DAY 634
The heating in the HPM Building ran on a schedule Yuki had never successfully decoded—hot until eleven, cold until three, unpredictably warm again in the late afternoon—and her standing desk near the radiator had the advantage of reliable warmth from November through March. She kept a second sweater on the back of her chair. The whiteboard wall behind her was covered in molecular diagrams that hadn't changed in two weeks, the problems those diagrams represented already solved and waiting on results, and in the meantime the lab's eleven workstations were running structural predictions for three active projects that had nothing to do with the dataset currently on her primary monitor.
The dataset had come from a man named Grigoriy Novikov, whom she'd never met, through a mutual contact at the Weizmann Institute who'd described him as "a mathematician, not actually a biologist, he knows this is an unusual request." The email had arrived thirteen days ago. The attachment: a compressed archive, raw numerical output in a format she didn't recognize, 40 gigabytes compressed. No context, no hypothesis, no methods section. Just the data and one sentence: Can you tell me what structure this describes?
She'd nearly deleted it. She had three grant reports due in December and a manuscript under review at Nature Computational Science and a standing desk already covered in papers she'd promised to read. But the data was clean—no noise artifacts, no signs of synthetic generation—and it was large, which meant someone had put serious resources into producing it. Her folding prediction pipeline had reacted to the preliminary sample she'd run before deciding whether to engage: anomalous structure signatures, it flagged, which meant the distribution pattern matched a known class of molecular shapes rather than random data.
Thirteen days. She ran analysis in the gaps—early mornings before the lab filled, one late Tuesday when a grant deadline cleared. She ran the preliminary sample again. Then a larger sample. Then she was redesigning her input normalization to handle whatever format this data was encoded in, and by that point she was going to run the full dataset.
She touched the scar on her left index finger: a thin line from a kitchen accident at age eight, white and slightly raised. Her hands found it when she was thinking hard. She noticed where they'd gone and kept working. At 09:12, she launched the full 40-gigabyte dataset through the pipeline and went to get coffee.
The pipeline was called STRUCT-3D, built over four years with two postdocs and a graduate student who'd since taken a faculty position in Singapore. It worked by decomposing input distributions into three-dimensional coordinate clusters and running those clusters against her database of known protein conformations. For the Novikov dataset, the input format was unusual—not amino acid sequence data, which was STRUCT-3D's normal intake, but something that looked like energy landscape mappings: millions of data points describing conformational state transitions. The adaptation she'd built to handle this format had taken four days and she wasn't fully confident in it.
She came back with the coffee. The pipeline was still running. She answered two emails, reviewed a PhD student's draft introduction, and opened the output interface at 11:04 when the completion notification arrived.
The first result was a three-dimensional coordinate set. STRUCT-3D rendered it as a ribbon diagram—protein backbone shown as a curved ribbon, alpha helices in red, beta sheets in yellow. She didn't recognize the fold immediately. She rotated the structure in the visualization window, looking for the shape's logic. Then the pipeline's annotation engine returned the identification.
Beta-amyloid peptide, Aβ42, conformational state: pre-aggregation fold. Confidence: 0.94.
She stopped rotating. She read the annotation again. Put down her coffee. Beta-amyloid. The amyloid precursor protein cleavage product whose aggregation into plaques was the defining pathology of Alzheimer's disease. A protein that researchers across six continents had spent four decades and tens of billions of dollars trying to understand at this structural level. The pre-aggregation fold was the target state—the configuration that had to be mapped before anyone could design a molecule to interrupt the cascade toward plaques. She ran beta-amyloid studies herself, different methodology and smaller scale, and she knew the published conformational literature with the specificity of someone who had built grant proposals around its gaps.
She sat with this for a moment. Not the discovery—the name. Beta-amyloid was not an abstraction in her field. It was the subject of 40,000 published papers and seventeen failed drug trials and a clinical graveyard full of compounds that had looked promising and then hadn't. It was the protein that had resisted her field's best tools for longer than some of her graduate students had been alive. The idea that a cryptocurrency mining operation had generated a pre-aggregation fold at 0.94 confidence was the kind of statement she would have annotated as methodologically implausible if she'd encountered it in a submitted manuscript. She scrolled to the second result.
TNF-alpha receptor 1, extracellular domain, active conformation. Confidence: 0.91.
TNF-alpha receptor. Tumor necrosis factor signaling, the pathway central to rheumatoid arthritis and ankylosing spondylitis and the autoimmune spectrum that pharma had turned into seventy billion dollars of annual biologics revenue. A different mechanism entirely from beta-amyloid. A different disease family. She touched her finger scar without deciding to.
There were forty-seven results in the queue. She scrolled through them carefully. The names came up one by one, and she knew most of them—the way you knew the names of problems your field had been working on for decades—and she recognized what they had in common before she was halfway through the list.
The Protein Data Bank contained just over 200,000 experimentally verified structures, accumulated by X-ray crystallography and cryo-electron microscopy over five decades of structural biology. It was the ground truth. Computational predictions were only as good as their agreement with it. She opened a split screen: the pipeline's beta-amyloid render on the left, PDB entry 1IYT on the right. She ran the RMSD alignment. Root mean square deviation: 0.23 angstroms.
She ran the number again. Not wrong in the error sense—wrong in the sense of being too good. AlphaFold's published benchmark on this structure class was 0.89 angstroms. The Novikov pipeline output was beating the best published prediction by a factor of four.
She cross-referenced the TNF-alpha result. RMSD 0.19. Pulled three more from the queue at random: 0.21, 0.17, 0.28. For forty-five minutes she worked through the list, running each output against the closest PDB entry she could find. Some of the queue's structures didn't have PDB entries—they represented predicted conformational states that hadn't been experimentally verified. But where verification was possible, the match was consistent. The Novikov dataset wasn't producing protein structure approximations.
She sat down. She worked standing almost always, the desk calibrated to her height, standing through the whole day a habit since her fellowship. She sat down now, pulled her secondary chair to the desk, and looked at the split screen—two structures overlaid, differing by less than the diameter of a hydrogen atom. Four decades of structural biology. Thousands of researchers. Experimental equipment worth hundreds of millions. The highest-accuracy prediction tool in the field, built by DeepMind with years of machine learning work. And underneath all of that, for 634 days, a cryptocurrency mining operation had been doing this. Faster. At higher resolution. To the angstrom. She pressed her thumb against the scar on her finger and held it there.
Riccardo was at the adjacent station, working through a revision on membrane protein folding kinetics—she told no one. Hana was running simulations at the far end. Jakub, the second-year PhD student, had his headphones on and wouldn't notice if the building shook. They were doing the work they were supposed to do. The lab had a rhythm today, and she didn't want to alter it, so she closed the split screen and walked out.
The hallway was long and tiled in the 1970s and cold in the way institutional hallways were cold regardless of outdoor temperature. A window at the far end looked north across campus, the rooftops of Höngg layered in the middle distance, November sky pale and uniform above them. She stopped at the window and stood there. Two students passed behind her. A professor from computational chemistry nodded. She nodded back.
The dataset from Grigoriy Novikov was the output of an Icelandic mining operation that had been running for 634 days. Forty-seven protein targets identified so far, covering disease pathologies from Alzheimer's research to the autoimmune and oncological literature she knew well. The predictions accurate at a resolution beyond any tool her field possessed. She needed to talk to the person who sent it. She had a mobile number in his original email. She used it, and there were three rings before he answered, the connection clear with a continuous low hum in the background—industrial machinery, far enough away to be texture rather than noise.
"Novikov."
"This is Yuki Tanaka. You sent me data thirteen days ago."
"Yes." A pause. "You've finished the analysis?"
"I have. I'm going to tell you what your data describes, and I need you to tell me what produces it."
"Go ahead."
"Your outputs are protein conformational state predictions—stable low-energy configurations for human proteins associated with disease pathologies. I've identified forty-seven protein targets. Beta-amyloid, tau, TNF-alpha receptor are furthest resolved. The prediction quality is better than anything currently published by a significant margin." Silence on the line. Not hesitation. Calculation.
"How far better?" he said.
"Factor of four on AlphaFold benchmarks. Point two angstrom RMSD range across verified proteins."
"Then it's been folding correctly," he said, after a pause. Not a question. A conclusion reached before the call.
"You explain what produces this output. Then I can tell you whether that's possible."
He explained it without ornamentation: the HashNet blockchain's proof-of-work algorithm, its pseudonymous designer. The Krafla facility in Iceland, 14,000 mining units at 2.4 exahashes per second. Hash outputs with distribution anomalies he'd flagged through statistical analysis—symmetry signatures inconsistent with random hash solving. He'd isolated the anomalous cluster and run his own pattern analysis. A mathematician's reading, not a biologist's.
"The proof-of-work computation," she said. "The internal loop that generates valid hashes."
"Yes."
"You're saying that loop maps to protein conformational states."
"The distributions match what you'd expect if it did."
She wrote nothing down. The proof-of-work algorithm found valid hashes by traversing an energy landscape. If that landscape was constructed to mirror protein folding dynamics, every valid hash would correspond to a low-energy protein configuration. The mining network's optimization pressure—thousands of competing pools racing toward faster solutions—would converge on correct folding solutions as a byproduct. The most powerful computational network on earth, powered by an economic incentive that guaranteed continuous operation, and no one understanding what they were computing.
"Who designed the algorithm?" she said.
"Listed as Takahashi. Pseudonymous. No other documentation."
"How many machines total on the network?"
"Rough estimate: one to two million."
One to two million machines. Running for 634 days.
"Every miner on the HashNet network," she said. "They've all been computing protein folding without knowing it."
"For 634 days," he said. "Yes."
She stayed at her desk after the call. The visualization she'd left open showed the beta-amyloid structure in its default rotation cycle—four seconds per revolution. The ribbon diagram turned: the pink alpha helix, the yellow beta strand, the grey loop of the pre-aggregation fold. It turned and came back and turned again.
Forty-seven million Alzheimer's patients worldwide. She'd cited that number in the introductions of her last four grant applications. It had always been a scale-setter, a justification for funding—the outer limit of relevance, kept at a useful distance.
She thought about David—not as longing, just as the natural adjacency of disease to the person who treated it. He worked at Boston Children's, pediatric oncology, which meant his patients were the number made particular. Mika was seven and lived with him during the school year. Mika had her mother's hair and her father's way of approaching a question from every angle before accepting the answer, and the structures on her screen represented diseases that could enter that world before Mika had the chance to understand them.
She was making the leap too early, and she knew it. Structural predictions were not synthesized molecules. Synthesized molecules were not tested interventions. Tested interventions were not approved treatments. The standard distance between this screen and clinical practice was measured in years, in trials, in regulatory frameworks that existed for reasons that didn't vanish because the prediction was good.
But the predictions were accurate to 0.23 angstroms. Forty-seven of them. Accumulated over 634 days inside a computational network the size of a national power grid, running on economic incentives that would keep it running whether anyone wanted it to or not.
She was not a person who used the word "cure" without precision. It was not a scientific term. But sitting alone in her office at ETH Zurich at 4:42 in the afternoon, looking at the data from a mining operation in Iceland, she let herself think it plainly.
Grigoriy Novikov, a mathematician in Iceland, had sent her data he didn't understand because he recognized a shape that shouldn't exist. Now she understood what it was. Two people in the world understood what it was.
The structure rotated. Beta-amyloid, pre-aggregation fold, 0.23 angstroms from verified reality. She watched it turn.