The shape of cures to come: AlphaFold’s structural revolution in drug discovery

The decades-long puzzle of protein folding

In the pantheon of scientific challenges, few have loomed as large or as stubborn as the protein folding problem. First posed in the 20th century, it asked how a mere string of amino acids — a one-dimensional sequence — can spontaneously contort into a precise three-dimensional shape. For decades, answering that riddle was akin to finding gold nuggets by panning endless rivers.

Biologists toiled with X-ray crystallography and cryo-electron microscopes, painstakingly determining one structure at a time. The stakes were high: protein shapes dictate what they do in the body, from ferrying oxygen in blood to triggering biochemical reactions (quantamagazine.org). Cracking the folding code promised profound payoffs for medicine and biology (quantamagazine.org) – illuminating the causes of diseases, revealing new drug targets, and enabling the design of novel therapies. Yet progress was slow. Generations of eminent scientists made only baby steps; predicting structure from sequence remained, as one observer put it, a holy grail just out of reach.

By the 1990s, an international competition called CASP (Critical Assessment of Structure Prediction) became the biennial Olympics of protein folding. Computational teams would pit their algorithms against unseen protein structures, usually scoring modestly. Then, in late 2020, something extraordinary happened. A relatively new entrant from Google’s DeepMind arrived and blew away the field. AlphaFold2, as it was known, achieved levels of accuracy that left seasoned researchers in shock (quantamagazine.org). Its models of protein 3D structures were correct to within atoms, scoring over 90 on the Global Distance Test – a leap five times better than the nearest competitor.

John Moult, CASP’s co-founder, declared that AlphaFold had “largely solved” the decades-old problem, heralding not an end but a beginning (quantamagazine.org). In one Zoom presentation, computational biology witnessed its “Eureka” moment. The breakthrough was so momentous that, just four years later, the architects of AlphaFold received the 2024 Nobel Prize in Chemistry for “developing a game-changing AI tool” (nature.com) – sharing the honor with a pioneer of protein design. It was the first Nobel ever to reward an achievement made possible by artificial intelligence, underscoring how profoundly this feat has transformed science.

From AI curiosity to structural savant: AlphaFold 2 and 3

What enabled AlphaFold’s tour de force? In essence, DeepMind’s researchers infused machine learning wizardry into protein biology. Traditional methods like homology modeling (borrowing structures from similar proteins) or physics-based simulations had inched forward but hit data and computing limits. AlphaFold2 took a different approach: it treated protein sequences as a language and their evolutionary relationships as clues. Using a neural network architecture heavily leveraging transformer models (the same AI engines behind modern language models), AlphaFold processed multiple sequence alignments and learned to predict which amino acids pair up in space (nature.com).

It iteratively refined a “prediction” of the protein’s shape, essentially imagining the folding process end-to-end. The result was an algorithm that, given any protein sequence, outputs a 3D structure often indistinguishable from one solved in a lab. In technical terms, AlphaFold2’s structures frequently achieve backbone accuracy above 85–90% (measured by the GDT score) even for challenging proteins (nature.com) – a jaw-dropping level of precision that turned skeptics into believers overnight.

AlphaFold2’s successor has since arrived on the scene: AlphaFold 3. Unveiled in 2024, this latest iteration expanded the AI’s repertoire from single proteins to the entire cast of molecular life. Not content with predicting lone protein structures, AlphaFold3 can now model complexes – situations where proteins bind DNA, RNA, small-molecule drugs or each other (genengnews.com).

Under the hood, it swaps out some of AlphaFold2’s engine for a new diffusion-based architecture, reflecting the rapid evolution of AI techniques (nature.com). The payoff is a tool that doesn’t just fold proteins in isolation but assembles molecular machines in silico. AlphaFold3 can, for example, predict how a regulatory protein wraps around DNA or how a candidate drug snugly nestles into an enzyme’s pocket. In benchmarks, it has vastly outperformed earlier specialized software for protein–ligand interactions, correctly placing small molecules in binding sites about 58% of the time – a remarkable success rate given the complexity of those predictions (collaborate.princeton.edu).

Demis Hassabis, DeepMind’s co-founder, hailed the upgrade, saying it can model “nearly all of life’s molecules” with cutting-edge accuracygenengnews.com. Put differently, AlphaFold’s brain has grown from a solitary puzzle-solver into a master Lego architect of biology, capable of virtually assembling the pieces of living systems.

There is, however, a catch: unlike its predecessor, AlphaFold3 was not open-sourced. DeepMind released it via a controlled web server, not as downloadable code, sparking an outcry among researchers (genengnews.com). Academic scientists, who had grown accustomed to tweaking AlphaFold2’s open code for their own experiments, bristled at being handed a black box. Over a thousand signed a protest letter complaining that they couldn’t test or reproduce AlphaFold3’s claims.

The server itself came with restrictions aplenty – initially only 10 predictions per day (later raised to 20) and a narrow menu of allowed ligand molecules (genengnews.com). Frustrated posts on social media noted that small custom drug molecules couldn’t be input at all. In effect, DeepMind and its sibling company Isomorphic Labs have kept AlphaFold3 on a short leash, limiting its use to non-commercial research and reserving the full power for their own drug discovery projects and partnerships. The contrast with AlphaFold2’s free-for-all availability is striking, and it speaks to the shifting sands of science and commerce – an issue we will return to. But first, it’s worth exploring how AlphaFold’s prowess is being put to work around the world.

Mapping the protein universe for all to see

If AlphaFold is the virtuoso, then the AlphaFold Protein Structure Database is its grand concert hall – one with open doors and free admission. Launched in mid-2021 as a partnership between DeepMind and Europe’s EMBL-EBI institute, this online repository began by posting AlphaFold’s predicted structures for the human proteome and several other organisms, about 350,000 proteins in all. Scientists were agog at the immediate impact: obscure enzymes and orphan receptors, previously known only by cryptic gene codes, suddenly had three-dimensional faces. And that was just the overture.

By the summer of 2022 the database expanded in a dramatic way, publishing around 200 million protein structures, essentially every protein known to science with a sequenced gene (nature.com). In the poetic phrasing of DeepMind, AlphaFold had charted “the entire protein universe” (nature.com). Researchers in need of a protein’s shape could now obtain it as easily as a Google search – no tedious lab experiments or costly crystallography required.

Crucially, this trove has democratized access to structural biology. Before AlphaFold, determining a protein structure was often the preserve of wealthy labs or pharma companies, given the expensive equipment and expertise needed. Now a grad student in a modest lab, or a biotech startup in a co-working space, can pull up high-quality structures on their laptop in seconds. The effect on drug discovery has been energizing. The PDB (Protein Data Bank), the long-standing archive of experimentally solved structures, contains about 200,000 entries accumulated over decades.

AlphaFold’s database is three orders of magnitude larger, covering proteins from Mycobacterium tuberculosis to the banana tree – a cornucopia for picking new drug targets. Even though these are predicted models, their accuracy is high enough in many cases to serve as the basis for virtual drug screening or rational design. As one review noted, the AlphaFold database helps meet the “vigorous demand” for structures in modern drug discovery, far beyond what the limited PDB could offer (nature.com). In short, the gates that once restricted structural data have been flung open. A researcher hunting a cure for a neglected disease or a company optimizing an enzyme can both drink from the same firehose of structural information. It is hard to overstate how much this leveling of the playing field resembles a paradigm shift – akin to going from snail-mail to the internet era in terms of data accessibility.

From code to clinic: AlphaFold accelerates drug design

The ultimate test of AlphaFold’s impact is whether it can speed up the making of medicines – and here the signs are promising, albeit sprinkled with reality checks. In the two years since AlphaFold2’s debut, labs and companies have eagerly applied its structures to structure-based drug design, the craft of crafting molecules that fit proteins like keys in locks. Medicinal chemists traditionally rely on experimentally-solved structures of targets (say, a cancer-causing kinase or a viral enzyme) to design inhibitors. But many important proteins had unknown structures, leaving drug hunters groping in the dark. AlphaFold changed that overnight by supplying modeled structures for just about anything with a sequence. Suddenly, previously “undruggable” targets became visible.

For example, scientists quickly used AlphaFold to model human proteins implicated in cancer that had eluded crystallography. One team examined a protein called WSB1, involved in tumor growth, by predicting its 3D shape with AlphaFold and identifying pockets that could bind a drug (nature.com). Another group took aim at Mycobacterium tuberculosis, the bacterium behind TB, and leveraged AlphaFold models of its enzymes to guide the design of new inhibitors. Across academia and industry, project after project began reporting that AlphaFold-filled knowledge gaps in their workflow.

Consider the case of neglected tropical diseases, illnesses that afflict millions in poorer regions but receive scant R&D investment. Leishmaniasis, a parasitic disease, is one example. A team working with the non-profit DNDi had a promising molecule against Leishmania parasites but didn’t know what protein it hit – a classic needle-in-haystack problem. Using AlphaFold, they homed in on an enzyme the parasite needs to survive, predicted its structure, and even modeled the human version of that enzyme for comparison (deepmind.google). This revealed how their molecule was interacting with the parasite’s enzyme and how it could be tweaked to avoid human proteins.

In essence, AlphaFold lit up a previously dark corner of the parasite’s biochemistry, reviving a stalled drug lead. “Before, we couldn’t look at how our compound bound and say ‘add a carbon here, remove a nitrogen there’ – that was off-limits,” one DNDi chemist recounted. “Except, now, it isn’t” (deepmind.google). For diseases like Chagas, leishmaniasis or malaria – where research funds are limited – having free, accurate protein models is a boon. It allows scientists to perform virtual experiments that would otherwise require expensive equipment, leveling the field between high-resource and low-resource settings (deepmind.google).

AlphaFold has also been conscripted in the fight against infectious diseases more broadly. During the COVID-19 pandemic, early AlphaFold models of SARS-CoV-2 proteins helped researchers identify potential binding sites for drugs and antibodies, complementing lab-based structure efforts. In antibacterial research, teams like one at MIT have used AlphaFold structures of bacterial proteins to run virtual screens for new antibiotics (news.mit.edu). James Collins, a biotech pioneer, noted that such AI-driven approaches could uncover antibiotics with unprecedented mechanisms to outwit resistant bacteria.

In one study, however, Collins’ group also highlighted a sobering lesson: using AlphaFold models for large-scale drug screens is not plug-and-play. They tried predicting which of hundreds of compounds would bind dozens of E. coli proteins, but initial docking simulations performed barely better than random guessing (news.mit.edu). The culprit, they suspect, is that these models (AlphaFold’s or even experimental ones) provide a single static snapshot of a protein. In real life, proteins jiggle and flex; a binding pocket might open only when the protein breathes. So, while AlphaFold gave us the parts, we still need better tools to simulate how those parts dance together in a living cell.

The MIT team improved accuracy by applying machine-learning corrections to the docking results (news.mit.edu), but the episode is a cautionary tale: having the map doesn’t mean we’ve conquered the territory. It underscores how AlphaFold, powerful as it is, becomes even more powerful when paired with other computational methods and experimental validation.

Meanwhile, a new frontier beckons: de novo protein design – creating brand-new proteins that perform useful tricks (like therapies, enzymes or nanomaterials). Here AlphaFold plays a different role, essentially as an AI structural guru that tells designers whether their invented protein sequences will fold up as intended. Pioneering labs (such as David Baker’s group, who shared that Nobel Prize) have used deep learning to “hallucinate” novel proteins and then run AlphaFold to see if the hallucination is plausible.

The result has been artificial proteins that actually work in the lab, from enzymes that munch plastic waste to novel therapeutics (deepmind.google). AlphaFold’s predictive accuracy gives designers confidence to leap into the unknown, accelerating a field that was once sluggish and hit-or-miss. In one instance, researchers designed plastic-degrading enzymes that were much faster than their natural counterparts, guided in part by AlphaFold confirmations of the enzyme structure after each design tweak (deepmind.google).

In another, scientists created minibinder proteins that glom onto targets like SARS-CoV-2’s spike protein, as precursors to new antiviral drugs – they generated candidates by the thousands using generative AI, then filtered them with AlphaFold to pick the ones most likely to fold and bind correctly. Such feats have led to a flurry of biotech startups focused on AI-driven protein design, confident that an AlphaFold-augmented pipeline can churn out therapeutic proteins or catalysts in weeks rather than years. For investors and pharma executives, this is an enticing prospect: the ability to digitally prototype drugs and biologics before ever moving to costly wet labs. No wonder DeepMind’s own spin-off, Isomorphic Labs, explicitly aims to marry AlphaFold’s capabilities with drug discovery, and companies from Insilico Medicine to Generate Biomedicines are raising hefty capital on similar promises.

When AlphaFold meets the wet lab: a new synergy

Rather than rendering laboratory scientists obsolete, AlphaFold has in many ways become their trusted ally. Structural biologists now routinely use AlphaFold to guide experiments, merging AI predictions with empirical data in a tag-team effort. A stunning example is the recent solving of the gigantic nuclear pore complex – the cellular gatekeeper that controls traffic into the nucleus. This massive assembly, about 120 million daltons in size and composed of ≈1,000 protein pieces, was like a 3D puzzle with half its pieces missing.

By 2021, cryo-electron microscopy had yielded only a partial structure (~50% complete) with many gaps. Enter AlphaFold2: by predicting previously unknown structures of numerous nucleoporin proteins, it provided the missing pieces needed to complete the puzzle (embl.orgembl.org). Researchers at EMBL and Max Planck combined cryo-EM maps, AlphaFold models, and integrative modeling software to fit together an almost complete nuclear pore architecture – covering over 90% of the core structure, up from 46% before. In the words of one scientist, “You cannot assemble a puzzle when you don’t know what the pieces look like. AlphaFold2 gave us the shapes” (embl.org). They even used a community-modified version of the AI (ColabFold) to predict how the pieces interact, arranging the pore’s subunits into a coherent whole.

Combining AI predictions with cryo-electron microscopy allowed researchers to solve the massive nuclear pore complex. The image shows the human nuclear pore core structure before (left) and after (right) integrating AlphaFold2 models, going from ~46% to >90% complete (embl.orgembl.org). Such AI-augmented assemblies reveal how proteins come together in cells, aiding drug discovery for complex targets.

This synergy between AI and experiment is being replicated across structural biology. Crystallographers facing a tricky structure now often turn to AlphaFold for molecular replacement – using the AI model as a starting point to solve X-ray diffraction phases. In many cases, what once required months of labor (or resorting to experimental phasing methods) can be cracked in a matter of hours with an AlphaFold model as the search template.

A survey in 2022 showed AlphaFold models could solve the crystal structures of dozens of proteins that previously stymied standard methods (journals.iucr.org). Similarly, NMR spectroscopists use AlphaFold predictions to interpret ambiguous data and accelerate structure determination. Far from putting structural biologists out of work, the AI has become the ultimate lab assistant – tireless, exceptionally knowledgeable, and unfazed by complexity. As one practitioner quipped, “AlphaFold won’t run your gel or align your beamline, but it sure tells you if those hours will be worth it.” By integrating predictions into the experimental workflow, scientists can focus their wet-lab efforts where it counts, testing hypotheses that emerge from the AI’s suggestions.

This is particularly useful for multicomponent assemblies, membrane proteins, or fleeting complexes that are challenging to capture experimentally. Even molecular dynamics (MD) simulations – long used to study protein motion – now often start from an AlphaFold structure when no crystal structure is available, giving a far more realistic initial model to simulate. MD can then explore the conformational dynamics that AlphaFold’s static snapshot doesn’t show, generating insights into different states, flexibility, and allosteric changes. Indeed, early studies indicate that running MD on AlphaFold models can reveal functionally important motions and even improve ligand docking predictions, as the protein is allowed to “breathe” beyond the single conformation (news.mit.edu).

In short, rather than AI replacing experiments, we are seeing a merging of the two into a powerful virtuous cycle. One may dub it a new “man + machine” model for drug discovery: AI generates a hypothesis (structure or interaction), experiments verify and refine it, and the refined data in turn inform better AI models. This synergy is accelerating the pipeline from target identification to lead compound optimization. A case in point: DeepMind’s AlphaMissense, a related AI model, uses AlphaFold-derived structural context to predict which genetic mutations are likely disease-causing – blending computation with medical genetics in ways that could streamline drug target validation. And pharmaceutical giants, historically cautious about unproven tech, are now fully onboard the AI train (or at least fearful of missing it).

The business of folding: from open science to competitive advantage

It is telling that what began as an open scientific endeavor has quickly entangled with commercial interests. The initial release of AlphaFold2 in 2021, complete with open-source code and a free database, was lauded as a triumph of open science – a gift to humanity. But as the implications for drug discovery became evident, the tone subtly shifted. DeepMind, which had invested vast resources in the project, spun out Isomorphic Labs to apply AlphaFold in pharmaceutical development (with the not-so-subtle implication that profits would eventually follow). By the time AlphaFold3 emerged, DeepMind chose a more guarded approach, likely mindful of the technology’s value.

The tension between openness and monetization was laid bare by the backlash to AlphaFold3’s limited release (genengnews.com). Many in academia grumbled that a powerful tool developed with the help of academic data (e.g. public protein structures and sequences) was now partly behind corporate walls.

Big Pharma, for its part, is not passively watching from the sidelines. Several heavyweight firms have inked collaborations with Isomorphic Labs or directly with DeepMind to gain early access to AlphaFold’s advanced versions. More intriguingly, a consortium of companies including AbbVie, Johnson & Johnson, Sanofi, and Boehringer Ingelheim recently took matters into their own hands – contributing their private troves of experimental protein structures (thousands of them, never seen in public databases) to help train a new AI model dubbed OpenFold (linkedin.com).

This initiative, coordinated via a federated learning platform, aims to create an AlphaFold-like tool with pharma’s insider data, explicitly tailored for drug discovery needs. The irony of the name “OpenFold” is not lost on observers: the project’s code may be open-source, but the superior version trained on proprietary data will stay behind closed doors, available only to the contributing companies. In other words, pharma is building its own private AlphaFold club. The motivation is clear – in drug development, a predictive edge can translate to huge competitive advantage (and by extension, future revenue). If adding confidential structural data on, say, a novel cancer target or a discontinued drug program can make the AI even 5% more accurate at predicting binding modes, that could be the difference between hitting or missing a viable drug candidate.

This raises a broader question: will the AI-for-drug-discovery boom turn scientific tools into trade secrets? Some worry that the spirit of global collaboration seen in AlphaFold2’s release might not last. For now, the original AlphaFold2 code and the vast database remain open, enabling countless discoveries by those who could never afford their own AI.

But as new and improved models (AlphaFold3 and beyond) come on line, access may depend on who you work for or how deep your pockets are. There is precedent in other fields: AI models for things like language or image generation started openly, then increasingly became proprietary as their capabilities grew lucrative. Investors might cheer this trend, eyeing returns from exclusivity, but scientists caution that biology moves fastest when knowledge is shared, especially in areas like pandemic response or neglected diseases where no single actor has incentive to invest alone. A brewing debate pits the ideal of open science against the reality that AI models cost millions to develop and thus will be monetized to recoup investments.

While the algorithms can be kept secret, the scientific insights ultimately cannot be walled off – a drug developed using AlphaFold3 still has to publish its trial results. Nonetheless, we are entering an era where having the best protein-prediction AI could become a strategic asset for companies, much like having the fastest supercomputer or the largest compound library.

Remaining mysteries: beyond static structures

For all its accomplishments, AlphaFold (and its cousins) still faces important limitations that present both technical challenges and business opportunities. One is protein dynamics: proteins are not statues but dancers, switching poses as they do their jobs. AlphaFold typically predicts a single, likely the most stable, conformation of a protein. But drugs often need to latch onto a transient shape – for instance, an enzyme’s open state versus closed state.

Allostery, where a molecule binding one site subtly reshapes a distant site, is fundamental to biology (and to many drug mechanisms, like turning an enzyme off by binding a remote “off switch”). A static AlphaFold model won’t reveal how an effector binding over here causes a loop over there to wiggle open. In industry speak, “conformational flexibility” remains a frontier. Advanced users are addressing this by generating ensembles of AlphaFold predictions or integrating them with molecular dynamics simulations to approximate the range of motion (news.mit.edu).

Some research groups are even developing algorithms to predict not just one folded structure, but a spectrum of conformations an intrinsically floppy protein might adopt (nature.com). Still, a full solution to accurately predicting conformational ensembles and allosteric shifts – essentially marrying thermodynamics with AlphaFold’s prowess – has yet to be achieved. When it is, it could truly revolutionize allosteric drug design, unveiling cryptic pockets that only exist in certain states.

Another limitation is the handling of non-standard biology. AlphaFold was trained predominantly on canonical proteins made of the 20 standard amino acids. Real biology is messier: many proteins have post-translational modifications (phosphates, sugars, methyl groups) or incorporate uncommon amino acids like selenocysteine – these were largely outside AlphaFold’s training scope. AlphaFold3 took a step in this direction by allowing “modified residues” and even small cofactors or ions in its predictions (nature.com).

But if you ask it to predict a protein with, say, a bulky glycan chain or an unnatural amino acid from synthetic biology, you’re likely out of luck. Similarly, membrane proteins in weird lipid environments, or multi-domain monsters with flexible linkers, can still confuse the AI or result in lower confidence predictions. There’s also the matter of protein-protein interactions in cells: AlphaFold-Multimer (a version of AF2) can predict complexes to a degree, but it doesn’t reliably sort out which of the thousands of possible pairings in a cell actually occur. So while AlphaFold can tell you the shape of individual puzzle pieces with astonishing accuracy, figuring out how all the pieces of the cellular puzzle fit and move – the orchestra beyond the soloist – is a grand challenge ahead.

Finally, it’s worth noting what AlphaFold doesn’t do: it doesn’t predict function per se. Structure is often an excellent guide to function (shape dictates function, as the dictum goes), but it’s not infallible. We still need biochemical experiments to confirm what a protein does and biological assays to confirm a drug actually works in cells or organisms. AlphaFold gives a head start – sometimes saving years of trial-and-error – but it’s not a miracle maker that eliminates the painstaking steps of drug development.

As any biotech investor knows, a lead compound still has to navigate toxicity, delivery, regulatory hurdles, and more. The optimism around AlphaFold has to be grounded in this reality: it solves one bottleneck (and brilliantly so), but the “bench-to-bedside” journey for a drug is long and complex.

AlphaFold might be best described as a phenomenal new pickaxe in the gold mine of drug discovery – greatly expediting the digging, but not guaranteeing that every strike finds gold.

Conclusion: a new era unfolds

In 2025, we stand at a remarkable inflection point. Artificial intelligence, in the form of AlphaFold and its progeny, has transformed protein science from a laborious art into a high-speed digital enterprise. Drug discovery, especially structure-based design, is arguably the most immediately rewarded. Life scientists now navigate a world where experimentalists and algorithms collaborate: AI predicts a protein’s contours, humans test and refine, and together they iterate toward new cures.

The language of drug R&D is peppered with phrases once confined to computer science – models, inference, training data, code release. Investors are taking note too. Dozens of startups are leveraging AlphaFold (and often recruiting its creators or users) to build AI-driven pharma companies, some boasting that they can go from target to clinical candidate in a fraction of the traditional time. Pharma giants, not known for technological agility, are partnering or pouring money into their own AI divisions, lest they miss the next penicillin because it was found by a machine.

And yet, hype and reality must be disentangled. We have a revolutionary tool, yes, but also a brewing contest over who controls it and how far it can go. The coming years will likely see AlphaFold-inspired AI integrated into every major pharmaceutical pipeline, accelerating preclinical research. We’ll also see rival methods (from Meta, from academic consortia, from who-knows-where) vying for the crown – perhaps predicting not just structure but protein behavior, drug interactions, even suggesting new biology that hasn’t been observed.

The race is as much collaborative as it is competitive; even as DeepMind keeps some cards close to the chest, the global scientific community has gained an indelible asset in the freely available predictions and code it did share. One might say the genie is out of the bottle, and it’s folding proteins at a breathtaking pace.

In drug discovery terms, the hit rate for finding new therapies should only improve from here. Drugs that were shelved because “we couldn’t get a crystal structure” may get revived. Targets once dismissed as too obscure might reveal an Achilles’ heel visible in an AlphaFold model. Researchers in fields as diverse as immunotherapy, neuroscience, and agriculture are already exploiting these predictions to innovate beyond what was previously possible.

The greater implication is that our fundamental understanding of biology – how life’s molecular machines are built – has leapt forward. For investors and biotech specialists, this AI-fueled structural revolution suggests that the coming decade could see faster discovery, smarter design, and perhaps a few paradigm-busting drugs born directly from silico insights. In drug discovery, as in biology, structure is money.

AlphaFold’s story is thus a rare one: a scientific breakthrough that hit overnight success, lived up to its immense promise, and is now changing how an industry works – all in a span of a few years. Not even the most optimistic observers in 2019 would have predicted that by 2025, a free online database of 3D protein models would become a staple tool for every pharma company and university lab, or that an AI would help map a colossal nuclear pore complex that puzzled scientists for decades (embl.org).

But here we are. To borrow a phrase from a certain tech mantra, this is not the end, it’s just the beginning. Biology has a way of surprising us, and with AI as a partner, more surprises surely lie ahead – perhaps the next one being that elusive molecule which folds disease into defeat.