AI Longevity Daily

Daily Digest

Penn's Percival, trained on large-scale CT, predicts cross-organ disease better than image baselines but wasn't compared to radiologists; some predictions may reflect existing undocumented disease.
Multiple preprints integrate histopathology slides, spatial transcriptomics, and brain imaging into multimodal models; Brain2Vec has open-sourced weights and is accessible via cloud platform.
First, verify the ATH434 phase-2 post-hoc reanalysis: statistical model was added after trial launch, original primary endpoint was not met—this does not equal confirmed efficacy.

🔥 Today’s Top 10 Focus

1. Percival vision-language model extracts cross-organ health signals from CT scans

According to Lifespan.io’s coverage , Percival, a vision-language model developed by the University of Pennsylvania team, was trained on over 400,000 CT image volumes from 52,000+ participants (npj Digital Medicine). The model embeddings showed significant associations with 35 lab markers (including kidney function, liver function, blood glucose, lipids, and inflammation markers) and correlated with 351 cross-organ diagnostic categories; it outperformed image-only baseline models in predicting subsequent heart failure, cirrhosis, acute kidney injury, and other diseases. Researchers note that some “future diagnoses” may be existing but undocumented conditions, so the predictive capability here differs from genuine early warning years in advance. The study has not been directly compared to radiologists, so no conclusion can be drawn that Percival outperforms routine clinical imaging interpretation.

2. Tessera: spatial transcriptomics representation learning framework with directional sensitivity

Paper index/abstract shows that Tessera introduces geometric inductive bias into spatial transcriptomics representation learning while preserving directional information of inter-cell relationships rather than modeling distance alone. The framework employs a translation-invariant design and outperforms existing spatial foundation models across multiple tissues and sequencing technologies, with zero-shot generalization to contexts 16× larger than training maximum. Currently a preprint; specific datasets and evaluation metrics await peer review, and causal mechanisms cannot yet be concluded.

3. HistoLoop: iterative spatial gene expression prediction from histopathology slides

Paper index/abstract introduces HistoLoop, a looped Transformer that iteratively refines spot states in H&E images to predict spatial gene expression. Under the HEST-Bench HVG50 evaluation protocol, HistoLoop achieved the highest macro-average gene Pearson correlation across 6 cancer tasks in 5 organs on the HEST-1K dataset. This is observational prediction; it’s a preprint and has not yet been validated for actual impact on clinical diagnosis or treatment decisions.

4. Brain2Vec: brain structure foundation model pretrained on 74,425 T1-weighted MRIs

Paper index/abstract shows that Brain2Vec, when initialized with pretrained weights for brain age prediction, outperforms random initialization in both final accuracy and convergence speed; frozen embeddings beat anatomical volume feature baselines in 4 out of 5 classification tasks. Model code and trained weights are open-sourced (see CBICA/Brain2Vec ) and also accessible via the NiChart cloud platform. Pretraining data aggregated from multiple studies; downstream tasks are all academic evaluations and not yet validated in clinical deployment.

5. Alterity’s post-hoc reanalysis of MSA drug ATH434 reports 52% slowing of functional decline

According to Longevity.Technology’s coverage , Alterity Therapeutics presented a post-hoc reanalysis of its phase-2 trial at the 2026 Seoul International Parkinson and Movement Disorder Congress (as a poster), adjusting baseline disease severity by cerebrospinal fluid neurofilament light chain (NfL); the 50 mg group showed 4.64 fewer points of worsening on the disability scale at Week 52 versus placebo (p=0.032; minimal clinically important difference for the scale is 1.5 points). Key limitations: the statistical model was added after trial initiation, the original primary endpoint (substantia nigra iron deposition) was not met, and the 75 mg dose did not reach statistical significance; this figure comes from exploratory reanalysis and does not equal confirmed clinical efficacy.

6. VGL-stGen: multimodal large language model for lung histopathology and spatial transcriptomics

Paper index/abstract introduces VGL-stGen, which starts from a pretrained medical vision-language model and combines paired lung H&E images, spatial transcriptomics, and single-cell reference data to support 13 tasks (tissue classification, cell type classification, gene prediction, and conditional expression imputation). Image permutation experiments show that correctly paired tissue images contributed additional information beyond observed expression. Researchers explicitly note that external classification depends on computational references sharing annotated lineages, and that some H&E sources were already exposed during training; results should be extrapolated cautiously.

7. GCLDock: curriculum learning improves protein-ligand flexible docking success rates

Paper index/abstract proposes introducing curriculum learning into flexible docking training, progressively adding samples with large protein conformational changes. On the PDBBind ESMFold apo benchmark, GCLDock achieved a top-1 success rate of 46.52% (ligand RMSD < 2Å), and 40.26% with additional PoseBusters validity constraints, surpassing existing baselines. This task targets protein-ligand binding pose prediction; it’s a computational benchmark evaluation and has not been independently validated in wet-lab experiments or drug discovery pipelines.

8. Duke dementia AI tools clinical translation initiative

According to News-Medical’s coverage , Duke University has launched an initiative aimed at bringing dementia AI tools into routine clinical practice, noting that while many research findings exist, few tools have actually been deployed. Specific tool names, evaluation metrics, enrolled populations, and pilot scale are all at the short-abstract level; detailed data awaits verification, and no inference can be made about clinical diagnostic accuracy or improved patient outcomes.

9. LoRA fine-tuning of vision foundation models converges in performance but diverges in internal mechanisms on brain MRI

Paper index/abstract applied LoRA fine-tuning to four ViT-B/16 backbones—DINOv3, SigLIP2, MerMED-FM, BiomedCLIP—for brain tumor segmentation (6 datasets, average Dice 78–81%). The study found that the four models achieved similar se