. Structural biologists didn’t become obsolete; they became more productive. The forty percent increase in structure submissions, the shift toward dissimilar and uncharted proteins: these suggest researchers doing more of what they were already trying to do, just without the multi-year overhead. Turkish undergraduates can generate structures, but expert structural biologists determine what those structures mean.
In 2033, graduate students will laugh at the idea of spending years trying to crystallize proteins. They’ll look at our methods the way we view medieval alchemy (quaint, inefficient, and slightly mad). What they won’t see is how our constraints shaped not just what we studied but how we thought about questions themselves.
The Science We Can Actually Do Now
The question “will it crystallize?” has largely disappeared from grant proposals. What replaced it is just different.
Now the questions are: Does this structure help us understand the disease mechanism? Can we design molecules against this target? What does this protein family tell us about evolutionary history? Does this enzyme have potential for plastic degradation? Which genetic variants affect structure in ways that matter for function?
These questions are harder. Not harder than crystallography (crystallography was hard because it was slow, expensive, and failure-prone). Harder because they require actual scientific judgment. “Will it crystallize?” was a logistics question. You could answer it with money, time, and grad student labor. The new questions require hypothesis formation, experimental design, interpretation of ambiguous results. They’re intellectually demanding rather than resource-constrained. AlphaFold didn’t make science easier. It made the easy part instant and exposed how much harder the actual thinking is.
AlphaFold accelerated answers to one question (what does this protein look like?), which revealed how many other questions we’d been implicitly avoiding because we couldn’t afford the prerequisite information. The malaria vaccine research didn’t fail before AlphaFold; it just couldn’t proceed past a certain point. The CDK20 inhibitor discovery wasn’t impossible; it was merely too risky. The neglected disease work wasn’t beyond reach; it was beyond budget.
Now it’s not. That changes the shape of scientific ambition. Graduate students can pursue structural biology projects without betting their entire thesis on crystallization success. Drug developers can explore targets previously considered undruggable. Researchers in low-income countries can participate in structure-based drug discovery without access to synchrotron facilities.
None of this makes science easy. It makes different science possible. The fifty-year-old protein folding problem is solved, but protein folding was never really the problem. It was a prerequisite. The actual problems (disease mechanisms, drug interactions, evolutionary relationships, functional dynamics) remain. We just have better starting conditions now.
Two hundred fourteen million structures. More than technical achievement. They’re a map of questions we never asked because we couldn’t afford the toll. Exploring that map (figuring out what these structures mean and do), that’s the science we can finally do. The question is whether we can afford what we’ll discover.
AlphaFold didn’t just solve protein folding. It revealed how much of science has been shaped not by importance but by affordability. The protein folding problem was solved not by better microscopes but by better mathematics, and that shifted the bottleneck from physical resources to intellectual ones.
How many other fields have their own invisible maps of questions rationed by tractability rather than importance? That’s the pattern worth watching.
Research Notes: AlphaFold and the Science We Can Actually Do Now
What happens when a field’s fundamental bottleneck just disappears?