This will be a growing challenge for authors. At @ResearchHub we regularly received reviews that were longer than the manuscript itself. Whether a review was written with AI is secondary. What we need is an agreement on what is expected from authors. And obviosuly with AI we can have way more detailed reviews now! I am totally ok with a preprint co-existing with its reviews. I do not think authors should be expected to do every requested experiment, or even be asked for new experiments at all, unless a key control is missing. The preprint, the reviews and any follow-up experiments can simply co-exist to give the maximum context to the reader/AI agent Historically it has been a flex to do every experiment a reviewer asks for, but I personally do not see the logic. I never received funding for those experiments, and I may not even want them in my manuscript.

Human brain cellular xenografts in neuroimmunology and beyond: @ImmunityCP cell.com/immunity/fulltext/S…

Really looking forward to SciCon 2026 to discuss the future of scientific research and funding! @ResearchHubF @ResearchHub

We have another update on our preprints on preprints 🤠 Tracking claim changes from preprint to publication across 72,644 biomedical studies using large language models "Among 399,928 original research articles published in 47 journals b/w 2021 & 2024, retraction occurred at 0.79 per 10,000 article-years among 19,692 articles linked to a bioRxiv preprint (6 events) and 5.20 among 380,236 without an identified preprint (715 events; hazard ratio 0.22...)" 😎 Thanks again for leading this powerful project, Dr. Ruslan Rust @rust_ruslan 🙏 biorxiv.org/content/10.64898…

AlphaCell: Towards building a World Model to simulate perturbation-induced cellular dynamics biorxiv.org/content/10.64898…

We 'll be hiring very actively at Berkeley! Assistant/ Associate/ Full Professor – Electrical Engineering & Computer Sciences (Deadline: Dec. 7, 2026 at 11:59 pm Pacific Time). Apply here: aprecruit.berkeley.edu/JPF05… Assistant Professor – AI for Biology – Electrical Engineering & Computer Sciences and Center for Computational Biology (Deadline: Nov. 9, 2026 at 11:59 pm Pacific Time) Apply here: aprecruit.berkeley.edu/JPF05… Assistant Professor – Intersection of Statistics and CS – Departments of Electrical Engineering & Computer Sciences and Statistics (Deadline: Nov. 10, 2026 at 11:59 pm Pacific Time). Apply here: aprecruit.berkeley.edu/JPF05… Assistant/ Associate/ Full Professor – Computational Biomedicine - Center for Computational Biology (Deadline: Nov. 18 2026 at 11:59 pm Pacific Time). Apply here: aprecruit.berkeley.edu/JPF05… Assistant / Associate Professor – Electronic, Functional, and Quantum Materials – Electrical Engineering & Computer Sciences and Materials Science & Engineering (Deadline: Dec. 7, 2026 at 11:59 pm Pacific Time). Apply here: aprecruit.berkeley.edu/JPF05… Coming Soon! Assistant/ Associate/ Full Professor - AI and the Physical Sciences -Bakar Institute of Digital Materials for the Planet (BIDMaP) and home department(s) to be determined. (Deadline: Dec. 7, 2026 at 11:59 pm Pacific Time) Apply here: aprecruit.berkeley.edu/JPF05… Coming Soon! Assistant/ Associate/ Full Professor-in-Residence - Gladstone Institute of Data Science and Biotechnology and home department(s) to be (Deadline: TBD at 11:59 pm Pacific Time) Apply here: aprecruit.berkeley.edu/JPF05…

A coordinated and reproducible multicellular response to AD pathology biorxiv.org/content/10.64898…

Your APOE genotype is fixed at birth. But 8.6% of its link to Alzheimer's diagnosis appears to pass through DNA methylation, a layer of chemical tags that can change. A new Nature Aging study of 1,250 postmortem brains maps where those tags sit. APOE ε4 is the strongest common genetic risk factor for late-onset Alzheimer's, but a gene variant alone can't be changed or easily targeted. If part of that risk runs through chemical tags on DNA, those tags could become a new set of biomarkers and drug targets, though this paper only tests that statistically. First, some background. DNA methylation is a small chemical tag (a methyl group) attached to DNA letters. It doesn't change the genetic code. It acts more like a dimmer switch, helping decide how strongly a gene is turned on in a given cell. Scientists measure it at sites called CpGs, places in the genome where a C is followed by a G. The researchers used brain tissue from the Rush Alzheimer's Disease Center cohorts. These are older adults who agreed to donate their brains after death, so doctors could pair years of clinical records with a direct look at the tissue. The average age at death was 89.5 years, 69% were women, and 66% met the criteria for a pathological Alzheimer's diagnosis. They measured methylation at about 930,000 CpG sites in the prefrontal cortex, a region involved in planning and memory. Then they tested which sites tracked three things: a formal Alzheimer's diagnosis, amyloid-β load (the sticky plaques that build up between neurons), and tau tangles (twisted protein fibers that build up inside neurons). The key findings: • 107 methylation sites were linked to at least one of the three measures. 9 tracked diagnosis, 65 tracked amyloid, and 46 tracked tau tangles. • Only 12 sites were linked to two or more measures, and just 1 (cg08550059) was linked to all three. • 38 of the 107 sites replicated in an independent set of 139 brains from the Netherlands Brain Bank. • DNA methylation statistically mediated 8.6% of the effect of APOE genotype on Alzheimer's diagnosis (P = 2 × 10⁻⁶). • 39 of the genes tied to these sites, including RIN3, LGMN, and LHPP, showed different activity across major brain cell types such as neurons and astrocytes. The low overlap between measures is worth noticing. Plaques, tangles, and diagnosis are often treated as one disease process. At the methylation level, they look more like partly separate signals, which fits the idea that Alzheimer's is several overlapping processes and not one. The mediation result is the central claim, so it helps to know what the method actually does. Mediation analysis asks how much of the link between A (APOE genotype) and B (diagnosis) travels through C (methylation). Think of it like tracing a river. If APOE is the source and diagnosis is the sea, mediation measures how much of the water flows through one particular tributary. Here, that tributary carried 8.6% of the effect, and the result was statistically strong (P = 2 × 10⁻⁶, meaning a result this size would be very unlikely by chance alone). But a small P value says the signal is probably real. It doesn't say the signal is large, and it doesn't say which way the arrow points. That matters because 8.6% is a modest share. The large majority of APOE's effect runs through other routes, such as how APOE protein handles cholesterol in the brain, how it clears amyloid, and how it shapes inflammation. Methylation is one channel among several, and this study can't say how much of the 8.6% is cause and how much is consequence. A brain that already has more pathology could change its own methylation pattern, and the statistics would look much the same. Some caveats. This is postmortem tissue taken at one moment, at an average age near 90. So the study can't say whether methylation changes cause disease, respond to it, or both. Mediation is a statistical estimate, not proof of a causal chain. Only 38 of 107 sites replicated, and the replication brains were younger on average (80.7 years) and fewer in number. Everything was measured in the prefrontal cortex, so other brain regions could look different. Tissue also contains many cell types, which makes it harder to say which cells carry each change. This work is years from clinical use. The nearest realistic payoffs are better epigenetic biomarkers of pathology and a shortlist of genes worth testing in cells and animal models. What it does add is a reframing. APOE ε4 is often described as a fixed sentence written at birth. This study suggests at least part of that risk is expressed through a regulatory layer that is dynamic. A gene tells you what the brain could do. Methylation helps decide what it does.

Ever wonder if eventually cells could be trained out of disease states? We're on the path; first steps: New #preprint - @WesleyPClawson "Goal directed closed-loop training of network bursts in cortical cultures produces improved performance and error correction over days" biorxiv.org/content/10.64898…

Svante Pääbo and colleagues used AlphaGenome as a time machine to reconstruct the evolution of human gene regulation. An impressive new bioRxiv preprint! 🧬⏩ Fifty years ago, scientists proposed that the differences between humans and chimpanzees might be driven less by changes in genes themselves and more by changes in how those genes are regulated. The authors revisited this hypothesis using deep learning and identified more than 71,000 genomic regions predicted to have gained regulatory activity along the human lineage. But the most interesting finding is how these changes happened! In 71% of cases, evolution repurposed existing regulatory elements rather than creating new ones. A single mutation could introduce a tissue-specific transcription factor binding site, activating an existing element in another tissue. To investigate this, the researchers used AlphaGenome to predict the regulatory effects of 17 million nucleotide substitutions, linking them to gene expression differences between humans and other great apes. Interestingly, the vast majority of these regions fell outside known Human Accelerated Regions, genomic regions that evolved unusually rapidly along the human lineage. This suggests that the rate of sequence evolution does not necessarily reflect its functional impact. Comparisons with Neanderthal and Denisovan genomes revealed that 99.6% of the regulatory gains fixed in modern humans and successfully dated arose before our evolutionary split from these archaic lineages. There are also fascinating examples involving individual genes. One is SLC47A1, which encodes a transporter involved in eliminating drugs and toxins. The researchers identified five regulatory gains associated with increased expression in the human liver. Reverting these changes to their ancestral states in silico reduced predicted gene expression by 60%. Of course, many of these predictions still require experimental validation. But the broader idea is remarkable: ancient DNA has allowed us to discover which mutations occurred in the past. Now, deep learning is beginning to reveal how those mutations may have changed the way genes worked. biorxiv.org/content/10.64898… #aDNA #evolution #regulation #AlphaGenome

We have several postdoc positions available in my lab at @EPFL_en to work at the intersection of AI and biology, spanning foundation models for single-cell biology, spatial biology, multimodal AI and AI agents for scientific discovery. Consider applying, more details here: brbiclab.epfl.ch/join-us/

Human lymph node CD8+ T cell memory is nested within drainage regions biorxiv.org/content/10.64898…

Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro

Recordanelle plasmids now available on @Addgene! addgene.org/browse/article/2…

Where a cell sits in a tissue matters. This Cell Genomics paper uses spatial transcriptomics to reveal altered cell communities and drivers of aberrant epithelia and pro-fibrotic fibroblasts in interstitial lung diseases. hubs.li/Q04B7hWK0 #ImpactfulPaper

Red blood cell-derived extracellular vesicles as a potent autologous adjuvant go.​nature.​com/4ijWhur

The hard part about AI in bio is that by the time you can test the predictions from model X, you already have better predictions from model X+2 or X+3.

Cool paper from Gladyshev lab - mice, rats, macaques and humans age with a shared transcriptomic signature. Shared enough to predict time to death. They pooled 11,000 transcriptomes from over 25 tissues across those four species. Genes tied to senescence, inflammation and apoptosis went up with age in all of them. Two hits, CDKN1A and LGALS3, also tracked mortality and multimorbidity at the protein level in UK Biobank. The resulting clocks predicted time to death in people about as well as second-generation epigenetic clocks, and in rodents they picked up known lifespan-extending interventions. What I find most interesting is the module clocks. You get separate readouts for separate parts of the transcriptome. Chronic disease mostly aged the inflammatory module. Caloric restriction and Klotho deficiency acted on mitochondrial and metabolic modules. A single "biological age" number can't tell you which part of aging an intervention touched. I think this matters most for animal research. If the molecular readout is conserved, a mouse study and a human trial can report aging on the same scale. That's the translation bridge we've mostly been missing. It would be very cool to link this with the automated Intrinsic Capacity panel we're building with ARPA.

We need to do experiments to figure out exactly how these work. But there are way too many candidates! So we created an explorable website (compass.arcinstitute.org/) and MCP (compass.arcinstitute.org/age…) for the research community to mine the COMPASS clusters and make their own discoveries!