Patterns最新文献

In a recent issue of Cell Reports Physical Science, Zhao et al. introduced ChemDFM, a foundational large language model designed specifically for chemistry. The model bridges the gap between general-purpose language models and specialized chemical knowledge, including the integration of multimodal capabilities for spectroscopic data interpretation, improved numerical reasoning, and connectivity with chemical tools and databases to enhance practical research applications. This approach demonstrates how domain adaptation can transform AI tools into collaborative research partners for scientific discovery.

We introduce MediSim, a multi-modal generative model for simulating and augmenting electronic health records across multiple modalities, including structured codes, clinical notes, and medical imaging. MediSim employs a multi-granular, autoregressive architecture to simulate missing modalities and visits and iterative, reinforcement learning-based training to improve simulation in low-data settings. Additionally, it utilizes encoder-decoder model pairs to handle complex modalities like notes and images. Experiments on outpatient claims and inpatient ICU datasets have demonstrated MediSim's superiority over baselines in predicting missing codes, creating enriched data, and improving downstream predictive modeling. Specifically, MediSim improved over 74% on missing code prediction, enabled up to 65% better downstream predictive performance compared to original deficient records missing either some visits or entire data modalities, and successfully produced realistic note and X-ray samples for use in downstream tasks. MediSim's ability to generate comprehensive, high-dimensional EHR data has the potential to significantly improve AI applications throughout healthcare.

This review explores the role of prompt engineering in unleashing the capabilities of large language models (LLMs). Prompt engineering is the process of structuring inputs, and it has emerged as a crucial technique for maximizing the utility and accuracy of these models. Both foundational and advanced prompt engineering methodologies-including techniques such as self-consistency, chain of thought, and generated knowledge, which can significantly enhance the performance of models-are explored in this paper. Additionally, the prompt methods for vision language models (VLMs) are examined in detail. Prompt methods are evaluated with subjective and objective metrics, ensuring a robust analysis of their efficacy. Critical to this discussion is the role of prompt engineering in artificial intelligence (AI) security, particularly in terms of defending against adversarial attacks that exploit vulnerabilities in LLMs. Strategies for minimizing these risks and improving the robustness of models are thoroughly reviewed. Finally, we provide a perspective for future research and applications.

We describe NLSExplorer, an interpretable approach for nuclear localization signal (NLS) prediction. By utilizing the extracted information on nuclear-specific sites from the protein language model to assist in NLS detection, NLSExplorer achieves superior performance with greater than 10% improvement in the F1 score compared with existing methods on benchmark datasets and highlights other nuclear transport segments. We applied NLSExplorer to the nucleus-localized proteins in the Swiss-Prot database to extract valuable segments. A comprehensive analysis of these segments revealed a potential NLS landscape and uncovered features of nuclear transport segments across 416 species. This study introduces a powerful tool for exploring the NLS universe and provides a versatile network that can efficiently detect characteristic domains and motifs.

Gender bias in machine translation (MT) has been studied for over a decade, a time marked by societal, linguistic, and technological shifts. With the early optimism for a quick solution in mind, we review over 100 studies on the topic and uncover a more complex reality-one that resists a simple technical fix. While we identify key trends and advancements, persistent gaps remain. We argue that there is no simple technical solution to bias. Building on insights from our review, we examine the growing prominence of large language models and discuss the challenges and opportunities they present in the context of gender bias and translation. By doing so, we hope to inspire future work in the field to break with past limitations and to be less focused on a technical fix; more user-centric, multilingual, and multiculturally diverse; more personalized; and better grounded in real-world needs.

Natural products (NPs) play a vital role in drug discovery, with many FDA-approved drugs derived from these compounds. Despite their significance, the biosynthetic pathways of NPs remain poorly characterized due to their inherent complexity and the limitations of traditional retrosynthesis methods in predicting such intricate reactions. While template-free machine learning models have demonstrated promise in organic synthesis, their application to biosynthetic pathways is still in its infancy. Addressing this gap, we propose the graph-sequence enhanced transformer (GSETransformer), which leverages both graph structural information and sequential dependencies to achieve superior performance in addressing the complexity of biosynthetic data. When evaluated on benchmark datasets, GSETransformer achieves state-of-the-art performance in single- and multi-step retrosynthesis tasks. These results highlight its effectiveness in computational biosynthesis and its potential to facilitate the design of NP-based therapeutics.

Neural networks have long strived to emulate the learning capabilities of the human brain. While deep neural networks (DNNs) draw inspiration from the brain in neuron design, their training methods diverge from biological foundations. Backpropagation, the primary training method for DNNs, requires substantial computational resources and fully labeled datasets, presenting major bottlenecks in development and application. This work demonstrates that by returning to biomimicry, specifically mimicking how the brain learns through pruning, we can solve various classical machine learning problems while utilizing orders of magnitude fewer computational resources and no labels. Our experiments successfully personalized multiple speech recognition and image classification models, including ResNet50 on ImageNet, resulting in an increased sparsity of approximately 70% while simultaneously improving model accuracy to around 90%, all without the limitations of backpropagation. This biologically inspired approach offers a promising avenue for efficient, personalized machine learning models in resource-constrained environments.

Multisensory perception produces vast amounts of data requiring efficient processing. This paper focuses on the multisensory example of touch in biological and artificial systems. We integrate philosophical theories of multisensory perception with neuromorphic hardware and demonstrate how classical sensory integration concepts can enhance artificial sensory systems. This approach bridges theoretical neuroscience and computational applications using philosophical tools. We contrast human touch perception, involving feature binding, with artificial perception in neuromorphic computing, where such integration is absent. Two theoretical frameworks are of interest, feature binding and modalities as conventional kinds, in evaluating their relevance to artificial touch. Our findings suggest that a hardware-tailored adaptation of the conventional modalities approach accurately reflects artificial touch perception. Unlike human perception, artificial systems process sensory data separately, lacking binding mechanisms. We explore the implications of these differences, highlighting challenges in replicating human sensory experiences and the role of subjective experience in perception.

Biomarkers play a central role in medicine's gradual progress toward proactive, personalized precision diagnostics and interventions. However, finding biomarkers that provide very early indicators of a change in health status, for example, for multifactorial diseases, has been challenging. The discovery of such biomarkers stands to benefit significantly from advanced information processing and means to detect complex correlations, which quantum computing offers. In this perspective, quantum algorithms, particularly in machine learning, are mapped to key applications in biomarker discovery. The opportunities and challenges associated with the algorithms and applications are discussed. The analysis is structured according to different data types-multidimensional, time series, and erroneous data-and covers key data modalities in healthcare-electronic health records, omics, and medical images. An outlook is provided concerning open research challenges.

Many diseases are the end outcomes of multifactorial risks that interact and increment over months or years. Time-series AI methods have attracted increasing interest given their ability to operate on native time-series data to predict disease outcomes. Instantiating such models in risk stratification tools has proceeded more slowly, in part limited by factors such as structural complexity, model size, and explainability. Here, we present RiskPath, an explainable AI toolbox that offers advanced time-series methods and additional functionality relevant to risk stratification use cases in classic and emerging longitudinal cohorts. Theoretically informed optimization is integrated in prediction to specify optimal model topology or explore performance-complexity trade-offs. Accompanying modules allow the user to map the changing importance of predictors over the disease course, visualize the most important antecedent time epochs contributing to disease risk, or remove predictors to construct compact models for clinical applications with minimal performance impact.