Patterns最新文献

Deep neural networks (DNNs) are practical and effective but, despite the name, they lack biological validity. The recent study by Kang et al.¹ in Patterns takes a step toward rectifying this deficit by hard-wiring receptive fields into the first layer of a visual DNN, and the authors show that their network can generalize across image types. Training on photographs, for example, resulted in good performance on sketches; conventional DNNs did not match this behavior.

Many oncology predictive models fail to improve care. Issues include risks of bias, underpowered radiomics studies, and limited clinical impact. A path forward involves an emphasis on clinically actionable questions, rigor, and generalizability.

Foundation models (FMs) have been built to analyze single-cell data with different degrees of success. Here, we present scELMo (single-cell embedding from language models), a method for analyzing single-cell data with the help of large language models (LLMs). LLMs can generate both the description of metadata information and the embeddings for such descriptions. We then combine the embeddings from LLMs with the raw data under the zero-shot learning framework to further extend its function by using the fine-tuning framework to handle different tasks. We demonstrate that scELMo is capable of cell clustering, batch effect correction, and cell-type annotation without training a new model. Moreover, the fine-tuning framework of scELMo can help with more challenging tasks, including in silico treatment analysis or modeling perturbation. scELMo has a lighter structure and lower requirements for resources, suggesting a more promising path.

Biological brains can effortlessly adapt to continuously changing stimulus environments, whereas conventional deep neural networks (DNNs) remain highly susceptible to domain shifts. Here, we demonstrate that static, hard-wired receptive fields, which spontaneously emerge in the early visual pathway, facilitate environment-agnostic object recognition in the brain. To test this mechanism, we introduced pre-developed Gabor filters in the early layers of DNNs, keeping them fixed during training. Despite the reduced learning flexibility, our networks exhibited robust continual learning capabilities under significant domain shifts, unlike conventional DNNs, which fail to generalize under similar conditions. Our network achieved generalized representations across domains in the latent space, while conventional DNNs only captured domain-specific variance. The static visual filters helped prevent local texture biases, leading to shape-based perception similar to that of primates. These findings highlight an intrinsic biological strategy that enables reliable continual learning in dynamic and unpredictable environments.

Complex systems can often be analyzed at either the microscale of their individual components or the macroscale of their collective organization, yet it remains debated which level of description offers the most meaningful causal understanding. Hoel's recent study in Patterns addresses this challenge by introducing Causal Emergence 2.0, a novel formalization showing that a system's causal workings are best described by how causal influence is distributed across its hierarchy of scales.

Complex systems can be described at myriad different scales, and their causal workings often have a multiscale structure (e.g., a computer can be described at the microscale of its hardware circuitry, the mesoscale of its machine code, and the macroscale of its operating system). While scientists study and model systems across the full hierarchy of their scales, from microphysics to macroeconomics, there is debate about what the macroscales of systems can possibly add beyond mere compression. To resolve this long-standing issue, here, a new theory of emergence is introduced that can distinguish which scales irreducibly contribute to a system's causal workings. The theory's application is demonstrated in coarse grains of Markov chains, revealing a novel measure of emergent complexity: how widely distributed a system's causal contributions are across its hierarchy of scales.

The integration of generative AI (GenAI) into academic workflows represents a fundamental shift in scientific practice. While these tools can amplify productivity, they risk eroding the cognitive foundations of expertise by simulating the very tasks through which scientific competence is developed, from synthesis to experimental design to writing. Uncritical reliance can lead to skill atrophy and AI complacency. We propose a framework of essential AI meta-skills: strategic direction, critical discernment, and systematic calibration. These constitute a new form of scientific literacy that builds on traditional critical thinking. Through domain-specific examples and a pedagogical model based on situated learning, we show how these meta-skills can be cultivated to ensure that researchers, particularly trainees, maintain intellectual autonomy. Without deliberate cultivation of these meta-skills, we risk creating the first generation of researchers who serve their tools rather than direct them.

Generative artificial intelligence can be used to create realistic new data, even for complex real-world processes that cannot be exhaustively modeled: the model is simply learned from preexisting data. Generative artificial intelligence is therefore expected to be a game changer in omics research, where data collection is hampered by considerable experimental constraints. However, it can also "hallucinate"-i.e., create data that are too original to be realistic-which is a critical issue in molecular biology, as hallucinated inferences could have devastating consequences. The author thus explores various use cases to mitigate hallucination-induced risks and to safely unleash the full potential of generative methods.

Multiplexed spatial proteomics profiling platforms expose the intricate geometric structure of cells in the tumor microenvironment (TME). The spatial arrangement of cells has been shown to have important clinical implications, correlating with disease prognosis and treatment response. These datasets require new statistical methods to test whether cell-level images are associated with patient-level outcomes. We propose the topological kernel association test (TopKAT), which combines persistent homology with kernel testing to determine whether geometric structures created by cells predict continuous, binary, or survival outcomes. TopKAT quantifies the topological structure of cells in each image using persistence diagrams and compares the similarities between persistence diagrams on the basis of the number and lifespan of the detected homologies among cells. We show that TopKAT can be more powerful than existing approaches, particularly when cells arise along the boundary of a ring and demonstrate its utility in breast cancer and colorectal cancer applications.

Large language models (LLMs) excel in textual tasks but often struggle with physical-world reasoning tasks. Inspired by human cognition-where perception is fundamental to reasoning-we explore augmenting LLMs with enhanced perception abilities using Internet of Things (IoT) data and pertinent knowledge. In this work, we systematically study LLMs' capability to address IoT-sensory tasks, by augmenting their perception and knowledge base, and then propose a unified framework, IoT-LLM, to enhance such capability. In IoT-LLM, we customize three steps: preprocessing IoT data into suitable formats, expanding LLMs' knowledge via IoT-oriented retrieval-augmented generation, and activating LLMs' commonsense knowledge through chain-of-thought prompting. We design a benchmark comprising five real-world tasks with varying data types and reasoning complexities to evaluate the performance of IoT-LLM. Experimental results reveal that IoT-LLM significantly improves the performance of IoT-sensory task reasoning of LLMs, with models such as GPT-4o-mini showing a 49.4% average improvement over previous methods.