US federal agencies searched for active scientific grants by keywords before terminating funding. Agreements with federal agencies, attached to a court petition, describe the general procedure: first, grants were found by words and themes, then some of them were selected for termination of funding. In individual cases, artificial intelligence tools were used when compiling lists. On July 15, the plaintiffs in the case of Thakur v. Trump filed a petition in federal court and attached agreements with agencies. The National Science Foundation, the National Foundation for the Humanities, the Department of Defense and Transportation, and the US National Institutes of Health described in them how they selected grants for consideration for termination. "Search terms, keywords or phrases" highlighted grants for review. According to the text of the agreements, then projects that expressed or allegedly expressed positions not supported by the administration were selected from the list. The agencies applied general criteria and standard letters, rather than separately checking whether each grant recipient had met the conditions. The same agreements state that the recipients did not violate the terms of funding. At the National Institutes of Health and the Department of Health, the list of themes was expanded: from projects on diversity, equality, and inclusion to gender, vaccine skepticism, and COVID-19. The lists included "health equity", "structural racism", and "sexual orientation". The materials do not explain what data the AI tools processed and how their results influenced the decisions. The agencies only indicated that AI could participate in preparing the lists provided to it. The Associated Press reported on July 21 that the case concerns more than a thousand grants from the University of California. The plaintiffs estimate the previously awarded funding at approximately $2 billion. A hearing on the petition is scheduled for October 20, and the court has not yet decided whether this procedure is lawful. This differs from the White House plan to experimentally test the rules for issuing new grants, where it is proposed to compare the method of selecting applications with research results. The Thakur agreements describe the review of already issued grants by thematic words and the alleged position of the project. For long-term biomedicine research, continuity is important: a team is hired for a grant, participants are recruited, and experiments are conducted for years. The materials do not allow us to establish the consequences for each laboratory, but the described procedure creates a risk for such programs: the funding of an already started study may be revised without a separate assessment of whether the laboratory has met the conditions of the grant. This applies to aging research as one of the long-term biomedical areas; individual gerontology programs are not named in the documents.

🔗 Read original →

On July 22, Haowei Man's team published an article in Nature about ContactSeek. The method combines AlphaFold3 predictions for the protein-RNA-DNA complex with sequencing data and suggests amino acid substitutions for more selective DNA editors. A base editor changes one DNA letter without a double-strand break. A guide RNA leads it to the desired sequence, a CRISPR protein holds the complex on the DNA, and an attached enzyme performs the chemical substitution. Similar sequences sometimes also hold this complex, and then the editor changes letters outside the chosen target. Usually, engineers screen amino acid substitutions in the protein and measure the result. ContactSeek starts with the error traces of the original editor: the authors found genome-wide sites where it came together with the guide RNA and submitted the target and these site sequences to AlphaFold3. The model built complex variants from protein, RNA, and DNA. Here, AlphaFold3 builds a map of likely contacts within the complex. ContactSeek compares the differences between the target and off-target sites with the sequencing signal. This is how the program identifies amino acids whose contacts with RNA or DNA change along with the frequency of off-target edits and suggests substitutions for cellular experiments. The authors tested this route on the adenine editor Cas9 and the cytosine editor Cas12a in HEK293T cells. The ABE8e-DD variant had a total guide-dependent off-target editing signal 99.2% lower than the original ABE8e: on the ABEsite16 target, the comparison covered 270 off-target sites. For the cytosine editor variant on Cas12a, the signal decreased by 82.1-95.1% for four guide RNAs. The authors also measured target editing, off-target RNA changes, and editing without guide RNA. Base editors have already reached patients: in personalized therapy, the KJ Malton editor was delivered to the liver with lipid nanoparticles, and after a year, clinical improvement was noted in a child without serious side effects. ContactSeek has been tested only in HEK293T cells, so its variants still need to be tested for delivery to tissues and in long-term observations. ProteinMPNN helped obtain prime editor variants that accumulated better in cells. ContactSeek solves a different problem: it narrows down the list of places where the editor can trigger outside the target. The contact map turns a broad screening of amino acid substitutions into specific hypotheses for experimentation.

🔗 Read original →

User LessWrong bits suggested that people will start making plans for future anti-aging therapy before it appears. On July 20, bits published an essay on when radical life extension will enter people's personal plans. This refers to increasing healthy and overall life beyond what current medicine and habits provide. In the essay, bits asks when a person will decide that they will be able to take advantage of future therapy. Then the planning horizon changes: decisions about retirement, savings, children, and career depend on the new understanding of how much time is left. To explain this, bits uses the concept of Turkish-American economist Timur Kuran - preference falsification. A person may want to live significantly longer, but publicly repeat the usual norm as long as such a desire seems like a strange fantasy. When future therapy seems plausible, people start talking about it openly, and it becomes easier for the next person to do the same. This model is based on two assumptions: many people already want a long life, but hide this desire; a change in norms will allow them to speak out. In the essay, this is an explanatory hypothesis, not a measurement of public opinion: the author does not provide data on the scale of hidden demand or the speed of the cascade. In his scenario, the cascade can be triggered by signals that are easy to see and retell: a recognized scientific result of rejuvenation in humans, a notable discovery using AI, a new therapy in the public eye, or a regulator's permission to measure aging in clinical trials. Such signals make future therapy a subject of ordinary conversation. Then, bits suggests, demand for research, money, and political decisions change. In a comment to the essay, Dagon suggests a different sequence. "The strongest factor in expected lifespan - measured or at least claimed life extension or significant health extension in old age," he writes. A wide audience, in his opinion, will change expectations after a result; early supporters are able to believe earlier and be wrong about the timing. The dispute concerns the order of events. According to bits' model, public expectation is able to gather support for future research. According to Dagon's model, a visible result appears first, which people trust. If people start changing plans before such a result, bits' hypothesis will gain support. If expectations shift after results, Dagon will be proven right.

🔗 Read original →

Colossal is in talks for funding at a valuation of $20-30 billion. On July 20, Axios reported that Colossal Biosciences is in negotiations for new funding at a valuation of $20-30 billion. In January 2025, the company raised $200 million at a valuation of $10.2 billion; Axios writes that over the past year, it has begun to generate revenue. Colossal is known for its programs to bring back the mammoth, dodo, and woolly wolf. At the same time, the company is building biobanks, genetic medicine programs, and reproductive technologies. These areas have grown out of work with the genomes of extinct animals, but they already have their own customers and products. Axios writes that Colossal is creating repositories of genetic material of endangered species together with the US government and the Dubai Museum of the Future. In June, a memorandum with the U.S. Fish and Wildlife Service on a cryobank of tissues, cells, and genomes gave this work a specific shape. This is a conservation service that can be developed in parallel with long-term de-extinction programs. Colossal also has a biomedicine line. Form Bio, spun out of it, helps teams creating genetic medicines: selecting designs for gene therapy and parsing production data. TechCrunch also mentions companies Breaking and Astromech, which have grown out of Colossal. Thus, work with animal DNA is being transformed into tools for drug development and services for government conservation programs. According to the author's model, the $20-30 billion range refers to a company where de-extinction remains the most notable project, and biobanks, genetic medicine services, and spun-out companies are already working alongside it. The negotiations assess the entire set of technologies, not just the promise of one day bringing back an extinct animal.

🔗 Read original →

Antonio and Pablo Acuaviva described five results in Banach space theory for which the model proposed ideas and drafted proofs. On July 19, Antonio and Pablo Acuaviva posted a preprint on five results in Banach space theory. The authors write that the model found key moves and wrote draft proofs, which they then checked, corrected, and formalized into final arguments. Banach spaces allow for measuring distances between objects and working with infinite sequences; in this language, mathematicians describe functions and function transformations. One missing hypothesis here breaks the proof, so the authors read each step, corrected errors, and verified the arguments against previous theorems. In four of the problems, according to the Acuavivas, the model's original answers already contained the basic proof: the authors had to correct references, individual errors, and the presentation. In the fifth, the model proposed a plan for a long proof, and the authors filled in the transitions and assembled a coherent argument. In the same work, the authors describe the search for problems in the scientific literature. Scripts took original article texts from arXiv and marked phrases like "question", "problem", and "hypothesis". Then, an agent read a fragment, searched for a proof, counterexample, or already known answer, and saved the result in a package for a mathematician. The package contains the original article, the exact formulation of the question, the course of the argument, and the found references. The main run went through a queue of 1,433 articles on functional analysis. Mathematicians spent individual packages on parsing: 31 received the status of verified, 10 were rejected. The five results from the first part of the preprint grew out of problems that the authors chose themselves, not from this queue. Two different modes emerged: in one, the model works on a task selected by people, and in the other, it searches for questions in articles and prepares material for analysis. In the story with the lower bound for convex optimization, the argument is recorded in a formal language, and Lean checks it line by line. With the Acuavivas, the proofs are checked by the mathematicians themselves: their package assembles material for analysis, rather than replacing verification with a formal certificate. With FunSearch from Google DeepMind, the model also proposes options, but in the form of programs: an automatic evaluator launches the code and selects successful ones. A proof cannot be launched in this way: a mathematician must check each step, match the theorem with the question and with the published literature. The model quickly iterates through moves, and the mathematician chooses a question, checks the argument, and searches for its place among known results. In this work, the model's answer turns into a research result only after such verification.

🔗 Read original →

Kolon TissueGene reported preliminary results of the ACTiVION-II trial on July 20: in 531 patients, a single injection of TG-C did not improve knee pain and function better than saline placebo after one year. The company also reported that the therapy did not achieve all pre-specified secondary endpoints. TG-C combines donor cartilage cells and genetically modified cells that produce TGF-β1 protein. Developers tested whether a single injection into the joint could alleviate symptoms of knee osteoarthritis, a disease that makes ordinary movements painful and stiff. In ACTiVION-II, doctors randomly assigned patients with knee arthritis to either TG-C or saline placebo. After 12 months, they compared pain on the Visual Analog Scale (VAS) and knee function on the WOMAC questionnaire, which asks about pain, stiffness, and everyday movements. According to Kolon TissueGene's release, TG-C did not provide a statistically significant benefit in either pain or knee function. The company also reported a comparable frequency and severity of adverse events in both groups. This is a separate outcome: patient safety and benefit require different data. So far, the company has only disclosed a preliminary summary without effect size, p-values, and subgroup analysis. ACTiVION-II tested a single injection of TG-C in patients with knee arthritis. Its result describes this therapy and this patient group; data from a second Phase III trial, ACTiVION-I, Kolon TissueGene expects in October.

🔗 Read original →

Insilico Medicine released 3D-Fit for testing language models in molecule design. On July 20, the Insilico Medicine team posted the 3D-Fit preprint. In the test, language and diffusion models create molecules for 1,453 protein-small molecule complexes. The drug molecule must occupy the protein's pocket - a cavity between its amino acids. A chemist can ask to preserve a fragment of a known molecule, attach a group for a hydrogen bond at a given point, or reach a specific amino acid. Each such contact gives the entire molecule additional conditions: the bonds and angles must remain valid, and the molecule itself must fit among the protein's atoms. In 3D-Fit, the authors give the model the coordinates of the pocket's atoms and a text condition: an anchor fragment, a pharmacophore point - a place for the necessary chemical function - or a mandatory contact with the protein. The model outputs the coordinates of the ligand, i.e., the candidate molecule. Then the authors separately measure two properties. The first is whether the model fulfilled the local spatial instruction. The second is whether the molecule was successfully restored, whether its geometry withstands physical checks, and whether its pose fits in the pocket. Language models often cope with anchor fragments and pharmacophore points. The authors associate this with the task format: the text condition is similar to a record that the model should generate. But the entire pose requires simultaneously agreeing on the shape of the molecule and its position among the protein's atoms. In this test, language models gave way to specialized diffusion models. In the initial poses, all the tested language models received a UniDock score above -6 kcal/mol; the authors use this threshold for weak binding. Local optimization shifted most results to approximately -6...-7 kcal/mol. The best diffusion models received lower scores, meaning more favorable poses in this metric, and more often passed the geometry checks in PoseBusters. Local contact is not enough to make a molecule a good candidate. 3D-Fit checks whether this contact is preserved when the entire molecule passes through the constraints of chemistry and the three-dimensional protein pocket. The physical check separates hitting a given point from a molecule that entirely fits in the pocket.

🔗 Read original →

TGN-073 reduced tau pathology in mice by activating the AQP4 water channel. In PS19 mice with tau pathology, the AQP4-activating compound TGN-073 enhanced the influx of MRI tracer from cerebrospinal fluid. The brain tissue showed less pathological tau, better preservation of neurons, and weaker gliosis, a reaction of support cells to damage. This effect was not observed in mice lacking AQP4. The article was published on July 18 in Molecular Neurodegeneration. PS19 is a line of mice with human mutant tau protein that aggregates into clumps inside neurons, damaging them and reproducing part of the tauopathy picture. The authors chronically administered TGN-073 to these mice, a compound that activates the AQP4 water channel. The brain has no separate pipe for cleaning up proteins: cerebrospinal fluid moves along vessels, exchanges with intercellular fluid, and carries away dissolved substances, a route called the glymphatic pathway. AQP4 is located in the processes of astrocytes around vessels and directs the flow of water through their membranes. Therefore, the authors measured not only the total amount of AQP4 in the brain but also how densely the channel is assembled around vessels. In PS19 mice, fluid exchange weakened even before the severe stage of the disease and worsened with age. After TGN-073, more MRI tracer entered the brain from cerebrospinal fluid. The tissue showed less pathological tau, neurons were better preserved, and gliosis was weaker. At the same time, more tau appeared in cerebrospinal fluid, and AQP4 again assembled more densely around vessels. The authors then repeated the experiment in PS19 mice lacking AQP4. TGN-073 no longer enhanced tracer influx and did not reduce tau pathology, gliosis, or neuronal loss. In this model, the entire observed effect of TGN-073 required AQP4. The connection between TGN-073 and fluid movement did not come out of nowhere: in a 2023 study, MRI also showed wider contrast distribution in healthy rats after this compound. The new article adds a model of tauopathy: here, one target is associated with both fluid movement and tau accumulation and neuronal preservation. For tauopathy, this represents two different directions of intervention: reducing tau production or changing fluid movement, which removes it from tissue. In chimeric mice, AAV vectors have already been used to deliver genetic cargo to human glial cells through cerebrospinal fluid movement. And VY1706 is being tried to reduce tau production in the brain with a single viral injection. In the PS19 model, TGN-073 worked through a different route: through AQP4 and fluid flow around vessels.

🔗 Read original →

ABILITY first recorded human brain signals with its system during an operation. On July 21, Swiss ABILITY Neurotech reported on the first operation in which its system recorded human neural signals. The Munich series involves people who have their brain tumors removed: during the procedure, electrodes will record cortical activity for 20-30 minutes. When a person with paralysis tries to say a word or move their hand, the intention may arise in the cortex, although the signal no longer reaches the muscles and voice. The neurointerface reads the electrical activity of the cortex, and the program learns to associate recurring patterns of activity with commands for a computer, speech synthesizer, or prosthesis. ABILITY's first procedure took place in the neurosurgical department of the Technical University of Munich's university clinic. The first participant was under general anesthesia. The team checked that the electrodes, equipment, and data recording worked together. In subsequent operations, the company plans to offer awake participants speech and motor tasks. Such tasks allow matching the pattern of brain activity with the attempt to utter a word or make a movement. "We can record known phenomena - evoked potentials and high-frequency gamma activity - and compare the quality of the ABILITY signal with clinical electrophysiology systems," says neurosurgeon Simon Jacob. Evoked potential is a brief response of the brain to a stimulus; high-frequency gamma activity is one type of its electrical activity. In the Munich series, the team will compare these signals with recordings from clinical systems. In parallel, in Utrecht, recruitment is underway for the INTRECOM study for people with severe paralysis and communication disorders. There, during the operation, four arrays of electrodes will be placed on the surface of the brain, and then researchers will train participants to use the system at home. The Munich series checks the recording during the operation, while INTRECOM builds a communication channel beyond it.

🔗 Read original →

The Xinhua Hospital commission linked the death of a 6-year-old participant to a gene editor introduced into the spinal fluid. On July 23, Science and Retraction Watch reported on the death of the girl after an early trial of therapy for the CHD3 gene variant. The hospital's emergency commission called the death "definitely related" to the treatment and cited thrombotic microangiopathy - damage to small vessels with microscopic thrombi - as the cause. In March 2025, doctors administered two AAV9 viral vectors into the spinal fluid of the 6-year-old participant, whom the investigation calls Mei. AAV is a viral shell capable of delivering genetic instructions to a cell. Seven days later, the girl died; the commission linked this outcome to the treatment, and the investigation describes a severe immune reaction. The treatment attempted to correct the R1025W substitution in the CHD3 gene, which is involved in regulating the function of other genes during brain development. To do this, Zilun Cui's team chose a base editor - a CRISPR variant that changes one "letter" of DNA without cutting both its strands. The entire editor did not fit into one AAV shell, so it was divided into two parts. Both parts should enter the same nerve cell, assemble a working protein, and modify the necessary DNA segment. An article in Nature, published on February 18, 2026, described this scheme in mice with the human CHD3 variant: after editing, the level of CHD3 protein in the animals increased, and behavioral indicators changed. In an experiment on two macaques, researchers saw parts of the editor in neurons and measured their assembly. To cover a sufficient number of brain cells, hundreds of trillions of viral particles are needed; the investigation links such a scale of delivery to an immune risk, and the commission named TMA as the cause of death. "Death should always be reported in a trial where treatment is first introduced to a person," said bioethicist Hank Greely. The trial card still indicates the recruitment status, with one participant planned, and the primary safety endpoint - treatment-related serious adverse events over 26 weeks. After the first injection, a severe outcome should become part of the decision on the next administration for the family and the clinical team. A personalized base editor for a baby with a CPS1 deficiency was delivered by lipid nanoparticles to the liver; after a year, the child had clinical improvement, and no serious side effects were registered. In the CHD3 trial, a large genetic construct required two viral vectors, and delivery to the brain required hundreds of trillions of viral particles. The risk of the next injection is composed of tissue, carrier, dose, immune protection, and reports of a severe outcome. Preclinical results become the basis for a clinical decision only together with these data. Sources: - investigation by Science and Retraction Watch - full text of the article in Nature

🔗 Read original →

A-Alpha Bio launched a consortium of five participants for joint antibody measurements on July 22. A-Alpha Bio announced the Atlas Consortium, which includes GSK, Boltz, Cradle, Dyno Therapeutics, and one other participant. They will jointly plan experiments for antibody-antigen binding and then obtain combined results quarterly. For a model to propose a new antibody, it's not enough to see the antibody sequence and its target. A response from the lab is needed: how strongly this pair binds. The strength of this binding is called affinity. The "antibody-antigen" pair, along with the measured affinity, becomes an example for the model to learn from. In the May protein model test, the computer search also ended with laboratory measurements: does the designed protein bind to the target? Such examples are difficult to compare when labs work under different protocols. One group tests one pair of molecules, another changes the conditions of the experiment, and a third uses a different method. The model receives many results, but it's harder to separate the properties of the molecule from the differences between experiments. Atlas participants will jointly determine which antibody-antigen pairs to measure. A-Alpha Bio will conduct these experiments on AlphaSeq under the same conditions, combine the results with their own experimental series, and release the array to all subscribers quarterly. Each new series will simultaneously serve as material for training and a common test for the participants' models. A-Alpha Bio estimates that with five founders, each participant will receive approximately 28 million affinity measurements per year. If the number of participants grows to twenty, the annual volume for each will increase to 100 million measurements, as stated in the Atlas Consortium announcement. Atlas already has material to work with. In a release on July 7, the company reported that it is transferring over 450 million affinity measurements and more than 7,000 lab-confirmed pseudostructures of antibody-antigen complexes to Atlas. The consortium will add to this set with regular series that participants will plan together. Each quarter, participants will agree on a new set of molecules, and AlphaSeq will measure them under one protocol. Then, all participants will compare the predictions of their models with the same results. Thus, laboratory measurements will become a common material for the next round of antibody design.

🔗 Read original →

On July 23, the US National Institute on Aging registered NECTAR, a phase II trial. The study plans to include up to 200 people over 65 years old: individuals with early Alzheimer's disease and cognitively healthy elderly. In NECTAR, all participants will perform daily tasks on a tablet for memory, attention, and other mental operations for four weeks. Some of them will additionally receive two doses of psilocybin at 25 mg with a two-week interval; after each dose, the participant will spend a day in the clinic. The rest will follow the same training program. Researchers will check if psilocybin changes the effect of daily practice. Neuroplasticity is the brain's ability to change connections between nerve cells under the influence of experience. Training provides such experience in the form of daily tasks. Since all participants follow the same program, the difference between the groups will show what the two doses of the drug add. This comparison tests the addition to training, not the difference between activities and the usual daily routine. In a review of 37 studies on digital cognitive training in mild cognitive impairment and dementia, the results of active engagement remained uncertain. In NECTAR, both groups receive training, so researchers will be able to separately assess the addition of psilocybin. The primary measurement after four weeks is a composite measure of neuroplasticity, which the registry calls NPCS. The team will also conduct cognitive tests, brain scans, sleep recordings, and safety assessments over six weeks. Enrollment is set to begin on October 1, 2026.

🔗 Read original →

Engineer James Orlando published exact states for two vertices of a quantum polytope on July 21. James Orlando posted a preprint, code, and data for the generalized Pauli constraints problem. His program gives explicit quantum states for two vertices that a 2008 paper had confirmed numerically: the calculation approached the answer closely, but there was no formula for the exact state at the time. Electrons belong to fermions, and the Pauli principle limits the number of particles in each quantum state. Generalized Pauli constraints add rules for the entire distribution of electrons across states. Mathematicians depict allowed distributions as points in a multidimensional figure; the vertices of this figure specify the limiting combinations allowed by quantum mechanics. In the work of Murata Altunbulak and Alexander Klüters from 2008, for a system of four fermions and nine orbitals, two such vertices remained a numerical result. Orlando published a preprint with explicit formulas and a repository where they can be verified. For one vertex, a combination of seven electronic configurations with integer weights is sufficient. For the other, quantum phases are required: amplitudes add up or cancel each other out, and it is this interference that creates the necessary distribution. What is interesting in this story is the way it works. Claude suggested moves, and the program listed vertices in rational arithmetic, built candidates, and checked them with solver certificates. Then the final state was verified again without rounding. In the author's account, Orlando writes that Claude "often and confidently made mistakes": the model got a digit wrong when reading a PDF. The acceptance criterion check detected an even more important failure: the search had silently limited itself to real amplitudes. After accounting for complex amplitudes, the state for the second vertex was found. The verification code and audit caught both errors. The model makes hypothesis searching cheaper; trust is created by a procedure capable of rejecting them. A week earlier, a project on Lean had checked a new lower bound proposed by GPT-5.6 Sol: the program's core re-derived each logical step. Here, the role of such a core is played by rational arithmetic and solver certificates. Open data allows another person to run the same chain and verify a specific formula, instead of evaluating the convincingness of the model's answer. The feedback from a profiled mathematician entered the same chain: after comments, Orlando clarified the formulations and re-checked the reasoning.

🔗 Read original →

Neuralink demonstrated how a brain implant controls an electric wheelchair. On July 23, Neuralink published a video featuring a participant in a clinical study: the implant translates imagined movement into a cursor on a screen, which in turn sets the direction and speed of the wheelchair. The PRIME study - the company's first human trial - is testing the N1 implant as a means of controlling external devices. Electric wheelchairs are typically controlled by a joystick, which is operated by hand, head, or another part of the body. In the Neuralink video, the participant gives a command through an implanted neurointerface. The device reads the brain's electrical activity, and a program converts the pattern of this activity into cursor movement. The cursor moves in an application that displays an image from a camera in front of the wheelchair. Upward movement means moving forward, downward means backward, and sideways means turning. The farther the cursor is from the center, the higher the speed. When the participant stops giving the command, the cursor slowly returns to the center, and the wheelchair stops. The on-screen pointer conveys the person's intention to the machine in the room. "At first, I was jerking around, but after a few minutes, the control became almost natural," the participant says in the video. According to him, the usual joystick required him to tilt his head and caused pain, while with the implant, he could sit up straight. The interface gives the person a way to choose where to go and when to stop. The PRIME study plans to enroll 15 adults with severe paralysis after spinal cord injury or ALS. The study is testing the safety of the N1 implant and the robot for its installation, as well as the device's operation with external devices. In this demonstration, the person sets the direction and stopping moment, and the wheelchair performs the movement. The signal from the brain goes all the way from intention to movement in physical space.

🔗 Read original →

The US Department of Energy (DOE) selected 278 projects for negotiations on awards under the Genesis Mission on July 22. These projects involve 342 institutions: national laboratories, universities, companies, and non-profit organizations. The Genesis Mission provides research groups with shared resources to work with AI. The DOE reported that 87 of the selected projects are led by national laboratories, 168 by universities, 19 by companies, and four by non-profit organizations. The selection launches negotiations on awards; after agreements, teams will gain access to the program's resources. The DOE reviewed applications in two steps: first, checking their relevance to one of the agency's 26 scientific priorities, and then - how AI can accelerate this work. More than 5,000 applications were submitted, according to Nextgov. Teams will receive tools for working with AI agents, models, software from industrial partners, and computing power from national laboratories. The task of this infrastructure is to reduce the cycle of scientific work: to formulate an approach, test it, and refine it based on the results. Projects from Princeton provide specific examples. One group will train a neural network to reproduce measurements in plasma physics that are too expensive or complex for laboratory experiments. Another will build a model of mobile DNA transfer between microbes to predict which bacteria will receive a genetic fragment. The portfolio also includes nuclear energy, extraction of critical minerals, chip design, and thermonuclear fusion. In March, the Genesis Mission connected 17 national laboratories, supercomputers, and scientific instruments, and Argonne demonstrated AI-assisted, autonomous laboratories, and robotic assistants in this system. Now, the DOE is distributing this infrastructure among 278 research tasks.

🔗 Read original →

At Texas A&M, pig kidneys were stored at -4 °C for three days and then transplanted back into the same animals. On July 23, MIT Technology Review wrote about the results of Matthew Powell Palm's team, presented in June at the American Transplant Congress. After 72 hours of storage, the kidneys were returned to the pigs; one of them was observed for 200 days. Typically, after removal, a kidney is kept on ice for 18-24 hours. During this time, doctors assess the organ, select a recipient, transport the kidney, and prepare for surgery. The organ continues to be damaged even in the cold, so long-distance transportation narrows the options. The Texas A&M team attempted to extend this window by cooling below zero. The kidney was flushed with a standard transplant solution and placed in a sealed chamber with constant pressure. At -4 °C, the pressure prevents water from forming ice: the temperature is already below zero, and crystals have not yet damaged the cells. A cryoprotectant, a substance like antifreeze, was not needed. The experiment tested the function of the transplanted organ. Some kidneys were kept on ice for two or 24 hours, while others were kept in the chamber for 24, 48, or 72 hours. Then, each was returned to the same pig, and the second kidney was removed: the preserved kidney was to cleanse the blood and produce urine on its own. After a day in the chamber, the kidneys immediately began to produce urine, and approximately ten days later, their indicators returned to normal levels. According to Powell Palm, organs after 48 and 72 hours recovered similarly. One pig was observed with such a kidney for 200 days. In the first month, the animal grew by about a third, and the kidney almost doubled in size, taking on the load of the removed paired organ. After removal, the team described it as healthy. In the pig experiment, the chain had already worked as a whole: the tissue was preserved without ice, warmed up, and a functioning kidney was obtained after transplantation. For transplant logistics, the extra two days provide time to compare several recipients, re-evaluate the organ, and choose a transportation route.

🔗 Read original →

The journal Science has published a map of aging of 40 human hippocampi: with age, microglia, astrocytes, and DNA layout change. On July 23, a peer-reviewed article by Nathan Sestan's team was published in Science. The authors studied post-mortem tissues of 40 people from 20 to 100 years old and simultaneously examined gene function, chemical DNA marks, and how chromosomes fold inside the cell nucleus. The hippocampus helps to remember events and navigate in space. Usually, age-related changes in the brain are discussed as a single general indicator, but the authors divided the material by cell type. They found that aging affects both the composition of cells and the way these cells manage genes. In samples of people 50-75 years old, the proportion of embryonically derived microglia decreased. Microglia are immune cells of the brain: they clean up cellular debris, respond to damage, and participate in tuning nerve connections. At the same time, cells with a profile similar to microglia from peripheral blood monocytes became more noticeable. There were also fewer astrocytes, which support signal transmission in synapses. This result is consistent with the SuperAgers map: in people over 80 years old with preserved memory, support programs were highlighted in astrocytes and CA1 neurons. The new work looks at the same cellular environment across the age scale and records a decrease in the proportion of astrocytes in the hippocampus. The most unusual observation concerns the three-dimensional layout of DNA. The chromosome in the nucleus is not stretched into a straight thread: distant DNA segments can be located nearby and together influence gene function. In many cell types, this spatial organization blurred with age. The authors' open version also describes the transition of microglia from a calm state to a state prepared for inflammation. In one human sample, the authors linked age-related changes in microglia and astrocytes with the rearrangement of gene regulation and DNA layout. The map shows the aging of the hippocampus as a set of specific cellular shifts, rather than as a single indicator for the entire brain.

🔗 Read original →

A 67% reduction in valine extended the lifespan of male mice by 23%. On July 24, a study on the lifelong diet of C57BL/6J mice was published in Nature Aging. Researchers reduced only valine in the diet, preserving calorie, fat, and carbohydrate content; median lifespan increased in males, and several health indicators improved in both sexes. Valine is an essential amino acid obtained from food, part of the BCAA group along with leucine and isoleucine. The Nature Aging article reported that the C57BL/6J mouse diet, starting at four weeks of age, contained 67% less valine; the deficient amount was replaced with amino acids that the body can synthesize itself. This single shift changed several indicators: mice of both sexes gained less fat, better controlled glucose, and received lower frailty index scores - a set of tests for age-related disorders. Median lifespan increased by 23% in males. In females, health indicators improved, but lifespan did not change. The authors measured different tissue responses: mice with valine restriction consumed more calories per unit of body mass and expended more energy at 18 months. In males, mitochondrial respiration intensified in the liver: mitochondria more actively converted food energy into molecules that cells use to fuel their work. In females, the proportion of neuroinflammatory glial cells - cells that support neuron function - decreased more strongly. A 2021 experiment with lifelong restriction of all three BCAAs extended the lifespan of male mice by approximately 30%. The new study separates valine from leucine and isoleucine: a single change in the diet was enough to improve health indicators in both sexes and extend the lifespan of males. This allows researchers to compare the response of individual tissues to each amino acid, rather than reducing all proteins at once.

🔗 Read original →

AcroCyte announced funding for ARPA-H kidney organoid growth technology from patient cells on July 23. AcroCyte Therapeutics reported that ARPA-H is supporting its R3CE technology. The company grows small three-dimensional tissue fragments - kidney organoid modules - from a patient's kidney cells. The company describes R3CE as a method for multiplying rare human cells in three-dimensional culture. In a July 23 release, AcroCyte writes that kidney cells are obtained through a minimally invasive procedure, expanded in culture over weeks, and grown into organoid modules. The company is attempting to make the transition from a small cell sample to organoid material repeatable. The organoid has several engineering challenges. Two weeks ago, a USC team showed how the WNT signal gives nephrons in a kidney organoid direction for growth. AcroCyte's application relates to a different stage: obtaining the cellular material from which tissue modules are then grown. AcroCyte also named project partners: University of Chicago Medicine, National Taiwan University Hospital, and the National Center for Biomodels. According to the release, they will work in the US and Taiwan. ARPA-H support is focused on R3CE - a technology that should provide comparable batches of cells for growing organoids. ARPA-H funds risky medical technologies and solutions that can be scaled up. For AcroCyte, scaling begins with reproducible growth of rare cells: it depends on whether the same growth method can provide comparable tissue modules.

🔗 Read original →

Scribe has set the price for its IPO: 8.58 million shares at $15 for the development of CRISPR programs. On July 23, Scribe set the price for its increased IPO: 8.58 million shares at $15. The company estimates gross proceeds, before expenses, to be $128.7 million; trading of SCTX on Nasdaq and the deal closure are scheduled for July 24 and 27. Scribe's main program is already undergoing its first human trial. Scribe Therapeutics has set the price for its increased IPO: 8.58 million common shares at $15. The company expects gross proceeds of $128.7 million, before expenses related to the offering; the deal closure is scheduled for July 27, subject to customary conditions. In Scribe's preliminary filing on July 2, the price and volume had not yet been specified. In its S-1 filing, the company indicated that it plans to use the net proceeds, together with its existing funds, for STX-1150, two other programs, and ELXR/XE technologies. STX-1150 is being developed for the long-term reduction of LDL-C, "bad" cholesterol. The liver produces the PCSK9 protein; it removes receptors that capture LDL-C from the blood from the cell surface. If PCSK9 is suppressed, more receptors remain, and they remove more LDL-C. STX-1150 is delivered to the liver in lipid nanoparticles, i.e., fatty shells for the drug payload. The ELXR system places chemical marks on the PCSK9 gene and reduces its activity without changing the DNA sequence. Thus, Scribe is trying to suppress PCSK9 for a long time by regulating gene function. The company's prospectus describes the first trial in Australia: it may include up to 64 adults with elevated LDL-C and a risk of atherosclerotic cardiovascular disease. In the first half of 2027, Scribe expects to receive data on safety, tolerability, and LDL-C reduction. The prototype in a non-human primate experiment showed that a single dose maintained a reduction in LDL-C of more than 50% for two years. BioPharma Dive calls Scribe the first gene-editing drug developer to go public in more than two years. The set price ties the development of the clinical STX-1150 program and early CRISPR programs to the public offering, which the company plans to close on July 27.

🔗 Read original →

On July 13, the Sakana AI, IT University of Copenhagen, and Autodesk Research team published an experiment in Nature Communications with four assemblies of 26–197 identical modules. A plane, guitar, boat, and round table in all three repeats came to the correct answer about their shape in 23 cycles, approximately 70 seconds. Each module is a printed circuit board in a cube enclosure. A microcontroller runs the same small neural network, an LED displays its current answer, and connectors on six faces transmit messages to neighbors. The cube has no coordinates and no overall map of the assembly: it knows only its internal state and what the attached cubes send it. The authors trained a neural cellular automaton — a network where each cube applies the same update rule. The cube changes its internal state, receives new messages from neighbors, and counts again. Signals spread through the assembly until different parts come to one label: plane, guitar, boat, or table. The network first learned in simulation on seven classes of objects, and then the same rule was transferred to physical modules. In its hidden channels, the authors saw stable spatial gradients: the state of the cube changed depending on its place in the assembly. Such a signal gives the cube a relative landmark, although no one reports coordinates to it. In simulation, the authors assigned the network another task: to simultaneously name the shape and indicate one of the six directions to a missing cube. The network found the direction with an average accuracy of 94.8%, while maintaining 98.9% accuracy of shape recognition. Then the program added a virtual cube in the indicated direction and repeated the calculation. An assembly that needs to find a place of failure must first agree on what object it forms. The team demonstrated this computational step on almost two hundred cubes connected only with nearest neighbors.

🔗 Read original →