Ask or search anything...

This paper introduces TITAN, a multimodal foundation model that effectively processes whole slide pathology images and text through a three-stage pretraining approach

HEST-1k introduces a large, meticulously curated dataset of paired spatial transcriptomics and histology data, alongside a supporting library and benchmark, enabling the evaluation of deep learning models for predicting gene expression from tissue morphology. This resource helps advance the capabilities of foundation models in pathology and facilitates the exploration of morphomolecular relationships.

Researchers developed VORTEX, an artificial intelligence framework that predicts 3D spatial gene expression patterns across entire tissue volumes using 3D morphological imaging data and limited 2D spatial transcriptomics. This method successfully mapped complex expression landscapes and tumor microenvironments in various cancer types, overcoming the limitations of traditional 2D spatial transcriptomics.

Researchers from Harvard, Mass General Brigham, MIT, and collaborating institutions introduce MedBrowseComp, a benchmark that evaluates AI agents' ability to navigate live medical knowledge bases and synthesize multi-hop evidence from sources like HemOnc.org, PubMed, and ClinicalTrials.gov, revealing that frontier agentic systems achieve as low as 10% accuracy on complex medical information retrieval tasks with performance degrading monotonically as navigation depth increases, while Computer Use Agents show superior performance when starting from domain-specific pages rather than general search engines.

Harvard University and affiliated researchers develop Generative Distribution Embeddings (GDEs), a framework that combines distribution-invariant encoders with conditional generative models to learn representations of probability distributions, demonstrating through systematic benchmarking that GDEs outperform Kernel Mean Embeddings and Wasserstein Wormhole on synthetic datasets while achieving successful applications across six computational biology problems including cell population modeling, perturbation effect prediction, DNA methylation pattern analysis, synthetic promoter design, and viral protein sequence modeling.

PANTHER introduces an unsupervised framework for learning whole-slide image representations in computational pathology, leveraging Gaussian Mixture Models to capture morphological prototypes and their prevalence. This approach generates task-agnostic slide embeddings that perform comparably to or better than supervised methods across 13 diverse diagnostic and prognostic tasks, offering inherent interpretability.

Researchers discovered that while Large Reasoning Models (LRMs) can be prompted to generate thinking steps in non-English languages, doing so consistently reduces their accuracy on complex math and science problems. They introduced a new challenging multilingual benchmark, XReasoning, to evaluate this trade-off, showing a clear performance degradation despite improved language alignment from prompt modifications or targeted fine-tuning.

This research systematically investigates Sparse Autoencoders (SAEs) for extracting interpretable features from Gemma 2 models, establishing benchmarks for SAE-based classification and evaluating their transferability across languages and modalities. It demonstrates that SAE-derived features consistently outperform baselines for various classification tasks.

This research introduces TANGLE, a self-supervised learning framework that guides the creation of high-quality, generalized whole-slide image embeddings using paired transcriptomics data. The framework, applied to both human cancer and rat toxicological pathology, achieved superior performance in few-shot classification, slide retrieval, and demonstrated biologically interpretable representations, including the development of iBOT-Tox, the first large-scale SSL model for rodent histology.