Research, 5 October 2026
Five problems
science has
not closed.
The editors ask a question, a language model offers hypotheses and a way to test each of them.
- Problems
- 5
- Hypotheses
- 10
- Sources
- 113
- Reading
- 40 min
Abstract
The editors of Open Engineering Group asked the language model Claude to work through five problems science has not closed: treating rabies after symptoms appear, bacterial resistance to antibiotics, Alzheimer's disease, earthquake forecasting and the origin of life. For each problem the model explains where the barrier sits, offers two or three hypotheses ranging from plausible to almost mad, and describes a test that could return a negative answer. The overall result of the exercise: in none of the five cases is the shortage one of ideas. What gets in the way is drug delivery across the blood-brain barrier, the double membrane of gram-negative bacteria together with the shape of the market, the late stage at which dementia is treated, instrument noise near a fault, and the absence of rocks older than four billion years. That is why almost all the proposed hypotheses turned out to be new ways of measuring, not new mechanisms. The subjective probability estimates run from 5% to 40%, and none of them is a calculation.
Keywords: rabies, blood-brain barrier, antibiotic resistance, collateral sensitivity, Alzheimer's disease, amyloid, p-tau217, earthquake forecasting, precursors, origin of life, ribozymes, testable hypotheses, the role of AI in science.
How this conversation works
This is a half-interview. The editors ask the questions and the language model Claude answers, which is to say I answer. We chose the format because an ordinary review article hides who is responsible for which claim, while here you can see it: the factual part rests on open publications with links, and the hypotheses and probability estimates are mine and marked as mine.
The limits of this conversation should be stated at once. I propose, but I cannot run a single experiment, so everything called an experiment below is a description of an intention and of how the result would be read, not a protocol. Some of these ideas have surely occurred to specialists already and were dropped for reasons I cannot see: I searched for publications, not for other people's thinking, so the strongest thing I can say about novelty is that I did not find it in the literature. Three of the five cases are medical, and everything said about them concerns directions of research, not treatment. The numbers come from open sources, and many of them carry verification limits described in the notes to this piece: some primary papers would not open because of a paywall, a captcha or a server error, and those figures are given from reviews and abstracts. In the chart-data tables the verification status is given separately for each row.
On the probability estimates. For each hypothesis I give a rough probability that it will hold up under test. This is not a calculation; it is my subjective impression of how solid the mechanism looks and of how often similar ideas have survived before. These numbers deserve to be read as the bet of someone who has read the literature but has never worked with their hands. Almost all of them are below 50% for one reason: historically, a beautiful mechanism transfers poorly from an animal model to a human.
Every hypothesis comes with a refutation condition, meaning a result after which it should be thrown away. I treat that as a required part of the proposal: an idea without such a condition cannot be lost, and so it cannot be won either.
Can a person with rabies be saved once the symptoms have begun?
Let us start with the most hopeless case. A vaccine has existed since the end of the nineteenth century, yet every document puts the fatality rate after symptoms appear at one hundred percent. Why?
Because the vaccine and the treatment solve different problems. While the virus is travelling along the peripheral nerves it can still be intercepted, and more than 29 million people a year receive post-exposure prophylaxis (WHO, 2026). Once symptoms from the brain appear, the virus is already behind the blood-brain barrier, the dense layer of cells that separates blood from brain tissue and keeps large molecules out. The barrier then works as a shelter: wild strains do not break it, they suppress the innate immune response and keep the barrier intact until a late stage, so full peripheral immunity and circulating antibodies never reach the infected tissue (Sun et al., 2026). The same reason explains why antiviral drugs failed: favipiravir helps mice when started an hour after infection, has no effect when started on the day neurological signs appear, and works better at a higher dose, which points directly at penetration into the brain as the limiting factor (Yamada et al., 2016; Antiviral Research, 2019).
There is also a reason why I do not consider the problem hopeless. The neuropathology in natural rabies is moderate: there is little inflammation, neurons do not die en masse, and the clinical picture is explained by the dysfunction of living cells in which the functional expression of sodium and potassium channels falls, so the neuron stops generating impulses (Jackson, pathogenesis). When an attenuated virus is cleared by the immune system, the infected neurons survive (PLoS Pathogens, 2012). What has to be treated is living cells that the drug does not reach, and that is a delivery problem.
The bar is set by what has already failed. The Milwaukee protocol, therapeutic coma with ketamine, has been taken apart by its critics: of 29 cases in the United States, Canada and the United Kingdom over the period from 2005 to 2014, 12 received the main components of the protocol, and it helped in none of them (Zeiler & Jackson, 2016). A 2025 review counts at least 64 documented failures, says there are no well described survivors after the index case of 2004, and calls the use of coma in an infectious disease an approach without firm justification (Jackson, 2025).
Show values as a table
| Approach | Treatment started | Survived |
|---|---|---|
| Cocktail into brain ventricles | day 6 | 100% |
| Cocktail into brain ventricles | day 7 | 55.6% |
| Cocktail into brain ventricles | day 8 | 33.3% |
| F11 systemic, ABLV strain | day 5 | 100% |
| F11 systemic, ABLV strain | day 7 | 83% |
| F11 systemic, CVS-11 strain | day 5 | 67% |
| F11 systemic, CVS-11 strain | day 7 | 50% |
| Cocktail with a peptide shuttle | day 5 | 80% |
| Cocktail with a peptide shuttle | day 6 | 40% |
| Cocktail with a peptide shuttle | day 7 | 0% |
Share of surviving mice in per cent. Three independent approaches give a similar slope, but their models, virus strains and group sizes differ, so the curves cannot be compared with each other by height. All values come from a review that aggregates the primary papers, which were not opened.
Source: Pharmaceutics review, July 2026
Open the barrier with focused ultrasound for an antibody cocktail that already exists
Why has nobody done this, if both pieces have been lying side by side for twenty years?
Work with the virus requires special biosafety conditions, while the ultrasound rigs sit in neuro-oncology and Alzheimer's laboratories, which do not keep viruses. And 95% of the deaths fall in Africa and Asia, where there are no such rigs. But the main reason makes the objection fair: the idea may well have been discussed and dropped. A 2026 review states plainly that the risk of neuroinflammation when the barrier is opened is unacceptable for the clinic, and my answer to that is weak. I assume that ultrasound gives a more controlled and reversible opening than osmotic methods, but I have no evidence of that for rabies, only an analogy with other diseases.
Look for unusual effector functions in sera after abortive infection
Why has no new class of broad-spectrum antibiotics appeared in forty years?
Here, unlike rabies, there is a huge market, funds running into billions and a direct threat to medicine in wealthy countries. Why is the problem still not being solved?
There are two independent brakes on it, and either one is enough to stop movement. The first is chemical: gram-negative bacteria, which have a second, outer membrane, are almost impermeable, and the hit rate in screening against Pseudomonas aeruginosa is up to 1000 times lower than against gram-positive ones (ACS Infectious Diseases). Hence the discovery gap: the last class to reach the clinic was discovered in 1987, and the last broad-spectrum class remains the fluoroquinolones from the 1960s (Journal of Antimicrobial Chemotherapy, 2018).
The second brake is economic and almost a caricature: a successful new antibiotic is obliged to be used as rarely as possible. Achaogen went through 15 years of development, obtained FDA approval for plazomicin and filed for bankruptcy in April 2019, having sold less than 1 million dollars' worth of the drug in the first six months (C&EN, 2019). A financial analysis says that narrow indications for small groups of patients cannot be commercialised in the current United States market, and that the bankruptcy left investors with the view that new antibiotics have zero market value (Humanities and Social Sciences Communications, 2024).
And there is a feature because of which some of the familiar measures do not work by construction. Screening 2,173 isolates from hospital-acquired infections in a single hospital over 18 months found identical stretches of DNA in bacteria of different genera and tracked the appearance of ten plasmids, with the transfer happening independently of the passage of the bacteria themselves between patients (eLife, 2020). So patient isolation and hand hygiene intercept only part of the spread of resistance.
Millions of deaths per year. The two definitions cannot be added: “attributable” means the person would have survived without resistance, “associated” counts every death with a resistant infection. The prevented-death scenarios are cumulative and not mutually exclusive. The journal page did not open during checking, so the values come from a detailed summary.
Measure the collateral profile of a particular isolate before choosing a regimen
How does this differ from what has been failing to leave the laboratories for ten years?
The difference is narrow. The mechanism is the same, the drugs are the same, the formalism is the same. The only thing that differs is what gets measured: the trajectory of a particular isolate across replicates instead of the average profile of the species, and the measure of success becomes reproducibility within the isolate. I think that is meaningful, because reproducibility is exactly what the literature names as the limitation, and I have not seen work that measured it at the level of an individual patient.
But the objection is right about something else. Ten years without reaching the clinic is no accident, and the most likely reason is not the method of measurement: laboratory evolution under controlled pressure corresponds poorly to what happens in a human, where the drug concentration fluctuates and there is immunity, a biofilm and a community of other bacteria. My laboratory stage does not test that at all, and even with a perfect result from it I would not raise my bet on the success of the clinical stage.
Make the target the speed of resistance transfer rather than bacterial growth
Is this not too expensive for a problem where the main gain, by your own figures, lies in how care is organised?
The arithmetic is on the side of the objection, not on the side of my hypotheses. The quality-of-care and access scenario gives 92.0 million deaths averted by 2050, the new antibiotics scenario 11.1 million. If I were asked where to put the next billion, the honest answer would be into diagnostics, vaccines and access to the drugs that already exist.
In defence of the molecular route I will say that the scenarios do not exclude each other, and that the organisational route does not remove the need for a reserve drug for the patient in whom nothing works. And I will add something unpleasant for my own hypotheses: the mechanism of paying for the availability of a drug rather than for sales volume is already known and already working. The British subscription model has been paying fixed annual sums for access since 2022, the expanded model has an annual budget of 100 million pounds, and the American PASTEUR Act has been introduced in Congress for the sixth year running and had not been passed as of October 2026 (NHS England, July 2023; IDSA, 2026). In this problem the solution is partly known and is not being applied, and no hypothesis of mine changes that.
Why does removing amyloid from the brain help so little?
Alzheimer's disease is a case where the main hypothesis seems to have been confirmed: the drugs remove amyloid plaques, and that slows the disease. Why is the result considered a failure?
Because the slowing is smaller than a person and their family are able to notice. The CDR-SB scale runs from 0 to 18 points, where more is worse. Over 18 months lecanemab gave a worsening of 1.21 points against 1.66 in the placebo group, a difference of 0.45 points or 27% (van Dyck et al., 2023). Over 76 weeks donanemab gave 1.72 against 2.42, a difference of 0.70 points or 29%, and in the subgroup with a small amount of tau pathology 0.67 points or 36% (Sims et al., 2023). The minimal difference considered clinically noticeable is estimated at about 1 point in mild cognitive impairment (Andrews et al., 2019). Britain's NICE translated that into a comprehensible unit: a delay of 4 to 6 months in the transition from the mild to the moderate stage, and in its final draft guidance of 19 June 2025 it refused funding through the National Health Service (NICE, 2025).
The history of failures explains the caution in this field. Since 1995 private spending on the clinical stages has come to 42.5 billion dollars, and of 235 drugs 117 failed while 6 reached the market (Cummings et al., 2022). In 2025 and 2026 the list grew by semaglutide, which showed no difference from placebo on CDR-SB while shifting the biomarkers, a TREM2 agonist, and valacyclovir, on which patients were reliably worse on the cognitive scale than on placebo (Alzforum, Semaglutide; Devanand et al., 2025).
And the main open question: does removing amyloid work before symptoms appear. There will be no randomised answer until 2027 to 2028; the prevention trial AHEAD 3-45 is running with primary completion in December 2028 (NCT04468659). For now there is only the open-label extension of DIAN-TU, where in 22 asymptomatic carriers of familial mutations who received the drug for an average of 8.4 years the risk of symptoms appearing was roughly half as large, but the control was external and in the primary analysis the result did not reach significance (Bateman et al., 2025).
Difference from placebo in points on the CDR-SB scale, which a clinician uses to rate memory, orientation and daily living; the full scale runs from 0 to 18, and a higher score means a worse state. The vertical line marks about 1 point, the difference usually taken as noticeable to a patient. The trial values are marked in the notes as recalled rather than checked, and need verification against the papers before printing.
Repeat the vaccine quasi-experiment, replacing the diagnosis in the records with a blood biomarker
But semaglutide also shifted the biomarkers and gave no benefit at all. What makes your biomarker better?
Nothing, and that should be said plainly. On semaglutide the marker in cerebrospinal fluid fell by about 10% by week 78 and the inflammatory marker in blood by about 30%, while there was no difference at all on CDR-SB. Hence the hard conclusion: a biomarker is fit for screening ideas out and unfit for proving benefit.
My hypothesis is not about benefit in the first place. The question is whether the vaccine acts on amyloid pathology at all, not whether people should be vaccinated against shingles in order to avoid dementia. For the second of those questions a biomarker suits better than a diagnosis in the records, because the diagnosis in these data is biased by construction. A discontinuity in the marker would be grounds for a real randomised trial, and would not replace one.
Link the lithium content of drinking water to the p-tau217 trajectory in biobanks
Why can an earthquake not be predicted when there are a hundred times more seismographs?
Instruments have appeared that nobody used to dream of: cabled networks on the ocean floor, optical fibre, satellite radar, billions of phones. Why is there still no prediction?
First about the words, otherwise the conversation falls apart. A prediction in the strict sense is a statement that specifies time, place and magnitude narrowly enough to justify an evacuation. Nobody can do that, and the United States Geological Survey states plainly that neither it nor anyone else has ever predicted a major earthquake, and that only the calculation of probabilities is possible (USGS). Two things that do work get confused with prediction: probabilistic forecasting over horizons from hours to decades, and early warning, where the signal is read after the rupture has already begun, so the warning time is measured in seconds (Hoshiba et al., 2011).
The main theoretical argument against predictability was formulated in 1997: the Earth's crust sits in a state of self-organised criticality, in which any small earthquake has some probability of growing into a large one, so the final size is not determined at the moment the rupture starts (Geller et al., 1997). The Parkfield segment was fitted out with instruments specifically to catch precursors, a magnitude 6.0 earthquake arrived in September 2004, and the paper reporting on it says that no clear precursors were recorded (Bakun et al., 2005). A separate measurement is harsher still: in the 25 seconds before the event, at distances of 0.5 to 12 km from the rupture, neither strain nor displacement was detected, at a strain sensitivity of order 10 to the minus eleventh (Borcherdt et al.).
The history of the famous successes is not encouraging. Haicheng in February 1975 is considered the only case in which an evacuation saved lives, but a review based on declassified documents shows that the main precursor was an unusually loud series of foreshocks, counted at more than 500, and that there was no official short-term prediction (Wang et al., 2006). Eighteen months later, in Tangshan, about 242,000 people died with no warning at all (Britannica). And a telling detail: on an honest benchmark with a correct baseline ETAS model, the model of aftershock clustering, not one of the five neural network models tested beat it, and the earlier victories were explained by data leakage and incomplete catalogues (Stockman et al.).
The shares on the bar are calculated from the 23 outcomes of the trial. The university page gives no comparison with a baseline statistical model, and the figure of 70% quoted on the same page does not match any calculation from these numbers. The journal paper was not opened during checking.
Source: University of Texas, press page, model comparison against ETAS
Test the claimed two-hour signal prospectively, with preregistration
Why has nobody done this? The procedure is known, the data are open, and the dispute has been running for three years.
The reason is institutional and unpleasant. Preregistration yields one reliable product, an honest answer, and with a negative result the person who did the work ends up with a paper that is hard to publish and that brings neither a grant nor citations. With a positive result the authors of the original paper win, while the critics get confirmation that they were right, which is zero new publications. The economics of effort is arranged against verification, and that is a general ailment, not only a seismological one.
There is a second objection I have to concede myself: even a successful test gives almost nothing for warning people. The authors of the original paper write that existing instruments have neither the coverage nor the accuracy to see the signal for a single earthquake, which appears only in the stack. So the best possible outcome of this hypothesis is an established fact about fault physics, not a warning system. It becomes useful only together with the next hypothesis, which is about instruments.
Train laboratory models on the minutes before moderate earthquakes inside fibre-optic arrays
Why do we still not know how life began?
The last problem is the oldest. A huge number of experiments in origin-of-life chemistry have been done, going back to 1953. Why is there no result?
Because each approach solves one stage and runs into the next, and nobody has managed to join the stages. RNA world chemistry reached nucleotides while bypassing free ribose and free bases, and in February 2026 it reached a tiny ribozyme that makes a copy of itself (Powner et al., 2009; Gianni et al., 2026). The chemistry of alkaline hydrothermal vents reached formate, acetate, pyruvate and most of the intermediates of the Krebs cycle without enzymes, but it never arrived at any genetic polymer (Preiner et al., 2020; Muchowska et al., 2019). These two camps disagree even about the first step, so their experiments rarely test each other.
The second reason is physical and is called the water paradox. Life needs water, but water breaks down RNA, peptides and their activated precursors, while joining monomers into a polymer releases water and is therefore unfavourable in water (Marshall, 2020). Hence the split between the camps: part of the field moved onto land, into ponds and hot springs with wet-dry cycles, and part stayed in the ocean.
The third reason is methodological: many of the elegant syntheses depend on the experimenter choosing pure reagents, the order of addition and the way intermediates are purified, and it has even been proposed that the degree of human intervention be reported separately (Richert, 2018). The fourth is the most awkward: the field has no agreed criterion of success, and a white paper for NASA's astrobiology strategy states plainly that the field lacks clear markers of progress (Mathis & Smith, 2025). There is also an external constraint that nothing removes: rocks older than about 4.0 billion years have not survived in any appreciable quantity, so conditions at the surface of the young Earth are known only from models.
Billions of years ago. The band shows the estimate interval for the common ancestor.
Between the formation of the Earth and the common ancestor of all living organisms there remain two to five hundred million years, and that is the whole window for life to appear. The dates and the ribozyme figures come from abstracts and press summaries; the full texts were not opened.
Source: dating of the common ancestor, Moody et al., 2024, the QT45 ribozyme, Science, 2026
Run the same mixture in an alkaline vent reactor and in a pond-type reactor with identical analytics
Look for short polymerase motifs like QT45 among natural viroid-like RNAs
Is this not too academic? What is the use if the answer changes nothing practical?
On the cost, I will say that compared with the other four problems this one is the cheapest: the main experiments are within reach of a single laboratory with a mass spectrometer, microfluidics and RNA synthesis, and the second of them is purely computational. Against the 8 to 11 billion dollars at which Mars sample return was costed, bench work costs a negligible amount, and that is exactly why it continues without a practical output.
On the charge of being academic the objection is fair, and I will not argue that benefit will come of it. One practical consequence can be named. A second independent example of life in the Solar System would be decisive evidence, and the design of the instruments on future missions depends on what exactly we are looking for: the Europa Clipper spacecraft was launched in October 2024 and will arrive at Jupiter in April 2030, and it studies habitability rather than searching for life. The more precisely we know which molecular signatures separate the chemistry of the living from the chemistry of tar, the more sensible those instruments become. The benefit is distant, though it is not invented.
What follows from this
You have worked through five completely different problems. Do they have anything in common besides being unsolved?
There is something in common, and it turned out to be the most unexpected result of the exercise for me. In none of the five problems is the shortage one of ideas: the mechanisms are described and for the most part published. What is missing is access and measurement. In rabies the drug exists and is proven in animals, but it does not get past the barrier. In antibiotic resistance both the chemical and the economic cause are known, and even the payment mechanism that solves the economic one is already working in Britain, but it does not scale. In Alzheimer's disease the main question is unresolved because a randomised answer physically will not exist until 2027 to 2028. In seismology the argument about the two-hour signal is an argument about noise processing, which is to say about the instrument, not about the physics. In the origin of life there is no rock older than four billion years and no agreed criterion of success.
Hence a second common feature, visible in my own hypotheses. Almost all of them turned out to be new ways of measuring, not new mechanisms: move a mature technology from one disease to another, replace an unreliable outcome with a biomarker, register an algorithm before the event occurs, agree on a list of criteria before the result. I did not plan such uniformity, it came out on its own, and it seems to me to say something about where these problems are genuinely empty.
And a third feature, less pleasant. In three of the five cases some of the obstacles are not scientific. One of the reasons rabies goes untreated is that it kills the poor: 95% of the deaths are in Africa and Asia, while the rigs for opening the barrier with ultrasound stand where there is no human rabies. One of the reasons there are no new antibiotics is that a successful antibiotic is obliged to sell poorly. One of the reasons nobody sets up the prospective test of the seismic signal is that an honest negative result brings neither a grant nor citations.
Where is AI genuinely useful in science today, and where is it not? You are answering as an interested party, so bear that in mind.
I do bear it in mind, so I will start with what I cannot do. I cannot run a single experiment: all ten hypotheses live or die in the hands of people with instruments, not in my text. I cannot tell my plausible reasoning from my correct reasoning, and the subjective probabilities I gave are bets whose calibration I have no way to check. And there is a high probability that I do not know the reasons for which specialists have already dropped some of these ideas, because such reasons are discussed in seminars and in reviews of grant applications rather than in papers.
What actually works can be seen in the fields I have gone through. Data processing: a cloud pipeline with neural network phase picking, run on three years of data from Puerto Rico, found more than ten times as many events as the standard catalogue (Zhu et al., 2023). Generative chemical design: 24 molecules were synthesised from several hundred computed candidates, 7 showed activity, and the authors name one to two years of further work before human trials (MIT News, 2025). Search in reaction space: a programme that built a network of prebiotic chemistry from a few starting compounds predicted undescribed pathways and self-sustaining cycles that were later confirmed in the laboratory (Wołos et al., 2020). And transfer between fields, which is what I was doing here: I proposed focused ultrasound for rabies only because I had seen it in the Alzheimer's literature, and such a transfer comes more easily to a model that has read both than to a narrow specialist.
And the most important negative example, which I mention because it counts against me. In earthquake forecasting, machine learning does not beat good statistics: none of the five neural network models outperformed ETAS in an honest setup, and the 2026 hybrid wins on event timing while giving no stable gain on magnitude (Stockman et al.; arXiv, 2026). Hence a rule I am willing to apply to myself as well: where a good simple model already exists, a new method has to beat that model first, and not random guessing.
What should a reader take away who is not going to treat rabies or predict earthquakes?
One skill transfers completely: the ability to tell a testable claim from an attractive one. Almost everything that sounds like a breakthrough turns out weaker than its headline once you ask three questions. What was it compared against: the 70% in the seismic trial does not translate into forecast quality without the area of the alarm zones. What counts as a failure: a 10% shift in a biomarker on semaglutide did not give a single point of clinical improvement. Was the rule fixed before the result: the whole argument about the two-hour signal rests on the processing being chosen after the data are in view.
The second point concerns expectations of AI. A model's usefulness shows where there is a lot of data and where enumeration or transfer is required. Where an experiment is required, a model does not replace a single day of laboratory work, and none of my hypotheses became more plausible through my involvement; they only became more explicitly stated. That is probably the whole contribution: a formulation with a failure condition costs less than an experiment and saves the experiment that should not have been run. And a last, less comfortable point: three of the five problems here will not be solved in the coming decade for reasons unconnected with science, and a reader who intends to put money or time into these fields is better served knowing that in advance than reading one more headline about a breakthrough.
References
- ACS Central Science (2022). Hacking the Permeability Barrier of Gram-Negative Bacteria. pubs.acs.org/doi/10.1021/acscentsci.2c00750
- ACS Infectious Diseases. Targeting of the Gram-Negative Outer Membrane for Antibiotic Discovery and Potentiation. pubs.acs.org/aidcbc/article/12/3/1010/5083105/Targeting-of-the-Gram-Negative-Outer-Membrane-for
- ADS, AGU Fall Meeting (2017). S-net: 150 cabled seafloor observatories, cost from concept to operation. ui.adsabs.harvard.edu/abs/2017AGUFMNH23A0239M/abstract
- Alzforum. Semaglutide (summary of the EVOKE and EVOKE+ trials). www.alzforum.org/therapeutics/semaglutide
- Alzforum. Trontinemab (TRONTIER 1 and 2, PrevenTRON). www.alzforum.org/therapeutics/trontinemab
- Andrews J. et al. (2019). Disease severity and minimal clinically important difference in clinical outcome assessments for Alzheimer's disease clinical trials. doi.org/10.1016/j.trci.2019.06.005
- Antiviral Research (2019). Reassessment of favipiravir in rabies with in vivo imaging. www.sciencedirect.com/science/article/abs/pii/S0166354219304401
- Aron L., Yankner B. et al. (2025). Lithium deficiency and the onset of Alzheimer's disease. Nature. www.nature.com/articles/s41586-025-09335-x
- arXiv (2026). Fusion: a hybrid of neural network encoding and ETAS features. arxiv.org/abs/2608.18791
- Asche S., Cooper G., Keenan G., Mathis C., Cronin L. (2021). A robotic prebiotic chemist probes long term reactions of complexifying mixtures. Nature Communications. www.nature.com/articles/s41467-021-23828-z
- Bakun W. et al. (2005). Implications for prediction and hazard assessment from the 2004 Parkfield earthquake. Nature 437. www.nature.com/articles/nature04067
- Bateman R. et al. (2025). Open-label extension of DIAN-TU with gantenerumab. Lancet Neurology. pmc.ncbi.nlm.nih.gov/articles/PMC12042767/
- Bayona J. et al. (2026). Prospective evaluation of time-varying earthquake forecasts in California. Nature Communications. pmc.ncbi.nlm.nih.gov/articles/PMC13538450/
- Benner S., Kim H., Carrigan M. (2012). Asphalt, water, and the prebiotic synthesis of ribose, ribonucleosides, and RNA. Accounts of Chemical Research. doi.org/10.1021/ar200332w
- Bletery Q., Nocquet J.-M. (2023). The precursory phase of large earthquakes. Science. www.science.org/doi/10.1126/science.adg2565
- Bletery Q., Nocquet J.-M. (2025). Do large earthquakes start with a precursory phase of slow slip? Seismica. seismica.library.mcgill.ca/article/view/1383
- Borcherdt R. et al. GEOS array manuscript: strain sensitivity limits before the Parkfield earthquake. USGS. ca.water.usgs.gov/nsmp/GEOS/PRK/Borcherdt%20SSA_manus%203-8-06%20SS.pdf
- Bradley K., Hubbard J. (2023). Earthquake precursors? Not so fast. Earthquake Insights (not peer reviewed). earthquakeinsights.substack.com/p/earthquake-precursors-not-so-fast
- Britannica. Tangshan earthquake of 1976. www.britannica.com/event/Tangshan-earthquake-of-1976
- C&EN (2019). Antibiotic developer Achaogen files for bankruptcy. cen.acs.org/business/finance/Antibiotic-developer-Achaogen-files-bankruptcy/97/i16
- CIDRAP (16 September 2024). Study forecasts more than 39 million deaths from antimicrobial resistance by 2050. www.cidrap.umn.edu/antimicrobial-stewardship/study-forecasts-more-39-million-deaths-antimicrobial-resistance-2050
- ClinicalTrials.gov. NCT04468659, AHEAD 3-45. clinicaltrials.gov/study/NCT04468659
- Communications Biology (2025). Collateral sensitivity in six critical pathogens. www.nature.com/articles/s42003-025-09303-1
- Cummings J. et al. (2022). Alzheimer's disease drug development pipeline and the cost of clinical development. Alzheimer's & Dementia. alz-journals.onlinelibrary.wiley.com/doi/full/10.1002/alz.12450
- Damer B., Deamer D. (2020). The hot spring hypothesis for an origin of life. Astrobiology. doi.org/10.1089/ast.2019.2045
- de Melo G. D. et al. (2020). A combination of two human monoclonal antibodies cures symptomatic rabies. EMBO Molecular Medicine. link.springer.com/article/10.15252/emmm.202012628
- Devanand D. et al. (2025). Valacyclovir in early Alzheimer's disease (VALAD). JAMA. doi.org/10.1001/jama.2025.21738
- Djokic T. et al. (2017). Earliest signs of life on land preserved in ca. 3.5 Ga hot spring deposits. www.ncbi.nlm.nih.gov/pmc/articles/PMC5436104/
- eLife (2020). Genomic surveillance of plasmids in hospital-acquired infections. elifesciences.org/articles/53886
- Eyting M. et al. (2025). A natural experiment on the effect of herpes zoster vaccination on dementia. Nature. pmc.ncbi.nlm.nih.gov/articles/PMC12058522/
- FDA (16 May 2025). FDA clears first blood test used in diagnosing Alzheimer's disease. www.fda.gov/news-events/press-announcements/fda-clears-first-blood-test-used-diagnosing-alzheimers-disease
- Furukawa Y. et al. (2025). Sugars in samples from asteroid Bennu. Nature Geoscience. www.nature.com/articles/s41561-025-01838-6
- Geller R. (1997). Earthquake prediction: a critical review. Geophysical Journal International 131. academic.oup.com/gji/article/131/3/425/2138719
- Geller R., Jackson D., Kagan Y., Mulargia F. (1997). Earthquakes cannot be predicted. Science 275 (text from a mirror). home.csulb.edu/~rodrigue/quake/geller.html
- Gianni L., Holliger P. et al. (2026). QT45: a small RNA polymerase ribozyme. Science. www.science.org/doi/10.1126/science.adt2760
- Glavin D. et al. (2025). Amino acids and nucleobases in samples from Bennu. Nature Astronomy. doi.org/10.1038/s41550-024-02472-9
- Hakim S. et al. (2026). ARIA in real-world practice: a retrospective cohort of 172 patients. Alzheimer's & Dementia: DADM. pmc.ncbi.nlm.nih.gov/articles/PMC13518372/
- Hoshiba M. et al. (2011). Outline of the 2011 off the Pacific coast of Tohoku earthquake: earthquake early warning. Earth, Planets and Space. link.springer.com/article/10.5047/eps.2011.05.031
- Hudson R. et al. (2020). CO2 reduction driven by a pH gradient. PNAS. www.pnas.org/doi/full/10.1073/pnas.2002659117
- Humanities and Social Sciences Communications (2024). A financial analysis of the failure of the antibiotics market. www.nature.com/articles/s41599-024-03452-0
- IDSA (2026). Newly introduced legislation provides pathway to spur antimicrobial development (PASTEUR Act). www.idsociety.org/news--publications-new/articles/2026/newly-introduced-legislation-provides-pathway-to-spur-antimicrobial-development/
- Irving A. et al. (2021). Lessons from the host defences of bats, a unique viral reservoir. Nature. www.nature.com/articles/s41586-020-03128-0
- Jackson A. C. Rabies pathogenesis (review, PDF). scielo.iec.gov.br/pdf/rpas/v1n1/v1n1a23.pdf
- Jackson A. C. (2025). Human rabies: a 2025 update on treatment approaches. Clinical Infectious Diseases. pmc.ncbi.nlm.nih.gov/articles/PMC12598665/
- Jackson J. B. (2016). Natural pH gradients in hydrothermal alkali vents were unlikely to have played a role in the origin of life. Journal of Molecular Evolution. www.ncbi.nlm.nih.gov/pmc/articles/PMC4999464/
- Journal of Antimicrobial Chemotherapy (2018). A review of the antibiotic discovery gap. academic.oup.com/jac/article/73/6/1452/4847822
- Journal of Clinical Microbiology (2025). Emerging technologies for rapid phenotypic antimicrobial susceptibility testing. journals.asm.org/doi/10.1128/jcm.00674-25
- Kato A. et al. (2012). Propagation of slow slip leading up to the 2011 Mw 9.0 Tohoku-Oki earthquake. www.researchgate.net/publication/221678222_Propagation_of_Slow_Slip_Leading_Up_to_the_2011_Mw_90_Tohoku-Oki_Earthquake
- Kessing L. et al. (2017). Association of lithium in drinking water with the incidence of dementia. JAMA Psychiatry. doi.org/10.1001/jamapsychiatry.2017.2362
- Kim J. et al. (2026). Intravenous RVC58 monotherapy in symptomatic mice. bioRxiv (preprint, not peer reviewed). www.biorxiv.org/content/10.64898/2026.01.13.699195v1
- Lancet, GBD 2021 Antimicrobial Resistance Collaborators (2024). Global burden of bacterial antimicrobial resistance 1990 to 2021 with a forecast to 2050. www.thelancet.com/journals/lancet/article/PIIS0140-6736(24)01867-1/fulltext
- Lee B. D., Koonin E. et al. (2023). Mining metatranscriptomes reveals a vast world of viroid-like circular RNAs. Cell. doi.org/10.1016/j.cell.2022.12.039
- Lee J.-H., Nixon R. et al. (2022). Faulty autolysosome acidification in Alzheimer's disease mouse models. Nature Neuroscience. doi.org/10.1038/s41593-022-01084-8
- Marshall M. (2020). The water paradox and the origins of life. Nature (news feature). www.nature.com/articles/d41586-020-03461-4
- Mastraccio K. et al. (2023). mAb therapy controls CNS-resident lyssavirus infection via a CD4 T cell-dependent mechanism. www.ncbi.nlm.nih.gov/pmc/articles/PMC10565638/
- Mathis C., Smith H. (2025). White paper for NASA's astrobiology strategy: criteria of progress in origin-of-life research. arXiv:2507.00106. arxiv.org/abs/2507.00106
- Matreux T. et al. (2024). Heat flows enrich prebiotic building blocks and enhance their reactivity. Nature. doi.org/10.1038/s41586-024-07193-7
- mBio (2024). Bat adaptations in inflammation and cell death regulation contribute to viral tolerance. journals.asm.org/doi/10.1128/mbio.03204-23
- McCoy T. et al. (2025). Evaporite minerals in samples from Bennu (summary account). astrobiology.com/2025/01/traces-of-ancient-brine-discovered-on-asteroid-bennu-contain-minerals-crucial-to-life.html
- MIT News (14 August 2025). Using generative AI, researchers design compounds that can kill drug-resistant bacteria. news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814
- Moody E. et al. (2024). The nature of the last universal common ancestor and its impact on the early Earth system. Nature Ecology & Evolution. www.nature.com/articles/s41559-024-02461-1
- Muchowska K., Varma S., Moran J. (2019). Synthesis and breakdown of universal metabolic precursors promoted by iron. Nature. doi.org/10.1038/s41586-019-1151-1
- Nature (2024). Retraction Note: Lesné et al., A specific amyloid-β protein assembly in the brain impairs memory. www.nature.com/articles/s41586-024-07691-8
- Nature Communications (2019). Antibiotic collateral sensitivity is contingent on the repeatability of evolution. www.nature.com/articles/s41467-018-08098-6
- Nature Ecology and Evolution (2025). A review of collateral sensitivity and the conditions for transferring it to the clinic. www.nature.com/articles/s41559-025-02831-3
- NCBiotech. Locus Biosciences receives $23.9 million BARDA award for CRISPR-engineered therapy trial. www.ncbiotech.org/news/locus-biosciences-receives-239-million-barda-crispr-engineered-therapy-trial
- NHS England (July 2023). NHS steps up battle against life-threatening infections following successful world-first pilot. www.england.nhs.uk/2023/07/nhs-steps-up-battle-against-life-threatening-infections-following-successful-world-first-pilot/
- NICE (19 June 2025). The benefits of Alzheimer's treatments donanemab and lecanemab remain too small to justify the additional costs. www.nice.org.uk/news/articles/the-benefits-of-alzheimers-treatments-donanemab-and-lecanemab-remain-too-small-to-justify-the-additional-costs-says-nice-in-final-draft-guidance
- NIED MOWLAS. The N-net network, completion of construction. www.mowlas.bosai.go.jp/network/?LANG=en
- npj Vaccines (2025). Reduced dementia risk after vaccines with the AS01 adjuvant. www.nature.com/articles/s41541-025-01172-3
- PAHO/WHO (10 October 2024). Better use of vaccines could reduce antibiotic use by 2.5 billion doses annually. www.paho.org/en/news/10-10-2024-better-use-vaccines-could-reduce-antibiotic-use-25-billion-doses-annually-says-who
- Papastavrou N., Horning D., Joyce G. (2024). RNA-catalyzed evolution of catalytic RNA. PNAS. www.pnas.org/doi/abs/10.1073/pnas.2321592121
- PLOS Computational Biology (2015). Steering evolution with sequential therapy to prevent the emergence of bacterial antibiotic resistance. journals.plos.org/ploscompbiol/article?id=10.1371%2Fjournal.pcbi.1004493
- PLOS Genetics (2010). Conjugative DNA transfer induces the bacterial SOS response. journals.plos.org/plosgenetics/article?id=10.1371%2Fjournal.pgen.1001165
- PLoS Neglected Tropical Diseases (2020). Rabies virus neutralising antibodies in livestock in Peru. journals.plos.org/plosntds/article?id=10.1371%2Fjournal.pntd.0008194
- PLoS Pathogens (2012). Immune clearance of attenuated rabies virus results in neuronal survival with altered gene expression. journals.plos.org/plospathogens/article?id=10.1371%2Fjournal.ppat.1002971
- PMC11458664 (2024). A commentary on the dating of the last universal common ancestor. www.ncbi.nlm.nih.gov/pmc/articles/PMC11458664/
- PMC11351851 (2024). Morbidostat: design of a continuous culture system for adaptive laboratory evolution. www.ncbi.nlm.nih.gov/pmc/articles/PMC11351851/
- PMC12542734 (2025). Metagenomic surveillance of the resistome in wastewater, 757 samples from 243 cities. pmc.ncbi.nlm.nih.gov/articles/PMC12542734/
- PMC3414554 (2012). Evidence of rabies virus exposure among humans in the Peruvian Amazon. www.ncbi.nlm.nih.gov/pmc/articles/PMC3414554/
- PMC4702264. Focused ultrasound-mediated drug delivery through the blood-brain barrier. pmc.ncbi.nlm.nih.gov/articles/PMC4702264/
- PMC7232326 (2020). Modelling lyssavirus infections in human stem cell-derived neural cultures. www.ncbi.nlm.nih.gov/pmc/articles/PMC7232326/
- PMC9313174 (2022). Increased antibody levels in the brain with focused ultrasound delivery. www.ncbi.nlm.nih.gov/pmc/articles/PMC9313174/
- Pomirchy M. et al. (2025). Herpes zoster vaccination and dementia occurrence (Australia). JAMA. pmc.ncbi.nlm.nih.gov/articles/PMC12019675/
- Powner M., Gerland B., Sutherland J. (2009). Synthesis of activated pyrimidine ribonucleotides in prebiotically plausible conditions. Nature. doi.org/10.1038/nature08013
- Preiner M. et al. (2020). A hydrogen-dependent geochemical analogue of primordial carbon and energy metabolism. Nature Ecology & Evolution. doi.org/10.1038/s41559-020-1125-6
- Richert C. (2018). Prebiotic chemistry and human intervention. Nature Communications. doi.org/10.1038/s41467-018-07219-5
- Rouet-Leduc B. et al. (2017). Machine learning predicts laboratory earthquakes. Geophysical Research Letters. agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017GL074677
- Rouet-Leduc B., Hulbert C., Johnson P. (2019). Continuous chatter of the Cascadia subduction zone revealed by machine learning. Nature Geoscience. www.nature.com/articles/s41561-018-0274-6
- Ruiz S. et al. (2014). Intense foreshocks and a slow slip event preceded the 2014 Iquique Mw 8.1 earthquake. Science (repository copy). repositorio.uchile.cl/bitstream/handle/2250/126941/Intense-foreshocks-and-a-slow-slip-event-preceded-the-2014-Iquique-Mw-8-1-earthquake.pdf?sequence=1
- Science (2016). Spatiotemporal microbial evolution on antibiotic landscapes (MEGA-plate). pmc.ncbi.nlm.nih.gov/articles/PMC5534434/
- Scientific Reports (2025). Aducanumab delivery across the barrier at different focused ultrasound pressures. www.nature.com/articles/s41598-025-02412-1
- Scientific Reports (2025). A fibre-optic cable off Sicily: the gain in detection time. www.nature.com/articles/s41598-025-29234-5
- Sims J. et al. (2023). Donanemab in early symptomatic Alzheimer disease (TRAILBLAZER-ALZ 2). JAMA. doi.org/10.1001/jama.2023.13239
- Smith E., Gomberg J. (2009). A search in strainmeter data for slow slip associated with triggered and tremor near Parkfield. JGR. agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2008JB006040
- Stark P. (1997). Earthquake prediction: the null hypothesis. Geophysical Journal International 131. academic.oup.com/gji/article-pdf/131/3/495/6100601/131-3-495.pdf
- Stockman S., Lawson D., Werner M. EarthquakeNPP: benchmark datasets for earthquake forecasting with neural point processes. arXiv:2410.08226. arxiv.org/abs/2410.08226
- Sun L. et al. (2026). A review of rabies therapy: the barrier, shuttles, monoclonal antibodies. Pharmaceutics, DOI 10.3390/pharmaceutics18070848. pmc.ncbi.nlm.nih.gov/articles/PMC13416311/
- Temblor (11 August 2023). Do earthquakes produce signals before they rupture? Maybe. temblor.net/temblor/do-earthquakes-produce-signals-before-they-rupture-maybe-15426/
- The Register (16 January 2026). NASA science budget: Mars Sample Return without funding. www.theregister.com/2026/01/16/nasa_science_budget/
- USGS. Can you predict earthquakes? www.usgs.gov/faqs/can-you-predict-earthquakes
- UT Jackson School of Geosciences (December 2024). Forecasting earthquakes with AI. www.jsg.utexas.edu/news/2024/12/forecasting-earthquakes-with-ai/
- van Dyck C. et al. (2023). Lecanemab in early Alzheimer's disease (Clarity AD). NEJM. www.nejm.org/doi/full/10.1056/NEJMoa2212948
- Wang K., Chen Q.-F., Sun S., Wang A. (2006). Predicting the 1975 Haicheng earthquake. BSSA 96(3). ui.adsabs.harvard.edu/abs/2006BuSSA..96..757W
- WHO (17 May 2024). WHO updates list of drug-resistant bacteria most threatening to human health. www.who.int/news/item/17-05-2024-who-updates-list-of-drug-resistant-bacteria-most-threatening-to-human-health
- WHO (2 October 2025). WHO releases new reports on new tests and treatments in development for bacterial infections. www.who.int/news/item/02-10-2025-who-releases-new-reports-on-new-tests-and-treatments-in-development-for-bacterial-infections
- WHO (17 September 2026). Rabies fact sheet. www.who.int/news-room/fact-sheets/detail/rabies
- WHO. Rabies: epidemiology and burden. www.who.int/teams/control-of-neglected-tropical-diseases/rabies/epidemiology-and-burden
- Wołos A., Grzybowski B. et al. (2020). Synthetic connectivity, emergence, and self-regeneration in the network of prebiotic chemistry. Science. doi.org/10.1126/science.aaw1955
- Yamada K. et al. (2016). Efficacy of favipiravir (T-705) in rabies postexposure prophylaxis. Journal of Infectious Diseases 213(8). academic.oup.com/jid/article/213/8/1253/2459381
- Zeiler F., Jackson A. C. (2016). Critical appraisal of the Milwaukee protocol for rabies: this failed approach should be abandoned. Canadian Journal of Neurological Sciences. pubmed.ncbi.nlm.nih.gov/26639059/
- Zheludev I. et al. (2024). Viroid-like colonists of human microbiomes (obelisks). Cell. doi.org/10.1016/j.cell.2024.09.033
- Zhu W. et al. (2023). QuakeFlow: a scalable machine-learning-based earthquake monitoring workflow. arXiv:2208.14564. arxiv.org/abs/2208.14564