AI Longevity Daily
Daily developments in AI, longevity and aging research.
Today’s Summary
Mouse study finds hypothalamic neuron cilia shorten with age; knocking out cilia growth genes partially disables caloric restriction benefits.
UW and Nvidia open-source protein design model RFD4-Proteina — lab tests show 5–10x higher design success rates over baseline.
The cilia-caloric restriction link was only demonstrated in mice; whether human hypothalamic cilia play the same role remains unknown.🔥 Today’s Top 7 🔬
1. Hypothalamic Neuron Cilia Length Affects Whole-Body Metabolism and Caloric Restriction Efficacy
According to a report by Lifespan.io , researchers found that cilia on hypothalamic neurons — those antenna-like structures cells extend from their surface — shorten with age, with declining NAD+ levels as the direct culprit. In vitro experiments showed that supplementing NAD+ precursors or inhibiting CD38, the enzyme that depletes it, could restore cilia length. The CHMP4B protein inside cilia normally cycles back to lysosomes to keep them intact, but that process breaks down when NAD+ runs low. More critically, when researchers knocked out cilia growth genes using silencing RNA, caloric restriction in 24-month-old mice stopped improving certain metabolic markers — pointing to these specific neuronal cilia as a necessary piece of the caloric restriction puzzle. The work was published in a peer-reviewed journal, but it’s all in mice so far, and whether human hypothalamic cilia play the same role is still an open question.
2. Nvidia and UW Open-Source RFD4-Proteina Protein Design Model 🧬
RFD4-Proteina, an open-source generative model for protein design, was just unveiled by the UW Institute for Protein Design — led by Nobel laureate David Baker — in collaboration with Nvidia, according to GEN . Built on Nvidia’s BioNeMo platform, the model can simultaneously optimize protein sequences, structures, and their interactions with other molecules. Lab tests across four task categories — protein binding, small molecule binding, enzyme scaffolding, and DNA binding — showed RFD4-Proteina hitting 5–10x higher design success rates than baseline methods, and generating results 5–10x faster. The model is open-sourced under a BSD-3 license and accessible via Amazon Bio Discovery, Hugging Face, and Google Cloud. Baker says it’ll help design proteins for drug discovery and enzymes that break down pesticides, though the road from computational design to lab validation is still a long one.
3. Cryo-EM and AI-Designed Scaffold Reveal How Andes Hantavirus Hijacks Human Cells 🦠
Researchers used cryo-EM to resolve the near-atomic-resolution structure of how the Andes virus spike protein binds to the human receptor protocadherin-1 (PCDH1), according to GEN . Andes virus is the only hantavirus known to spread person-to-person, and it can cause a deadly pulmonary syndrome. The team found that four Gn glycoprotein subunits on the virus surface form a deep pocket, into which PCDH1 inserts itself in a slender configuration — a geometry that most antibodies are simply too bulky to access. An AI-designed scaffold was used to stabilize this tetrameric complex. Mutagenesis experiments confirmed three Gn loops are involved in binding, and researchers even managed to transfer this binding capability to an Old World hantavirus. Published in Nature, the authors suggest designing slim molecules to block that deep pocket as a potential antiviral strategy — though no drug candidates exist yet.
4. Psilocybin Reshapes Brain Connections Through Distinct Cellular Pathways 🍄
New research from UC Santa Cruz digs into how psilocybin — the active compound in magic mushrooms and a serotonin receptor agonist — reshapes neural connections, according to News-Medical . The study identifies the specific cellular pathways through which psilocybin acts, which has been a key missing piece given the growing interest in using it for depression and PTSD treatment. That said, the media summary doesn’t spell out the experimental model, sample size, or the exact mechanisms involved. What we know is the research focuses on neuroplasticity — the brain’s ability to form new connections — which is widely thought to underlie psilocybin’s potential therapeutic effects. For the full picture, you’ll need to check the original paper; it’s not clear yet whether this was done at the cellular, animal, or human level, or how translatable the findings are to clinical use.
5. KARI Dataset Pools 6,000 Brain Scans to Study Alzheimer’s 🧠
The KARI multimodal imaging dataset brings together roughly 6,000 brain scans to explore the relationships between amyloid, tau protein, brain structure, and cognitive performance in Alzheimer’s disease, according to News-Medical . Multimodal imaging means combining different techniques — think PET (positron emission tomography) and MRI (magnetic resonance imaging) — so you can look at both pathological protein buildup and brain tissue atrophy at the same time. The media summary doesn’t detail the data sources, participant characteristics, or specific findings; it just says the dataset offers “critical insights.” You’d need to check the original paper for the full research design and results. It’s also not yet clear whether this is a retrospective analysis or a prospective cohort, or whether the imaging can actually improve early diagnostic accuracy.
6. The Postpartum Recovery Window May Shape Long-Term Maternal Aging 👶
A review paper by Dr. Jesse Poganik from the Genetics Division at Brigham and Women’s Hospital/MGH, published in Trends in Endocrinology & Metabolism, explores the postpartum recovery window as a potential determinant of long-term maternal aging, according to News-Medical . The postpartum recovery window — typically the weeks to months after birth when the body is repairing itself — involves metabolic, hormonal, and immune shifts that could have lasting effects on long-term health. The media summary only lists the author info without any concrete study data, cohort sizes, or biological age measurements. This is an opinion-style review, not an experimental report, so the specific evidence and mechanistic hypotheses are in the original paper. Don’t take this as proof that postpartum interventions can slow aging — the jury’s still out.
7. Short Gaps Between Pregnancies May Speed Up Biological Aging in Mothers ⏱️
A new study digs into the link between interpregnancy intervals and biological aging in mothers, according to News-Medical . Biological aging — distinct from chronological age — refers to the changes that accumulate in cells, tissues, and organs over time, measurable through markers like DNA methylation and telomere length, and associated with chronic disease risk. The media summary is pretty thin: no study design, sample size, specific interval definition (like under 12 or 18 months), or magnitude of the aging effect. All we know is the research touches on how reproductive history affects long-term health, which sits within the broader field of reproductive aging in longevity science. Full evidence is in the original paper, and this isn’t enough to draw any conclusions about pregnancy planning.
🧪 Cutting-Edge Research
Trion-1: A Diffusion Language Model Unifying Language, Protein Sequences, and Structure
This arXiv preprint introduces Trion-1, the first natively unified multimodal protein diffusion language model (dLLM). Trion-1 represents sequences, structures, and natural language in a single unified discrete token space, swapping out strict left-to-right generation for bidirectional or semi-bidirectional interaction — allowing global residues to communicate and model long-range contacts. This design unifies six core capabilities in one shot: folding, inverse folding, sequence understanding, structure understanding, sequence design, and structure design, while keeping the instruction-following ability of a general-purpose LLM. On benchmarks, Trion-1 hits a TM-score above 0.8 with a single sample (where TM-score measures predicted-vs-actual structural similarity, with >0.5 indicating the same fold) — comparable to existing three-modality models using 200 samples. On sequence understanding tasks it reaches 62.37%, putting it close to Claude Opus 4.7. It’s a preprint, so no peer review yet; code and model weights are available via the paper link.
Visual Grounding Benchmark Evaluates Multimodal LLMs on Retinal OCT 👁️
This arXiv preprint introduces RetinaGroundBench, the first benchmark for visual grounding of biomarkers in retinal optical coherence tomography (OCT). OCT is the go-to imaging technique in ophthalmology, capable of showing the layered structure of the retina. The benchmark includes 815 B-scan images annotated by ophthalmologists with 4,537 bounding boxes covering 13 biomarkers across 8 diseases. The team evaluated 27 open-source multimodal LLMs (0.8B–38B parameters) and found visual grounding is nowhere near solved: the best general model hit a macro F1 of just 23.7%, with consistently low recall — even though accuracy on biomarker presence/absence questions reached 73%. The authors’ own RetGrounder (8B parameters), fine-tuned with OCT-specific strategies, pushed F1 up to 33.9%, a 43% relative improvement. Still a preprint, and not yet validated in a clinical setting; code and model are available via the paper.
⌘ Top Open-Source Projects
ADAMMA-CDHI-ETH-Zurich/CosinorAge: Circadian Rhythm-Based Biological Age Prediction Package
CosinorAge is a Python package implementing cosinor-based biological age prediction. Circadian rhythms — the body’s ~24-hour physiological cycles — shift in amplitude and phase as we age, making them a viable aging biomarker. This tool is a solid fit for researchers who need to estimate biological age from time-series physiological data like body temperature or activity levels. The repo includes code for model training and prediction, but the README doesn’t spell out input data format, pretrained models, or validation datasets — so check the docs or example scripts before diving in. The project has 13 stars, was last active on October 23, 2026, and the license isn’t shown on the GitHub page, so worth double-checking.
JesperBoman/Butterfly_DNAm_age: Scripts for Training Epigenetic Clocks 🦋
Butterfly_DNAm_age is a collection of scripts for training epigenetic clocks using DNA methylation (DNAm) patterns to predict biological age. While most epigenetic clocks target humans or mammals, this project focuses on butterflies — making it a useful reference for anyone studying aging in non-model organisms or doing cross-species comparisons of aging mechanisms. The repo covers data preprocessing and model training workflows. The README is light on details around dependencies, input methylation data sources, and clock prediction accuracy, so you’ll want to lean on the code comments. The project has 1 star, was last active on October 23, 2026, and license and release status still need to be verified.