Muhammad Elsadany, he/him @brainiologist

Stratifying the autisms by a type I versus type II distinction in early development | Nature Mental Health nature.com/articles/s4422…

I couldn't find a tool to plot different #neuroimaging data in one consistent style, so I made one! Meet yabplot (yet another brain plot) - a #Python package for (sub)cortex & tracts.🧠 - Simple API - Built-in atlases - Custom atlas support 🔗 github.com/teanijarv/yabp… (drop a ⭐️!)

🧠✨ Older neurons get richer connections in the human brain! Our new paper is out in Nature Communications. We reveal a fundamental link between ᴅᴇᴠᴇʟᴏᴘᴍᴇɴᴛ ⛓️ ʙʀᴀɪɴ ꜱᴛʀᴜᴄᴛᴜʀᴇ & ꜰᴜɴᴄᴛɪᴏɴ in the human connectome 🧵👇

The cerebellum supports high-level language?? Now out in @NeuroCellPress, we systematically examined language-responsive areas of the cerebellum using precision fMRI and identified a *cerebellar satellite* of the neocortical language network! authors.elsevier.com/a/1mUU83BtfHC-… 1/n🧵👇

New in #JNeurosci from Kries et al: People with stroke-related aphasia have key differences in brain activity compared to healthy controls that may explain their impaired verbal speech processing. doi.org/10.1523/JNEURO…

𝗧𝗵𝗲 𝗘𝗻𝘁𝗮𝗻𝗴𝗹𝗲𝗱 𝗕𝗿𝗮𝗶𝗻 Annual reminder that the book is open access. How do we think of the brain as a deeply interconnected system with highly distributed, non hierarchical processing. Want to learn about the brain from a fresh perspective? mitpress.mit.edu/9780262544603/…

Online now: How heterogeneity shapes dynamics and computation in the brain dlvr.it/TQ4ccQ

Ritalin works, but not how we thought it did. Out today @CellCellPress : cell.com/cell/fulltext/… No effect on attention systems. Boosts arousal and anticipated reward. It doesn't help you pay attention better, but it lets you persist in boring tasks. bsky.app/profile/benjam…

𝗥𝗲𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝘁𝗵𝗲 𝗰𝗲𝗻𝘁𝗿𝗮𝗹𝗶𝘁𝘆 𝗼𝗳 𝗯𝗿𝗮𝗶𝗻 𝗮𝗿𝗲𝗮𝘀 New paper I anticipate will become a classic in neuroscience and a must-read for students at all levels. Understanding brain function beyond brain areas. doi.org/10.1038/s41593…

netneurotools: a trainee-oriented approach to network neuroscience | doi.org/10.1101/2025.0… Our lab’s internal toolkit for accomplishing everyday tasks in brain imaging ⤵️

fMRI signals “up,” but neural metabolism might be going “down.” In our @NatureNeuro paper, we show that about 40% of voxels with robust BOLD responses exhibit opposite oxygen metabolism, revealing two distinct hemodynamic modes. rdcu.be/eUPO8 funding @ERC_Research 🧵:

scplotter provides a set of functions to visualize single-cell sequencing data in an easy and efficient way github.com/pwwang/scplott…

New @Nature study maps the genetic risk shared across psychiatric disorders. In News & Views, I discuss what it means when that genetic risk overlaps with normal traits, including education-related outcomes. News & Views: nature.com/articles/d4158… Study: nature.com/articles/s4158…

CASSIA: Multi-agent LLMs for trustworthy single-cell annotation Single-cell RNA-seq studies live or die by good cell type annotation. Today, that often means stitching together multiple tools, curated marker lists, and a lot of expert time—only to end up with inconsistent labels that are hard to interpret and even harder to audit. LLM-based methods like GPTCelltype showed that large models can help, but they also inherit classic LLM failure modes: hallucinations, overconfidence, and opaque reasoning. Elliot Xie and coauthors present CASSIA, a multi-agent LLM framework designed specifically for automated and interpretable cell annotation. Instead of a single prompt, CASSIA orchestrates a team of agents: one for annotation, one for validation, one for formatting, one for quality scoring, and one for reporting. Optional agents handle harder cases—subclustering, uncertainty quantification, “annotation boosting” for low-confidence clusters, and a RAG module that pulls in cell marker databases and ontologies when local knowledge is thin. The result is not just a label, but a full reasoning trace plus a 0–100 quality score that tells you how much to trust it. On benchmarks spanning 970 cell types from GTEx, Tabula Sapiens, Human Cell Landscape, Mouse Cell Atlas, and Azimuth, CASSIA increases fully correct annotations by 12–41% and boosts overall accuracy by more than 20% compared with state-of-the-art methods, including GPT-based approaches. It holds up in harder settings too: cancer vs. non-cancer cells, metastatic signatures, fine-grained T-cell states, mixed cell populations, and even cross-species datasets (from shark to tiger). The quality scores and consensus metrics reliably flag low-quality or heterogeneous clusters and can even uncover errors in “gold-standard” reference annotations. Beyond single-cell analysis, CASSIA is a compelling blueprint for how multi-agent LLM systems can move from one-shot prompting to adaptive, expert-like workflows with built-in self-critique and uncertainty estimates. For single-cell biologists, it promises faster, more accurate, and auditable annotations. For AI-for-science more broadly, it points toward LLMs that behave less like black boxes and more like transparent collaborators in complex data analysis. Paper: nature.com/articles/s4146…

HiP-CT is a new technique that can hierarchically image whole human organs, beginning with the resolution of a human hair (25μm/voxel) down to a resolution where we can see single cells (1.5μm/voxel) [📹 Paul Tafforeau]

Check out our RNA fingerprinting preprint : a statistical framework that maps new data onto perturbation dictionaries to infer causal regulators of cellular responses. TL/DR: how to use others people's Perturb-seq to interpret your scRNA-seq experiment biorxiv.org/content/10.110…

This paper really is groundbreaking. It solves a long-standing embarrassment in machine learning: despite all the hype around deep learning, traditional tree-based methods (XGBoost, CatBoost, random forests, etc) have dominated tabular data—the most common data format in real-world applications—for two decades. Deep learning conquered images, text, and games, but spreadsheets remained stubbornly resistant. This paper's (published in Nature by the way) main contribution is a foundation model that finally beats tree-based methods convincingly on small-to-medium datasets, and does so very fast. TabPFN in 2.8 seconds outperforms CatBoost tuned for 4 hours—a 5,000× speedup. That's not incremental; it's a different regime entirely. The training approach is also fundamentally different. GPT trains on internet text; CLIP trains on image-caption pairs. TabPFN trains on entirely synthetic data—over 100 million artificial datasets generated from causal graphs. TabPFN generates training data by randomly constructing directed acyclic graphs where each edge applies a random transformation (using neural networks, decision trees, discretization, or noise), then pushes random noise through the root nodes and lets it propagate through the graph—the intermediate values at various nodes become features, one becomes the target, and post-processing adds realistic messiness like missing values and outliers. By training on millions of these synthetic datasets with very different structures, the model learns general prediction strategies without ever seeing real data. The inference mechanism is also unusual. Rather than finetuning or prompting, TabPFN performs both "training" and prediction in a single forward pass. You feed it your labeled training data and unlabeled test points together, and it outputs predictions immediately. There's no gradient descent at inference time—the model has learned how to learn from examples during pretraining. The architecture respects tabular structure with two-way attention (across features within a row, then across samples within a column), unlike standard transformers that treat everything as a flat sequence. So, the transformer has basically learned to do supervised learning. Talk to the paper on ChapterPal: chapterpal.com/s/a1899430/acc… Download the PDF: nature.com/articles/s4158…

Glad Anthropic bought the fast Js runtime.

Anthropic @AnthropicAI

7 months ago

Anthropic is acquiring @bunjavascript to further accelerate Claude Code’s growth. We're delighted that Bun—which has dramatically improved the JavaScript and TypeScript developer experience—is joining us to make Claude Code even better. Read more: anthropic.com/news/anthropic…

717 1K 9K 7.8M 1K

Academic flexes that I dislike: 1/🧵

Nature research paper: Shared and language-specific phonological processing in the human temporal lobe go.nature.com/3LGaIMO