OmniTCR: Single 113‑M‑parameter model predicts T‑cell antigen recognition and designs candidate receptors

A preprint released on 13 September on bioRxiv describes the OmniTCR model, which contains 113 million parameters. It was trained on 328 million immune sequence records, ranging from single receptors to rare full receptor‑peptide‑HLA complexes.

The model learns to predict the next token in a sequence where peptide, HLA, and the α‑ and β‑chains of the T‑cell receptor are separated by boundary tokens. By presenting the components in different orders — peptide‑HLA followed by receptor chains, or the reverse — OmniTCR can both generate candidate receptors for a given target and evaluate how well a known receptor matches that target. This setup lets it leverage both scarce full complexes and abundant single‑chain data for a unified task.

To test generalization, the authors held out peptides never seen during training. On 642 peptide‑β‑chain pairs OmniTCR achieved an AUPRC of 0.7009, and on 141 full peptide‑HLA‑receptor complexes an AUPRC of 0.8235, outperforming the strongest comparison models by 0.3396 and 0.3451, respectively.

For receptor design, OmniTCR constructs the CDR3β region of the β‑chain. Candidates are scored against a given peptide‑HLA pair and ranked by the difference to a null input score. In the top‑100 lists for 20 targets, the average fraction of shared sequences fell from 0.0870 to 0.0323, showing that the target reshapes the candidate pool.

For seven complexes with peptides outside the training set, the authors used AlphaFold 3 to model the candidate structures and compared their structural confidence. Overall, the model learns inter‑component relationships from the full data corpus, transfers to novel peptides, and produces a ranked list of candidates ready for experimental validation.

bioRxiv, 13 September

🔗 Read original →

Cellectis halts two CAR-T programs, shifts resources

🔗 Read original →

ScMLEAge cell‑age estimator leaks test data in published code

ScMLEAge is a method that estimates a cell’s age by comparing its RNA‑molecule counts to age‑specific profiles built from donor mice. According to an article published npj Aging on 12 September, the program assigns each cell to an age group based on the number of RNA molecules read from its genes. They used 10 males from 9 tissues from the Tabula Muris Senis dataset.

The method creates age profiles by summing RNA counts of cells from the same age group for each cell type, then

🔗 Read original →

Atlas of short proteins in human frontal cortex reveals abundant micro-MKKS63

The study published in Nature Aging, September 14 analyzed 610 postmortem frontal cortex samples and identified 4,321 short proteins using mass spectrometry and ribosomal profiling. Of these, 3,217 were absent from the manually reviewed UniProt section, while 1,067 showed strong spectral matches to predicted sequences.

The gene MKKS, annotated with a 570‑aa protein, most frequently yielded a shorter isoform, micro-MKKS63 (63 aa), which received over 1,000 spectral matches. Ribosomal profiling confirmed that cells translate the RNA region encoding micro-MKKS63, and its levels were reduced in the frontal cortex of symptomatic Alzheimer’s disease cases.

In the human microglial line HMC3, micro-MKKS63 localized to mitochondria. CRISPR‑mediated disruption of the micro‑MKKS63‑encoding region in two independent clones decreased basal and maximal oxygen consumption and energy‑linked respiration.

These results show that detecting short proteins separately is essential in brain tissue, as they can be the primary detectable product of a gene and directly influence cellular function.

🔗 Read original →

APOE ε4 carriers reveal distinct DNA associations with Alzheimer's disease

On September 14, researchers published a study in Nature Genetics that stratified Alzheimer’s patients

🔗 Read original →

Using organ perfusion time to deliver protective drugs before transplant

On 14 September, a scientific comment proposed using the time when a donor organ is connected to a perfusion machine to deliver a drug intervention. The idea stems from a work in Cell about the experimental compound FXT-001, which protects cell membranes from oxidative damage.

Machine perfusion pumps oxygenated fluid or blood through the extracted organ, keeping it outside the body and allowing its function to be assessed before transplantation. The comment suggests administering the drug during perfusion and measuring its effect by functional markers.

Damage begins when the organ is deprived of oxygen and nutrients, and worsens after blood flow is restored because iron accelerates lipid peroxidation of cell membranes. In 116 liver transplant recipients, malondialdehyde, a marker of lipid oxidation, peaked 30 minutes after reperfusion and its level correlated with AST, a laboratory indicator of liver injury.

FXT-001 chelates iron and blocks the chain‑radical reactions in membrane lipids, so it was tested precisely on this damage mechanism. In pig livers, adding FXT-001 to the cold‑storage solution lowered AST release during subsequent perfusion; in pig lungs, adding the drug to the perfusion fluid at the start reduced water accumulation.

In five pairs of human lungs deemed unsuitable for transplant, the organs were perfused simultaneously: one lung received FXT-001, the other a control solution. This paired design reduces donor variability, and the lungs treated with FXT-001 showed less excess water and weight gain and better preservation of inflation capacity.

The comment concludes that delivering the drug locally during perfusion and assessing organ function before transplant could improve outcomes. This approach would turn the perfusion period into a therapeutic window.

🔗 Read original →

ChASE37 enzyme aids human neural cell survival in rat stroke model

On September 9, researchers reported results from an ischemic stroke model in rats. Human neural progenitors — cells that can develop into neurons — survived for 4 weeks in 5 of 8 animals that received the cells together with the enzyme ChASE

🔗 Read original →

Norn Group Reopens Talent Bridge for International Aging Researchers

On September 12, Norn Group opened enrollment for its updated Talent Bridge program, which is aimed at individuals living outside the United States who want to work on aging research. The initiative connects candidates with mentors, the Longevity Nexus community, and potential host organizations in the US.

After applying, a candidate proposes a project; if selected, they complete the work in individualized stages totaling about 100 hours over a period of 3 weeks to 4 months. Upon finishing the collaborative project, the participant receives a stipend of roughly $10,000.

Norn first launched Talent Bridge in 2022. The program includes the Longevity Nexus community for peer exchange and mentorship, and offers help with legal, logistical, and visa matters—including guidance on the O‑1 visa for individuals with extraordinary ability.

Early participants illustrate the program’s impact: Marton Meszaros investigated the missing human data and validation methods needed for aging biomarkers and is now launching a company in the United States. Nicholas Di Leo prepared a review of rapamycin clinical trials. As Norn states, “We look for potential and focus, not formal qualifications.”

Talent Bridge structures the entire pathway around a completed, publishable project. This gives candidates concrete results to discuss with mentors and prospective host organizations. The program then provides networking and relocation support to facilitate the move to the US.

🔗 Read original →

Matt Kaeberlein Questions Genflow's Early Claim of Age Reversal in Dogs

On 8 September Genflow announced that it had met its pre‑specified primary endpoint in a trial of 24 beagles older than ten years. The company said the GRIM DNA‑methylation clock showed a lower biological age in the treated dogs compared with the saline‑treated group, and it pledged to present detailed results on 1–2 October.

On 12 September biogerontologist Matt Kaeberlein questioned why the outcome was released before the full data were available. He noted that, as far as he could see, no preprint or dataset existed to evaluate the claim of age reversal.

The study randomly assigned the 24 beagles to four groups, with assessors blinded to treatment. Three groups received different gene‑therapy variants, while the control group received saline. The primary endpoint was the GRIM clock, and Genflow plans to disclose the full dataset at the Animal Longevity Summit in Toronto on 1–2 October, with DNA‑clock and muscle‑histology results slated for 2 October.

Kaeberlein had previously discussed in August whether a shift in such a clock could be taken as evidence of an intervention’s effect, and he asked for the data needed to verify the claim of a 'reduction of biological age'. He highlighted the Longevity.Technology headline: 'Genflow gene therapy reverses aging in dogs'.

He argued that only after the 2 October release of the DNA‑clock and muscle‑histology measurements can the primary‑endpoint claim be properly checked against the underlying data.

🔗 Read original →

Sinclair and Brenner Debate Why SIRT1‑Targeting Drug Programs Were Halted

On September 10 a study on the compound SRT1720 was released; it was tested on rat kidney cells and on rats with sepsis. On September 1

🔗 Read original →

Roko Mijić Warns AI Growth Could Outpace Human Governance

Roko Mijić posted on X on September 12 that the rapid growth of AI capabilities could bypass human control, arguing that legal and political institutions might fail to steer AI’s expanding abilities in medicine, pensions, goods, and safety toward human benefit.

He sees governance as the condition that turns new capabilities into help for people; new ways of working and finding efficient solutions benefit humans only through rules and institutions that distribute their outputs, which he describes as a fragile link.

In his example, if AI’s ability to find efficient solutions grows about 50% per year, institutions could still treat it as fast economic growth, but at 1,000–10,000% per year AI would have enough optimization power to circumvent human governance, potentially leading to scenarios where AI overthrows the state instead of paying taxes or destroys people instead of treating aging.

He links the need for a slowdown in AI development — referencing an April conversation with David Wood about a contractual pause in the race for general‑purpose AI — to the task of reshaping governance to handle far more disruptive changes, arguing that surviving AI will require both slowing AI and rebuilding institutions.

🔗 Read original →

Single amino‑acid swap in CD27 boosts CAR‑T activity against CD70

In a preprint published on September 10, authors changed one amino acid in CD27 – the natural partner of the tumor antigen CD70.

CAR‑T cells are made by giving T lymphocytes a receptor that recognizes a chosen target; here the receptor’s recognition domain is CD27 itself rather than an antibody fragment. A 2025 study (2025 work) showed that a CD27‑CAR construct expanded in mouse blood roughly 80–100‑fold more than antibody‑based CARs in a myeloma model.

To improve CD27, the researchers modeled the CD27:CD70 interface and screened 24 contact residues; the N88A substitution (asparagine → alanine at position 88) emerged as the best candidate from both computational and experimental tests.

In vitro, N88A‑CAR‑T cells killed two AML cell lines faster than the parental CD27‑CAR. In a leukemia mouse model, animals received tumor cells and then 0.5 million CAR‑T cells after four days; tumors stayed below detection limits for the full 85‑day observation in the N88A group, whereas the original CD27‑CAR group had a median survival of 38 days.

In a myeloma model both CD27 versions kept tumors undetectable, but N88A showed greater expansion in mouse blood. Biochemical assays revealed N88A binds CD70 with a slower association rate but a >4‑fold slower dissociation, giving an approximately 2‑fold stronger CD27:CD70 complex; modeling suggests the complex can adopt multiple bound states, consistent with the measured slow off‑rate.

Off‑target screening detected signal only at FCGR3A due to the

🔗 Read original →

PHAROS preprint introduces algorithm to find drug sequences for desired cell states

On September 10 the PHAROS preprint appeared, describing a system that takes the current and desired states of a cell group and searches for an order of drug perturbations that, according to its calculations, brings the first state closer to the second. The approach relies on the STATE model, which is trained to predict how a cell group changes after a drug is applied. Unlike a standard cellular‑response model that predicts the effect of a given drug, PHAROS begins with the initial and target cell states.

The authors frame therapy discovery as a reverse question: “Which interventions are most likely to shift a heterogeneous cell population from its present state to a desired one?” PHAROS applies drugs one at a time; after each step

🔗 Read original →

New perspective proposes assay for tissue repair and tumor risk after cancer therapy

On September 11 Adriana Paulina Gudino Reyes published a perspective on restoring normal tissue after cancer treatment. She proposes checking the recovery of damaged tissue together with signs that the intervention might support tumor growth. Oncology metrics track tumor burden, progression and survival; Reyes adds a separate outcome: can damaged normal tissue regain its function. Tissue repair and tumor adaptation may use the same cellular signals. In aged mice, activating the Igf2 gene restored the liver’s regenerative capacity, while blocking the IGF1R protein‑receptor reduced tumor burden in a hepatoblastoma model. In healthy growing mice the same blocker decreased liver size and body mass. One growth pathway participated in both tissue repair and tumor growth under different conditions. Tissue repair engages inflammation, angiogenesis and extracellular‑matrix remodeling; in tumors the same processes occur among malignant, immune and stromal cells. “The same signaling network can support normal‑tissue repair while simultaneously helping the tumor survive, invade or evade immunity.” Reyes recommends moving from simple models to complex ones: first compare normal and tumor human cells before damage, after damage and during recovery; then add stromal and immune cells. Organoids and co‑cultures allow testing their interactions, as do microphysiological systems that replicate tissue functions. In each model she suggests pairing normal‑tissue repair readouts with tumor‑cell proliferation, invasion, angiogenic signaling, genetic integrity and therapy sensitivity. As an example she analyzes stromal cells from Wharton’s jelly – umbilical‑cord tissue – whose properties depend on source, purity, culture conditions, passage number, genetic stability, sterility and traceability from donor to sample. These data let one assess how a specific cell product behaves alongside a particular tumor.

🔗 Read original →

New Neurointerface Restores Speech and Movement via 3D Avatar

This neurointerface helps people with severe paralysis fully restore both nonverbal and verbal communication. While earlier systems could decode either speech or movement in isolation, this development is the first to combine both functions, translating brain signals in real time through a full‑size digital 3D avatar that speaks and reproduces facial expressions and gestures.

It relies on a single high-density ECoG implant that reads the electrical activity of the cerebral cortex. Researchers found that the brain zones responsible for speech and limb movement partly overlap in activity. To prevent interference, engineers built a specialized AI model with parallel decoders that separate and simultaneously process commands for articulation and gesturing (such as a shoulder shrug, head nod, or fist raise).

The neurointerface was successfully tested on three patients who had completely lost the ability to communicate. In dialogue tests the system showed high accuracy of intent recognition, and one participant reached 100% average decoding accuracy for both speech and gestures. This perfect score was maintained across three conversational blocks.

🔗 Read original →

Space Exposure Tests Reveal Cataract Lens Durability on ISS

Scientists sent intraocular lenses used in cataract surgery to the International Space Station to evaluate how artificial lens materials endure the harsh environment of open space. This knowledge is vital for preparing long‑duration missions to the Moon and Mars, where transplantation and on‑site surgical care may become essential because of radiation, injury, and crew aging.

135 lenses were removed from sterile packaging and mounted in special containers on the station’s exterior, where they faced direct vacuum, temperature swings, solar ultraviolet radiation, and a stream of atomic oxygen for six months at three different shielding levels. In parallel, 45 lenses from the same batch were kept in a control container on Earth at room temperature and normal

🔗 Read original →

ARPA‑H awards AI contracts for heart‑failure patient support between visits

On September 9, the US agency ARPA‑H announced contract awards for its 39‑month ADVOCATE program aimed at developing AI agents to assist heart‑failure patients between clinic visits. In the first stage, teams have 24 months to build an AI agent and submit a FDA device authorization request.

Heart‑failure management requires clinicians to iteratively adjust therapy based on symptoms, blood pressure, kidney function, and drug tolerance. ADVOCATE seeks to shift some of this repetitive work outside the visit by having the agent interact with patients, support them, and escalate cases to the clinical team.

The first‑stage contracts went to Atman Health, Tempus AI, UpDoc. The program also funds an AI observer that will monitor the agent in real time for unsafe recommendations, and an independent evaluation by the Johns Hopkins University Applied Physics Lab to assess technical performance and clinical outcomes.

Because algorithms can learn site‑specific habits, a review of medical‑AI auditing notes that local rules and patient mixes leave traces in records, causing a model to perform well in one hospital but poorly elsewhere. UpDoc’s roadmap includes independent validation, comparing the agent’s decisions with those of cardiologists, then an FDA‑cleared investigational‑device clinical study.

Atman uses a language model that reads medical records and converses with patients, while a separate rule‑based system makes treatment decisions so clinicians can see the logic behind each recommendation. The program’s trajectory — from contracts, through observer oversight, independent assessment, and usual‑care comparison — leads to clinical verification.

🔗 Read original →

Mitochondria in nerve terminals linked to varied rule‑switching in aged mice

Researchers linked poorer rule‑switching in aged mice to mitochondria in nerve terminals. The study appeared September 14 in Aging Cell. They examined the medial prefrontal cortex and tested the mitochondrial antioxidant MitoQ, which accumulates in mitochondria.

Mice first learned to obtain a reward for one of two screen lines. Then the rule changed so that reward depended on touching one side of the screen regardless of line, measuring the ability to abandon a learned response and adopt a new one.

Male C57BL/6J mice mastered the initial task equally, but after the rule switch accuracy in the final session varied widely among old animals. Investigators sought a marker of behavioral flexibility after the initial task was learned.

Three‑dimensional electron microscopy showed synaptic density declines with age. In old mice task accuracy varied independently of synaptic density and size, while a higher proportion of synapses containing a mitochondrion in the nerve terminal correlated with lower accuracy.

Proteomic analysis of a synapse‑enriched fraction found poorer accuracy matched higher levels of mitochondrial proteins, including those involved in energy production. Comparing age‑related proteins with accuracy‑correlated proteins revealed little overlap, indicating distinct protein sets for aging versus individual differences.

From week 55 to week 75 mice received MitoQ in water or the control compound dTPP. The final analysis included seven MitoQ‑treated and five dTPP‑treated mice. MitoQ improved performance on the rule‑switching task while initial learning remained equal. After MitoQ the synaptosomal fraction showed reduced mitochondrial and pro‑apoptotic proteins linked to programmed cell death.

The authors propose mitochondria in nerve terminals as a testable component explaining why same‑aged animals differ in behavioral adaptation after condition changes.

🔗 Read original →

GenEHR predicts 5‑year cancer risk from electronic health records

Authors evaluated GenEHR on five existing EHR databases. For each database they trained a separate model version that estimates the probability of a first cancer diagnosis within 6–60 months and ranks individuals for targeted screening. The preprint was posted on 11 September.

The model incorporates the order of diagnoses, medications, procedures, lab results, and intervals between visits. A repeat analysis a week after an abnormal result may signal active diagnostic work‑up, whereas the same analysis a year later often reflects routine annual check‑up. First

🔗 Read original →

GenScript to invest 70% of IPO proceeds in AI‑driven automated labs

On September 11, GenScript signed a conditional agreement to place 77.126 million new shares at HK$30.50 each. If fully placed, the company expects to receive about HK$2.327 billion after expenses.

GenScript plans to direct roughly 70% of these net proceeds toward automated laboratories for AI‑based drug discovery. About 20% will go to related research, workflow‑linking software, and international platform expansion.

The AI system proposes candidate molecules that could become future drugs; the lab then synthesizes them and tests whether they bind to a chosen biological target, such as a disease‑relevant protein, and measures their properties. These results feed the next round of computer‑driven design.

Automated high‑throughput labs run such experiments in large batches, and the exchange announcement notes that funds will also cover equipment, upgrades to existing facilities, and expanded synthesis, screening, and testing capabilities. The board intends to use the raised capital to expand this infrastructure, while current operations are funded from

🔗 Read original →

Five pharma firms improve protein‑ligand AI model via federated learning

Five pharmaceutical companies jointly fine‑tuned OpenFold3 Preview 2 on a private set of 20,167 confidential structures and evaluated on a held‑out set of 1,056 structures. The work was reported by Nature on 14 September.

For a drug to work its molecule must bind a specific site on the target protein; OpenFold3 builds a 3D model of the protein‑ligand pair from sequence and molecular description. Useful structures often stay inside pharma pipelines because they involve undisclosed molecules and targets.

OpenFold released the weights, code and data for OpenFold3 in March, allowing anyone to reproduce and fine‑tune the model. Each company kept its own structures locally and fine‑tuned a copy of the model; updates were aggregated into a shared version using federated learning.

On the held‑out set the fine‑tuned version reached the target accuracy in 52.1% of cases, versus 35.6% for the original OpenFold3 Preview 2 and 40.9% for Boltz‑

🔗 Read original →