In Navarre, 60 people aged 100 and over will be divided between a usual routine and 12 weeks of home exercises. On July 23, the BIOANCIENT_EX record appeared on ClinicalTrials.gov. The team will compare the physical function of participants before the program and after 12 weeks, when one group maintains their usual routine and the other exercises at home. When a centenarian walks and gets up better, observation alone does not separate the effect of movement from initial health: a stronger person more often preserves both function and the ability to train. BIOANCIENT_EX will test this by comparing two groups created by random distribution before the start of training. This will allow researchers to compare changes in physical function with the usual routine of participants. The trial plans to include 60 residents of Navarre who are 100 years old and can stand up and sit down on their own or with minimal assistance. One group will maintain their usual routine. The other will perform strength exercises, static and dynamic balance exercises, joint mobility, flexibility, and aerobic exercise at home. The team changes the complexity every three weeks, communicates with participants by phone or video, and visits them at home if necessary. The main measurement will be the SPPB - a physical function scale from 0 to 12 points. Participants will hold balance, walk a short distance, and stand up from a chair several times; the results of these tasks will be added up into one score. After 12 weeks, researchers will also compare hand and leg strength, autonomy in daily life, mobility, cognitive tests, and RNA profiles in plasma and extracellular vesicles. In a 2025 Spanish study, 12 participants completed 12 weeks of strength training in a nursing home: six trained under supervision, six formed a control group. In the training group, the median SPPB score increased from 2.3 to 5.0. BIOANCIENT_EX transfers the training to the home, plans to recruit 60 participants, and adds balance, flexibility, and aerobic exercise. The protocol will test the change in function in centenarians in this format of training.

🔗 Read original →

A neural network has restored an electrical model of a heart cell from a single recording. On July 23, a study was published in eLife on heart cells grown from reprogrammed stem cells. The authors recorded the response of two living cells to a specially calculated sequence of voltages and assembled a computational model from each recording. A cardiomyocyte contracts after an electrical impulse, which is created by ion channels in the membrane: some allow sodium and calcium to enter the cell, while others release potassium and return the voltage to its initial level. A typical recording shows the overall result of their work. To explain the shape of the impulse, a researcher needs the individual properties of each current. In the eLife article, the team of Pei-Chi Yang first created 1.1 million synthetic models of cardiomyocytes, modifying 52 parameters of six ionic currents and calculating how the cell would respond to different voltage commands. From this population, the authors selected a voltage clamp protocol: a device that holds a specified voltage on the membrane and measures the current in response to each command. The neural network was trained on synthetic recordings to find the parameters that generated them. Then, it was given a recording of a living cell and obtained a set of parameters for its model. From a single electrical recording, the authors determined the properties of the six ionic currents of a separate cell. In their calculations, such a model reproduced the measured electrical impulse of the cell. The authors tested this approach on two cardiomyocytes of one iPSC line, which are cells obtained by reprogramming an adult cell into a stem cell and then growing it towards heart muscle. Using the restored models, the team calculated the movement of calcium inside the cell, which triggers its contraction. The idea of modeling differences between heart cells emerged earlier. In a 2013 PNAS study, researchers matched a population of models to experiments on rabbit Purkinje fibers and compared predictions with responses to four concentrations of dofetilide. The new study automates the reverse path: from a single recording of a human iPSC cell to the parameters of its model. The authors obtain a cycle in which the cell measurement sets the model, and the model formulates a question for the next experiment. The parameters of the six currents turn the shape of the electrical impulse into a testable hypothesis about the cell's mechanism of action.

🔗 Read original →

On July 24, Norn Group and Effy Klimi published an essay "Sisyphus, MD" on the design of clinical trials. The authors write: an intervention may be calculated to affect several age-related diseases, but in a clinic, it is usually tested against one diagnosis. According to the authors' logic, a therapy that is supposed to affect several aging processes comes to the clinic through a specific disease: heart failure, Alzheimer's disease, or diabetes. The company recruits patients with this diagnosis, pre-selects a measurable outcome, and compares the treatment with standard care. A narrow group provides a clearer signal of benefit and risk, which is easier to discuss with regulators, doctors, and investors. Thus, the method of testing narrows the task of the drug to one diagnosis in advance. After the first approval, the term of market exclusivity limits the incentive to pay for trials for other diseases: they again require money, patients, and time. Therefore, a therapy designed for the general course of aging takes the form of a medicine for one disease. The authors suggest testing the treatment for two other outcomes. The first is to track the time until the first age-related disease. The second is to see if a person maintains mobility, avoids age-related frailty, and preserves their "internal potential": a combination of physical and mental abilities. Such a trial could measure not one diagnosis, but functional decline and several age-related risks. For such trials, methods are needed to reliably measure these outcomes. The ARPA-H PROSPR program is developing biochemical and physiological markers, home data collection, and protocols that should allow evaluating age-related outcomes over three years. In the Norn Group essay, this work is an example of infrastructure without which broad hypotheses about aging are difficult to test in humans.

🔗 Read original →

A White House report proposes infrastructure for mapping brain connections; neurobiologists debate whether it will crowd out fundamental science. On July 21, the White House Office of Science and Technology Policy released a report titled "Science: A New Golden Age". On July 24, The Transmitter gathered the reaction of neurobiologists to one of its examples: the Connectome, a map of contacts between neurons. In the OSTP report, its mapping is cited as a possible task for X-Lab, a National Science Foundation program that funds independent research organizations outside the traditional academic environment. Such teams receive operational autonomy and money based on achieved milestones. The document proposes a hypothetical mission: academia, industry, and philanthropists could together scale up maps of connections in several species of small mammals. It suggests starting with chains related to reward and motivation. The authors compare the desired effect to the Human Genome Project, after which sequencing became a routine service. Maps of such chains could help study depression, addiction, and autism. The debate is about who will be able to do science before the emergence of large infrastructure. Neurobiologist Jason Shepard fears that with a finite federal budget, large programs will narrow the space for research that laboratories choose out of curiosity. Jeff Lichtman, one of the pioneers of connectome visualization, notes that the methods of this field grew in universities, not from industrial demand. Jan Vesel supports long-term funding of missions, but according to him, a lot of fundamental work is still needed for the application of neuroscience in medicine. Former director of the National Institute of Mental Health Thomas Insel formulates the dilemma more simply: the country needs both industrial teams and university science. Otherwise, the future common machine may not have the ideas that it is supposed to scale.

🔗 Read original →

Frontier Bio has passed the first stage of the NIH competition with a blood-brain barrier model. On July 21, Frontier Bio announced its victory in the first stage of the NIH Complement-ARIE competition. The company presented NeuroTraX-AI: a human cell-based model of the blood-brain barrier, the state of which an algorithm evaluates from ordinary microscopic images. The victory brought a cash prize and advanced the company to the next stages of the competition with a total fund of $7 million. For a brain drug, it's not enough to just get into the bloodstream. It needs to pass through the blood-brain barrier - a layer of cells in the walls of brain vessels. It protects the brain, allowing some substances to pass through and retaining others. Therefore, it's crucial for developers to understand before human trials whether the candidate will reach brain tissue and damage the vascular protection. Frontier Bio grows a human cell model of the neurovascular unit: a fragment of the environment around a brain vessel. An ordinary microscope captures cells without stains, and NeuroTraX-AI quantitatively evaluates the state of the barrier from these images. According to the company's description, the system shows whether the barrier retains its properties, whether intervention disrupts them, and whether the test agent passes through it. The image of the cell culture becomes a measurement of the barrier's state. Instead of manual viewing of images, the developer receives an indicator for early verification of the candidate on a human cell system. This verification answers a practical question: how a substance behaves at the border between blood and brain. The Complement-ARIE competition supports laboratory and computational methods that can replace part of animal testing. The first stage selected a pair of a human barrier model and automatic analysis of its images. For developers of brain disease treatments, this is a way to earlier determine which candidates are worth moving to the next verification stage.

🔗 Read original →

On July 24, an article about UniPert-G2CP was published in Cell. The model teaches how a molecule changes the state of a cell by linking the results of genetic experiments with cell reactions to small molecules - chemical compounds that are often the starting point for drug discovery. A cell does not respond to a drug with a single button: a molecule binds to a target protein and then changes the work of many genes. The same drug can cause different reactions in different types of cells. Therefore, a chemical screen requires multiple separate experiments: for each molecule, researchers treat cells and measure which RNAs have changed in them. A genetic screen provides such a map of states much more widely. In a common variant, CRISPR turns off a single gene, after which researchers take a profile of the cell's RNA. The map shows what state the intervention in each gene causes. In May, TxPert predicted the transcriptomic response to new genetic interventions; UniPert-G2CP links such maps to the structure of small molecules. The Connectivity Map in the LINCS program is already collecting profiles of genetic interventions and compounds. These can be used to match the traces that genes and drugs leave in cells. UniPert-G2CP uses this map for chemical search. The model receives a target protein and the structure of a molecule, translates them into a common representation, and matches them with the measured cell profile. Genetic experiments show how a cell responds to the shutdown of a specific gene; chemical experiments link similar profiles to real molecules. The authors checked the transfer on LINCS data: 4,994 genes and 7,860 molecules in five cancer cell lines. When one-fifth of the chemical measurements were left for training, pre-training on genetic screens increased the average correlation of the predicted and measured transcriptome by 375.4%. This figure describes the coincidence of two RNA profiles, not the number of drugs found. To search for substances that change the age-related states of cells, the same sequence is needed: describe the cellular effect of a protein target and select molecules that can cause the desired profile. UniPert-G2CP provides a way to narrow down this selection to laboratory verification of candidates.

🔗 Read original →

Asimov is creating a system at Boston University to test whether a therapeutic protein can be manufactured. On July 24, Asimov announced that it will design variants of therapeutic proteins on a computer, produce them in the DAMP laboratory at Boston University, and measure the properties of the finished product. These results will be fed back into the model, which will select the next sequence. A therapeutic protein starts with a sequence of amino acids, but that's not the end of the work. It still needs to be produced in cells, purified, and tested. Sometimes, it's at this stage that it becomes clear that a design that was successful on the computer is difficult to turn into a product. In the July 24 announcement, Asimov describes the sequence of actions: the company designs protein variants, the laboratory produces them and measures the manufacturing-related properties, and the model receives the results. The result of each experiment will become data for selecting the next batch of variants. The manufacturability of the protein will thus be taken into account in the decision on its sequence even before the next cycle of experiments. DAMP is the design, automation, manufacturing, and processes laboratory at Boston University. It already performs remotely ordered experiments with biological materials, chemicals, and liquids. Asimov will create a section there for working with therapeutic proteins and will become one of the laboratory's first users. The BU laboratory is part of the Programmable Cloud Laboratory Test Bed network. On July 22, the US National Science Foundation allocated $380 million over four years to the network for 20 laboratory nodes; the Astera Institute adds up to $20 million. BU will receive up to $20 million for its node. The network is expected to allow laboratories to use common methods and experiment results. Asimov is linking protein design to what happens during its actual production. A protein that can be produced and measured provides the model with an example for the next selection; a protein with poor manufacturing properties helps to weed out similar sequences earlier.

🔗 Read original →

The Verge: Genesis Mission has selected 278 AI projects, and laboratories and universities are to verify their hypotheses. On July 24, The Verge published an analysis of the first projects of the American Genesis Mission - a government AI program for science. Robert Hart links it to the new White House plan and asks if science has enough people and places to test machine proposals. On July 22, the US Department of Energy selected 278 Genesis Mission projects for grant negotiations. Teams from national laboratories, universities, companies, and non-profit organizations will gain access to computational power, models, and programs for research. The White House plan "Science: A New Golden Age," published on July 21, suggests relying on individual researchers, new organizations, and partnerships with companies. The grant provides the team with calculations and models; hypothesis verification remains the work of researchers and laboratories. AI can propose a molecule, material, connection in data, or experiment scheme. A researcher sets up a control experiment, a laboratory obtains a measurement, and another group repeats it on new samples. In The Verge's analysis, physicist Andreas Karth describes the risk: AI will produce many plausible options, and there will be fewer people with experience to filter them. "AI may have several big ideas buried under mountains of useless material, and there will be far fewer people with experience to distinguish one from another." In biomedicine, after laboratory experiments come clinical trials: they determine whether a person's condition changes. Each transition from model to experiment, repetition, and clinic filters out some convincingly sounding hypotheses. The debate about Genesis Mission concerns this entire chain: calculations accelerate the search for options, and universities and laboratories teach people to set up experiments, verify results, and prepare the next generation of researchers.

🔗 Read original →

The Pan-human Azimuth project has unified data from 23 tissues into a single hierarchy of cell types. On July 21, the HuBMAP team released the Pan-human Azimuth preprint. The neural network annotates individual human cell profiles according to a unified hierarchy; to train it, the authors collected 27.04 million profiles and launched a cloud service, as well as tools for R and Python. Single-cell atlases show which genes are active in each cell, and based on this, a biologist determines the cell's type and state. However, one atlas may call a cell by one name, and another by a different name. When comparing organs, the difference in names can be easily mistaken for a difference in biology. In the preprint by Surava Sarkar and colleagues, the original names from different datasets were manually matched to a single tree of cell types. The authors checked questionable assignments based on gene activity and trained a classifier on this annotation. It reads a table of RNA activity, assigns a cell to one of the eight levels of the hierarchy, and shows the confidence of the response. Empty droplets and background RNA received a separate class. Now, similar cells from different organs receive names according to the same rules. The authors tested the classifier on 1.1 million profiles from Tabula Sapiens v2: the data came from 24 donors and covered 28 tissues; about 600,000 profiles from nine donors the model saw for the first time after training. In the same way, it annotated 85.9 million cells from scBaseCamp. On spatial kidney slices, the classifier reproduced the pattern of cortical matter and separated healthy and sclerotic glomeruli in agreement with the annotation of pathologists. The unified hierarchy allows querying data about a specific cellular state immediately in several tissues and in different people. For example, it is possible to check if the same state of fibroblasts, immune cells, or epithelium is repeated in different organs. The atlas of 7 million mouse cells has already shown synchronized age shifts between organs; Pan-human Azimuth gives human data a common language for such verification. The project documentation calls this approach comparing a new sample to a collected reference book. Pan-human Azimuth turns the chaos of cell names into a reproducible procedure. After such annotation, inter-tissue comparison can be checked on tens of millions of profiles.

🔗 Read original →

Insilico Medicine selected the AI-created molecule ISM9528 for pain treatment on July 22. Insilico Medicine named ISM9528 a preclinical candidate for pain treatment. The company plans to start the first human trial in 2027; the molecule is intended to be taken orally. In the Insilico release, the drug target is designated as Target Z. This is a provisional name for the biological target that the molecule is supposed to act on. The PandaOmics platform matched gene and protein function data, scientific publications, and other biological information, and then highlighted Target Z as a potential target for pain. After that, the team checked how this target is represented in pain-related cell types in humans and animals. Then, Chemistry42 proposed options for the chemical structure. The team selected molecules based on their effect on Target Z, ability to penetrate the blood-brain barrier, and properties needed for a future drug. The blood-brain barrier is a filter in brain vessels that only allows certain substances to pass through. After several cycles of selection, Insilico chose the molecule ISM9528. In animal studies, the company compared ISM9528 to pregabalin, a drug for neuropathic pain. In a nerve damage model in rats, the average dose of ISM9528, according to Insilico, relieved pain for up to six hours, just like pregabalin. In a postoperative pain model, the company reports that the effect began within half an hour and was stronger than that of an equal dose of pregabalin. Drug Target Review recounts these models and the plan for the first human trial in 2027. Insilico already has a program that has progressed further: rontosertib, for which AI chose the target TNIK and designed the molecule, showed results in Phase IIa for pulmonary fibrosis. ISM9528 is currently at the very first stage of this process. ISM9528 became Insilico's 31st preclinical candidate since 2021. The company reports that 13 previous candidates have received permission to start clinical trials.

🔗 Read original →

The Biomarkers of Aging Consortium will devote an entire session to replacement in aging, with the Replacement in Aging program scheduled for October 5 in Boston. The program will discuss cellular and tissue replacement, organ engineering, transplantation, immune system renewal, and neuroregeneration. On the session page, the organizers ask: can an aging organism be maintained by replacing failing parts with young, grown, or synthetic components? The discussion will take place throughout the day, from 9:00 to 18:00. The program features Vadim Gladyshev, Anthony Atala, Linda Griffith, and Hugh Herr. Transplantation, tissue engineering, immune system renewal, and neuroregeneration work with different parts of the aging organism. The session brings these areas together around one question: what part of the body can be replaced to restore the necessary function? In a separate operation, the focus is usually on whether a specific organ or tissue is working. An anti-aging strategy requires a longer check: what happens to cells, tissue, organs, and people over time. In their May roadmap on replacement-based interventions, Gladyshev, Atala, and colleagues propose evaluating replacement at each of these levels. The perspective in Aging Cell, on which this roadmap is based, distinguishes between the short-term outcome of a procedure and long-term functional restoration. Replacement can be evaluated as a strategy when measurements show: has the function returned and is it being maintained? Therefore, the biomarkers conference places cell, tissue, organ, and function replacement alongside aging measurement and intervention testing.

🔗 Read original →

Infinita has launched a council in Montana that accepts therapies after the first phase of clinical trials. The Montana ETRB will review access to experimental therapies before the usual FDA approval. A developer can submit an application for $12,500 after the first phase - the first human trial, where safety is primarily checked. On July 25, Infinita announced the launch of the Montana Experimental Treatment Review Board, or ETRB. In the announcement, the company stated its task directly: "Infinita has formed the first ETRB under Montana law. For $12,500, you can submit a therapy to our board after the first phase of clinical trials." A public entry point for early access to therapy has appeared: the board's website already leads to a submission form. The board checks if the safety of the intervention is sufficiently described, how the patient will be monitored, and how informed consent is formalized. Among the listed participants are aging biologists Matt Kaeberlein and Felipe Sierra, bioethicist Jessica Flanigan, Jamie Justice from XPRIZE, and oncologist James Burke. Montana's SB 535 law provides this route. It allows consideration of drugs, biological products, devices, and other interventions after a successfully completed first phase, if they are still being studied in an FDA-approved study or have a documented safety history. The patient discusses available approved options with their treating physician, receives their recommendation, and signs consent. The law divides the work between the board, clinic, and state. According to the Montana Department of Health's draft rules, the board reviews the protocol, safety data, monitoring plan, and patient consent. A licensed experimental center performs the treatment and reports serious adverse events to the department within five days. In June, Niklas Angering described the Próspera and Montana SB 535 connection as a future route: first, early safety testing in humans, then access through an American state. Now, this scheme has a board with a submission form and named participants. The developer collects a dossier on the therapy, the board reviews it, the clinic treats the patient, and the state receives safety reports.

🔗 Read original →

1000ExM increases the bio-sample by a thousand times; calculation shows differentiation of neighboring amino acids. On July 21, Helena Hu, Ed Boyden, and co-authors published 1000ExM on bioRxiv - a method that increases the fixed biological sample approximately a thousand times in length. The authors checked the preservation of structure on several proteins and a peptide, and then modeled protein recognition according to such a spatial map. Light microscopy merges close points into one spot. Neighboring amino acids in a protein chain are separated by about 0.38 nanometers; conventional optics cannot see such a distance. 1000ExM expands molecules to a scale visible to conventional microscopes. Researchers chemically attach side chains of amino acids to a swelling gel, cut the protein chain between the attachment points, and add water. The gel expands the attached fragments. Four polymer networks repeat this expansion: the sample grows approximately a thousand times along each axis, and its volume - approximately a billion times. The distance of 0.38 nanometers is transformed into approximately 380 nanometers - which can already be distinguished by a confocal microscope. Early variants of expansion microscopy provided a total linear magnification of about 16-22 times. After each round, they added a neutral gel to hold the stretched sample. In Hu's and colleagues' work, the next charged network is formed directly inside the already swollen gel. This allowed the expansion to be repeated four times. In the Eon experiment, neuron microcultures were expanded approximately 20 times to match their wiring with recorded activity. 1000ExM continues the same trend at the level of protein labels: after expansion, their geometry in the model becomes distinguishable by conventional optics. The authors compared the resulting maps with the known structure of nanotubes, the mCLING peptide, and GFP - the green fluorescent protein. Thus, they checked whether the labels preserve the geometry of the original molecule. Then, in a calculation for 23,391 canonical human proteins, they took into account missed labels and measured gel deformation; for most proteins, the label pattern turned out to be unique. In the calculation, the label pattern becomes a signature of the protein. Fragments with attached side chains leave spatial points; their arrangement can distinguish proteins. The gel increased the original distances to the scale of conventional confocal optics. On individual molecules, the authors checked the preservation of label geometry. According to their model, such signatures can be obtained inside cells and tissues.

🔗 Read original →

Sam Altman believes that the singularity has already begun and described a chain where robots build new data centers. On July 25, in an interview with Ti Mors, OpenAI CEO Sam Altman called the current moment the singularity. He suggested imagining automated production, where a data center controls robots, and the robots build the next center. Altman calls the singularity an already ongoing "crazy exponent": there is no single breakthrough moment, but the current period, he says, determines where the curve will go. Ten years ago, this perspective seemed distant and unlikely to him; now he says that humanity is already inside the process. The next stage, he associates with the physical infrastructure of computing. In a fully automated chain, a data center spends part of its power to control robots. The robots build another center, and its power returns to the same cycle. Altman suggests looking at this possibility as follows: "Too much attention is paid to algorithms that improve algorithms, and too little to data centers capable of creating new data centers." This scheme runs into electricity, microchips, and construction. The International Energy Agency compares the consumption of a typical AI data center to the consumption of 100,000 households. In its basic scenario, global data center consumption will grow from 415 TWh in 2024 to approximately 945 TWh by 2030. According to Altman's model, computations become part of the production of their own material base. In the same conversation, he named transistors and then electricity as the main limitations: they determine whether a new data center can launch another round of construction.

🔗 Read original →

The NIH has launched the Bio Genesis Mission, a program that connects AI, medical data, and clinical research. On July 22, the US National Institutes of Health announced the Bio Genesis Mission, its biomedical contribution to the federal Genesis Mission. The NIH aims to cut in half the time it takes to go from a scientific discovery to its impact on health over the next five to ten years, with the mission already accounting for more than $1.2 billion in commitments for the 2026 fiscal year and planned funding for 2027. To bring new therapies to patients, researchers need to match molecular and genetic information, data on observed signs of disease and patient health, select a candidate, and test it in a clinical trial. The Bio Genesis Mission is intended to bring this path - from data to the clinic - within a single program. The overall Genesis Mission has already selected 278 projects for grant negotiations, in which computations generate hypotheses and laboratories test them on samples; Bio Genesis adds clinical research to this chain, where changes in human health can be seen. The NIH describes six areas of focus for the mission: predicting the behavior of living systems, biomanufacturing, early detection of biological threats, childhood cancer, the development and clinical application of drugs, and the search for the causes of chronic diseases. These tasks define what the NIH plans to combine data, computations, and clinical infrastructure for. For drugs, the plan is described in more detail. According to a White House statement, the Department of Health and Human Services, the Department of Energy, and the Department of Defense will create an infrastructure that combines molecular, genetic, clinical data, and information from everyday medical practice. On this basis, AI should search for new applications of already known drugs and help bring new therapies to the clinic. NIH Director Jay Bhattacharya explained the idea through a patient: "People with cancer, chronic, or rare diseases feel the timeline of research as years of waiting for an answer." The entire Genesis Mission spans more than 15 federal agencies and relies on a common platform with data, computations, and AI tools. In the NIH director's statement, Bio Genesis links shared resources to the search for the causes of chronic diseases, the development of drugs, and their testing in humans.

🔗 Read original →

On July 21, Brian Johnson announced that induced pluripotent stem cells (iPSCs) had been derived from the cells of his blood. He referred to the culture as a "newborn clone" and linked it to future body repair: from testing therapies to growing tissues and organs. In the post, Johnson writes: "I just cloned myself... newborn." This refers to a cell line in a Petri dish: his blood cells had been reprogrammed into iPSCs. Since different cell types can be obtained from iPSCs in the laboratory, Johnson associates his line with testing therapies, growing tissues and organs for transplantation, and introducing "young cells." In a 2007 study, a group led by Shinya Yamanaka obtained such cells from adult fibroblasts - cells of connective tissue. The scientists applied four transcription factors, proteins that change gene function. After reprogramming, the cells acquired properties that allowed them to be used to obtain tissues of the organism in experiments. To obtain a specific tissue, the cells are directed towards the necessary specialization, the tissue is grown, and it is checked whether it works in the body. One original line can serve both as a disease model and as material for cell therapy: in these cases, it is given different tasks. This chain has already reached early clinical trials for nerve tissue: in a phase 1 study, four people with cervical spine injuries were administered two million neural precursors grown from iPSCs. The safety of the transplant was being tested. Johnson suggests considering this cell line as a personal reserve for body repair. The material was taken from himself, and he describes the replacement according to functions: vision, movement, blood formation, and organ function. According to his model, the body can be repaired consecutively, using one's own cellular material for each new task.

🔗 Read original →

After 30 minutes at -6 °C, 24 out of 30 newborn mice survived. On July 20, a study on supercooling was published in Scientific Reports: water in the body remains liquid at a temperature below zero. After half an hour at -6 °C, 24 out of 30 mice survived; after the same amount of time at 0-4 °C, none of the 28 mice survived. The authors of the article worked with mice aged one to five days. Each weighed about a gram and quickly lost heat. They compared 30 minutes of normal cooling at 0-4 °C with 30 minutes of supercooling at -6 °C. After normal cooling, all 28 mice died, while after supercooling, 24 out of 30 survived the first day. During supercooling, water remains liquid below the freezing point. Ice damages cells with growing crystals, so the researchers kept the mice dry in open plastic bags and cooled them with air: this way, there are fewer places where crystals can start to grow. Then, the animals were gradually warmed up in an incubator, oxygen was supplied, and breathing was manually stimulated. According to the authors' hypothesis, at -6 °C, stronger suppression of metabolism helps cells survive storage. In 2019, the same approach helped preserve human livers at -4 °C: after machine perfusion, they remained viable outside the body for 27 hours longer. Two days earlier, pig kidneys after three days at -4 °C started working again after transplantation. The liver and kidney are stored separately from the body; here, the researchers tested a short pause below zero in a whole mammal. The surviving mice were observed for up to three months. Until a month old, they gained weight at the same rate as the control animals. Then, the authors checked coordination, running endurance, sensitivity, gait, and learning of a fear response; the article and its abstract report that no differences with the control were found in these tests. In one experiment with a whole organism, the researchers checked whether it is possible to keep the body below zero without ice, warm it up, and monitor the return of breathing, movement, and development. For biostasis, this is the next step after preserving individual organs.

🔗 Read original →

In Alzheimer's disease, active and less active DNA segments are poorly separated in brain cells. In a Science article published on July 23, teams from Carnegie Mellon, the University of Pittsburgh, and the University of Washington measured gene activity and three-dimensional DNA contacts in the same brain cells. The Hicformer model showed that these contacts help predict disease-related changes in gene function. DNA in the nucleus is not just a long thread; segments with frequently read genes are usually located near other active segments, while less active segments are gathered separately. The fragments of DNA that end up next to each other determine whether a regulatory segment can influence gene function. The authors applied GAGE-seq to single cells from the post-mortem prefrontal cortex of people with Alzheimer's disease and those without the diagnosis. The method simultaneously reads gene activity and contacts between distant chromatin segments - the substance that chromosomes are made of. Then, a spatial map of the tissue showed where in the cortex cells with different gene functions and DNA layouts were located. In people with Alzheimer's disease, large active and less active zones of the genome were poorly separated from each other. The authors call this a mixing of compartments. In several cell types, they also found fewer close DNA contacts and more distant ones. The connections between genes and neighboring regulatory segments weakened; at the same time, in neurons, the function of genes associated with synapses decreased, and in microglia - the immune cells of the brain - stress, metabolism, and cellular aging programs changed. Hicformer combines DNA sequence, its overall layout, and local contacts to predict gene activity in different cell types. According to the authors, without knowledge of three-dimensional contacts, the model was worse at finding disease-related changes in gene activity. That is, the DNA contact map adds another way to search for cells and regulatory segments for future experiments to the list of included and excluded genes. Amyloid plaques and tau tangles remain noticeable signs of the disease. This work shows another level of its picture: in cells, not only the set of working genes changes, but also the arrangement of DNA segments that can control their function.

🔗 Read original →

Canadian surgeons transplanted islet cells from a pancreas that had been perfused with blood for 124 minutes after circulatory death. On July 24, a Canadian team described a clinical case: after confirming circulatory death of the donor, they restored blood flow to the abdominal cavity, isolated 400,000 islet equivalents from the pancreas, and infused them into a 69-year-old patient with type 1 diabetes. Pancreatic islets produce insulin and are isolated from donor organs and transplanted into people with type 1 diabetes who have lost the ability to sense dangerous drops in glucose. After circulatory death, tissue rapidly depletes its energy stores, and the islets must then survive enzymatic and mechanical processing in the laboratory. In the July 24 article, Vancouver doctors described how, after a mandatory five-minute wait, they connected the donor's aorta and inferior vena cava to a normothermic regional perfusion system, which pumped oxygenated blood through the abdominal organs at body temperature for 124 minutes. The aorta leading to the brain was clamped by surgeons, allowing blood to return to the tissues before the pancreas was cooled. The donor's lactate level, which accumulates with oxygen deficiency, decreased, and glucose levels remained stable. The pancreas was then cooled and transferred to the laboratory. From the organ, 400,000 islet equivalents were obtained - a standard measure of islet tissue volume. The cells preserved 92% viability, which was sufficient for a single infusion. The donor had a low body mass index, which usually reduces the yield of islet tissue. Four weeks after transplantation, the patient reduced her daily insulin dose from approximately 21 to 6-7 units. The time spent with glucose in a safe range increased from 74% to 96%, and the time spent below 3.9 mmol/L decreased from 24% to 3%. C-peptide in her blood increased from 184 to 832 pmol/L, a marker that appears along with insulin produced by the patient's own tissue. Normothermic regional perfusion is already used for liver and kidney retrieval after circulatory death. In Vancouver, doctors applied it to the pancreas: restored oxygen to the organ before cooling, isolated islet cells, and transplanted them into the patient.

🔗 Read original →

NVIDIA has released the weights of JEPA-DNA, a DNA model that learns to predict the representation of an entire sequence. On July 16, NVIDIA published the weights, code, and reproduction parameters of JEPA-DNA. The method adds a task to the usual training of genomic models: to restore the internal numerical description of the entire sequence from a visible fragment. A genomic model reads DNA as a string of four letters. Usually, it restores closed letters or predicts the next one. However, the work of a segment is often determined by a combination of elements on a long sequence: a promoter informs the cell where to start reading a gene, and a splicing site helps assemble RNA. In the JEPA-DNA preprint, first published on February 19, the authors left the task of guessing nucleotides and added a second one. One encoder receives a sequence with closed fragments; another sees the complete sequence and creates its internal numerical description. The predictor learns to obtain this description from the available environment. According to the authors' idea, this goal should help the model consider the function of the segment together with the data on neighboring letters. The verification covered three original models: DNABERT-2, Nucleotide Transformer v3, and HyenaDNA. Nine tasks measured the quality of features after training a simple classifier, and eight more checked DNA variants for similarity of representations of the original and modified sequences. For HyenaDNA, the AUROC in the promoter recognition task increased from 0.686 to 0.763. In the transcription factor binding task, the same metric decreased from 0.698 to 0.638. The same backbone gives different results on different tasks: the usefulness of the JEPA goal needs to be checked in a specific setting, rather than inferred from the average increase. NVIDIA has released the weights of HyenaDNA, and the repository contains instructions for three sets of weights and launches on GFMBench. This set of tasks tests DNA models on promoters, splicing, and the consequences of genetic variants. The weights are available under a non-commercial license. A laboratory can compare the original HyenaDNA and the JEPA-DNA version on its own task with DNA sequences. The open weights and reproduction parameters allow for testing the method beyond the tables of the preprint.

🔗 Read original →

On July 16, ProofAtlas published verifiable code in Lean, evaluating the time it takes for almost all Collatz trajectories to drop. On the same day, ProofAtlas released two formal declarations and tied them to a specific version of the source code. One of them asserts: for almost all positive starting numbers, the Collatz trajectory drops below a growing threshold in no more than 436 * log N ordinary steps. The Collatz problem sets a simple rule for an integer N: an odd number is replaced by 3N + 1, an even number is divided by two, and the process is repeated. The resulting sequence sometimes grows for a long time; therefore, it is difficult to estimate when it will first drop noticeably below the starting point. "Almost all" has a precise meaning here. Among numbers from 1 to X, the proportion of starts with such a drop tends to 100% as X grows. The threshold also grows with the starting point: for example, you can take √N. Then, for almost all N, the trajectory will drop below √N in 436 * log N steps. This constant counts each move, including divisions by two. In the same file, there is an estimate of 145 * log N for the Syracuse iteration. It immediately divides the result by all powers of two after the 3N + 1 operation until the next odd number. 436 and 145 cannot be compared directly: these are different ways of counting steps and different sets of starting points. Terence Tao proved a result of the same type for logarithmic density in 2019: the trajectory of almost every starting point reaches any growing threshold. ProofAtlas recorded a variant for ordinary natural density and ordinary Collatz steps in code that can be opened by a fixed version. The formal record gives the dispute a precise subject. Instead of retelling the result, one can check the formulation of the theorem, the condition on the threshold, and the way the steps are counted.

🔗 Read original →