AbstractBackgroundFeature selection is a critical step for translating advances afforded by systems-scale molecular profiling into actionable clinical insights. While data-driven methods are commonly utilized for selecting candidate genes, knowledge-driven methods must contend with the challenge of efficiently sifting through extensive volumes of biomedical information. This work aimed to assess the utility of large language models (LLMs) for knowledge-driven gene prioritization and selection.MethodsIn this proof of concept, we focused on 11 blood transcriptional modules associated with an Erythroid cells signature. We evaluated four leading LLMs across multiple tasks. Next, we established a workflow leveraging LLMs. The steps consisted of: (1) Selecting one of the 11 modules; (2) Identifying functional convergences among constituent genes using the LLMs; (3) Scoring candidate genes across six criteria capturing the gene’s biological and clinical relevance; (4) Prioritizing candidate genes and summarizing justifications; (5) Fact-checking justifications and identifying supporting references; (6) Selecting a top candidate gene based on validated scoring justifications; and (7) Factoring in transcriptome profiling data to finalize the selection of the top candidate gene.ResultsOf the four LLMs evaluated, OpenAI's GPT-4 and Anthropic's Claude demonstrated the best performance and were chosen for the implementation of the candidate gene prioritization and selection workflow. This workflow was run in parallel for each of the 11 erythroid cell modules by participants in a data mining workshop. Module M9.2 served as an illustrative use case. The 30 candidate genes forming this module were assessed, and the top five scoring genes were identified as BCL2L1, ALAS2, SLC4A1, CA1, and FECH. Researchers carefully fact-checked the summarized scoring justifications, after which the LLMs were prompted to select a top candidate based on this information. GPT-4 initially chose BCL2L1, while Claude selected ALAS2. When transcriptional profiling data from three reference datasets were provided for additional context, GPT-4 revised its initial choice to ALAS2, whereas Claude reaffirmed its original selection for this module.ConclusionsTaken together, our findings highlight the ability of LLMs to prioritize candidate genes with minimal human intervention. This suggests the potential of this technology to boost productivity, especially for tasks that require leveraging extensive biomedical knowledge. <<<

AbstractCancer diagnosis and therapeutics (theranostics) based on the tumor microenvironment (TME) and biomarkers has been an emerging approach for precision medicine. DNA nanotechnology dynamically controls the self‐assembly of DNA molecules at the nanometer scale to construct intelligent DNA chemical reaction systems. The DNA logic circuit is a particularly emerging approach for computing within the DNA chemical systems. DNA logic circuits can sensitively respond to tumor‐specific markers and the TME through logic operations and signal amplification, to generate detectable signals or to release anti‐cancer agents. In this review, the fundamental concepts of DNA logic circuits are clarified, the basic modules in the circuit are summarized, and how this advanced nano‐assembly circuit responds to tumor‐related molecules, how to perform logic operations, to realize signal amplification, and selectively release drugs through discussing over 30 application examples, are demonstrated. This review shows that DNA logic circuits have powerful logic judgment and signal amplification functions in improving the specificity and sensitivity of cancer diagnosis and making cancer treatment controllable. In the future, researchers are expected to overcome the existing shortcomings of DNA logic circuits and design smarter DNA devices with better biocompatibility and stability, which will further promote the development of cancer theranostics. <<<

Abstract<br> The societal shift toward greater gender equality has led to increased variability in people’s gender role attitudes, or the belief that men and women should occupy distinct family roles (i.e. men as breadwinners and women as homemakers). Existing evidence on the association between gender role attitudes and relationship well-being remains inconclusive with mixed findings, likely because past research has not adequately considered the direction and degree of (in)congruencies between partners within the relationship. Using longitudinal samples of 1,327 couples from the United States and 5,856 couples from Germany tracked over 2 and 13 years, respectively, we employed dyadic response surface analysis to examine how different patterns of partner (in)congruencies in gender role attitudes predict relationship well-being in mixed-gender relationships. The results showed that, for US men and German men and women, the direction of incongruence between partners’ gender role attitudes mattered: relationship satisfaction was higher when men adopted more egalitarian attitudes than women (or conversely, when women adopted more traditional attitudes than men) compared with the reverse. Relationship satisfaction was also higher when both partners showed congruence in extreme gender role attitudes (either strongly traditional or egalitarian) than when either partner endorsed more neutral attitudes. US women reported higher relationship satisfaction only when either partner endorsed more egalitarian attitudes. Although past research emphasizes the benefits of partner similarity for relationship well-being, our findings highlight the importance of both similarity and complementarity in gender role attitudes, potentially subject to cultural and contextual factors. <<<

This article is about the neural conundrum behind the slowness of human<br>behavior. The information throughput of a human being is about 10 bits/s. In<br>comparison, our sensory systems gather data at ~10^9 bits/s. The stark contrast<br>between these numbers remains unexplained and touches on fundamental aspects of<br>brain function: What neural substrate sets this speed limit on the pace of our<br>existence? Why does the brain need billions of neurons to process 10 bits/s?<br>Why can we only think about one thing at a time? The brain seems to operate in<br>two distinct modes: the "outer" brain handles fast high-dimensional sensory and<br>motor signals, whereas the "inner" brain processes the reduced few bits needed<br>to control behavior. Plausible explanations exist for the large neuron numbers<br>in the outer brain, but not for the inner brain, and we propose new research<br>directions to remedy this. <<<

<br> Understanding how numerous quantitative trait loci (QTL) shape phenotypic variation is an important question in genetics. To address this, we established a permanent population of 18,421 (18K) rice lines with reduced population structure. We generated reference-level genome assemblies of the founders and genotyped all 18K-rice lines through whole-genome sequencing. Through high-resolution mapping, 96 high-quality candidate genes contributing to variation in 16 traits were identified, including<br> <i>OsMADS22</i><br> and<br> <i>OsFTL1</i><br> verified as causal genes for panicle number and heading date, respectively. We identified epistatic QTL pairs and constructed a genetic interaction network with 19 genes serving as hubs. Overall, 170 masking epistasis pairs were characterized, serving as an important factor contributing to genetic background effects across diverse varieties. The work provides a basis to guide grain yield and quality improvements in rice.<br> <<<

AbstractRest and sleep not only strengthen existing memories but also reorganise memories to generate new knowledge that extends beyond direct experience. However, it remains unclear both<i>how</i>memories are reorganised and the effect of this reorganisation on behaviour. Here, we designed a novel protocol to casually manipulate memory consolidation during rest using awake, contextual targeted memory reactivation (TMR). We found that promoting memory consolidation during rest qualitatively reorganises memories by<i>forming ‘shortcuts’ between memories</i>which have not been experienced together. These shortcuts in memory extend beyond direct experience to facilitate our ability to make novel inferences. A series of control tests indicate that inference performance cannot be explained by quantitative strengthening of the experienced component links but are rather explained by qualitative changes in the cognitive map which involve formation of new shortcuts. Interestingly, we show that representing a shortcut may come with limitations, as shortcuts cannot be readily updated in response to rapid changes in the environment. Together, these findings reveal how memories are reorganised during awake rest to construct a cognitive map of our environment, while highlighting the constraints set by a trade-off between efficient and flexible behaviour. <<<

AbstractSchizophrenia is a prototypical network disorder with widespread brain-morphological alterations, yet it remains unclear whether these distributed alterations robustly reflect the underlying network layout. We tested whether large-scale structural alterations in schizophrenia relate to normative structural and functional connectome architecture, and systematically evaluated robustness and generalizability of these network-level alterations. Leveraging anatomical MRI scans from 2439 adults with schizophrenia and 2867 healthy controls from 26 ENIGMA sites and normative data from the Human Connectome Project (<i>n</i> = 207), we evaluated structural alterations of schizophrenia against two network susceptibility models: (i) hub vulnerability, which examines associations between regional network centrality and magnitude of disease-related alterations; (ii) epicenter mapping, which identifies regions whose typical connectivity profile most closely resembles the disease-related morphological alterations. To assess generalizability and specificity, we contextualized the influence of site, disease stages, and individual clinical factors and compared network associations of schizophrenia with that found in affective disorders. Our findings show schizophrenia-related cortical thinning is spatially associated with functional and structural hubs, suggesting that highly interconnected regions are more vulnerable to morphological alterations. Predominantly temporo-paralimbic and frontal regions emerged as epicenters with connectivity profiles linked to schizophrenia’s alteration patterns. Findings were robust across sites, disease stages, and related to individual symptoms. Moreover, transdiagnostic comparisons revealed overlapping epicenters in schizophrenia and bipolar, but not major depressive disorder, suggestive of a pathophysiological continuity within the schizophrenia-bipolar-spectrum. In sum, cortical alterations over the course of schizophrenia robustly follow brain network architecture, emphasizing marked hub susceptibility and temporo-frontal epicenters at both the level of the group and the individual. Subtle variations of epicenters across disease stages suggest interacting pathological processes, while associations with patient-specific symptoms support additional inter-individual variability of hub vulnerability and epicenters in schizophrenia. Our work outlines potential pathways to better understand macroscale structural alterations, and inter- individual variability in schizophrenia. <<<

We present an approach for estimating the fraction of text in a large corpus<br>which is likely to be substantially modified or produced by a large language<br>model (LLM). Our maximum likelihood model leverages expert-written and<br>AI-generated reference texts to accurately and efficiently examine real-world<br>LLM-use at the corpus level. We apply this approach to a case study of<br>scientific peer review in AI conferences that took place after the release of<br>ChatGPT: ICLR 2024, NeurIPS 2023, CoRL 2023 and EMNLP 2023. Our results suggest<br>that between 6.5% and 16.9% of text submitted as peer reviews to these<br>conferences could have been substantially modified by LLMs, i.e. beyond<br>spell-checking or minor writing updates. The circumstances in which generated<br>text occurs offer insight into user behavior: the estimated fraction of<br>LLM-generated text is higher in reviews which report lower confidence, were<br>submitted close to the deadline, and from reviewers who are less likely to<br>respond to author rebuttals. We also observe corpus-level trends in generated<br>text which may be too subtle to detect at the individual level, and discuss the<br>implications of such trends on peer review. We call for future<br>interdisciplinary work to examine how LLM use is changing our information and<br>knowledge practices. <<<

This paper shows that masked autoencoders (MAE) are scalable self-supervised<br>learners for computer vision. Our MAE approach is simple: we mask random<br>patches of the input image and reconstruct the missing pixels. It is based on<br>two core designs. First, we develop an asymmetric encoder-decoder architecture,<br>with an encoder that operates only on the visible subset of patches (without<br>mask tokens), along with a lightweight decoder that reconstructs the original<br>image from the latent representation and mask tokens. Second, we find that<br>masking a high proportion of the input image, e.g., 75%, yields a nontrivial<br>and meaningful self-supervisory task. Coupling these two designs enables us to<br>train large models efficiently and effectively: we accelerate training (by 3x<br>or more) and improve accuracy. Our scalable approach allows for learning<br>high-capacity models that generalize well: e.g., a vanilla ViT-Huge model<br>achieves the best accuracy (87.8%) among methods that use only ImageNet-1K<br>data. Transfer performance in downstream tasks outperforms supervised<br>pre-training and shows promising scaling behavior. <<<

Most robotic hands and grippers rely on actuators with large gearboxes and<br>force sensors for controlling gripping force. However, this might not be ideal<br>for tasks that require the robot to interact with an unstructured and unknown<br>environment. In this paper, we introduce a novel quasi-direct-drive<br>two-fingered robotic hand with variable impedance control in the joint space<br>and Cartesian space. The hand has a total of four degrees of freedom,<br>backdrivable differential gear trains, and four brushless direct current (BLDC)<br>motors. Motor torque is controlled through Field-Oriented Control (FOC) with<br>current sensing. Variable impedance control enables the robotic hand to execute<br>dexterous manipulation tasks safely during environment-robot and human-robot<br>interactions. The quasi-direct-drive actuators eliminate the need for complex<br>tactile/force sensors or precise motion planning when handling environmental<br>contact. A majority-3D-printed assembly makes this a low-cost research platform<br>built with affordable, readily available off-the-shelf components. Experimental<br>validation demonstrates the robotic hand's capability for stable force-closure<br>and form-closure grasps in the presence of disturbances, reliable in-hand<br>manipulation, and safe dynamic manipulations despite contact with the<br>environment. <<<

This paper proposes a new game-search algorithm, PN-MCTS, which combines<br>Monte-Carlo Tree Search (MCTS) and Proof-Number Search (PNS). These two<br>algorithms have been successfully applied for decision making in a range of<br>domains. We define three areas where the additional knowledge provided by the<br>proof and disproof numbers gathered in MCTS trees might be used: final move<br>selection, solving subtrees, and the UCB1 selection mechanism. We test all<br>possible combinations on different time settings, playing against vanilla UCT<br>on several games: Lines of Action ($7$$\times$$7$ and $8$$\times$$8$ board<br>sizes), MiniShogi, Knightthrough, and Awari. Furthermore, we extend this new<br>algorithm to properly address games with draws, like Awari, by adding an<br>additional layer of PNS on top of the MCTS tree. The experiments show that<br>PN-MCTS is able to outperform MCTS in all tested game domains, achieving win<br>rates up to 96.2% for Lines of Action. <<<