Uncontrolled proliferation is a hallmark of cancer, yet tumour cells can enter G0 arrest by halting the cell cycle reversibly (quiescence) or irreversibly (senescence) to survive under stress and in hostile tumour microenvironments (TME). G0 arrested cells contribute to drug tolerance, metastasis and recurrence, but their identification remains challenging due to their rarity and elusive regulatory pathways. Here, we quantify G0 arrest and proliferation decisions in single-cell and spatially profiled breast primary tumours to unveil adaptive responses driving immune compartmentalisation. We identify a genomically-constrained and prolonged G0 arrest state resembling that of dormant precursors of cancer progression. This state featured adaptive transcriptional reprogramming, including unfolded protein response stress, reduced mitochondrial translation and epithelial-mesenchymal plasticity enabled by semaphorin signalling. Spatial transcriptomics revealed G0 arrest pockets encapsulated by APOE+ lipid-associated macrophages, myofibroblastic CAFs and immature perivascular cells, suggestive of an immunosuppressive niche contrasting with cytotoxic environments in proliferative hotspots and displaying distinct drug sensitivities. To facilitate future research, we provide a foundation model capturing G0 arrest and proliferation with 94% accuracy in single cell data, available at https://github.com/secrierlab/G0-LM. Our findings provide new insights into the spatial organisation of cell cycle arrest in breast cancer, highlighting the role of G0 states in tumour heterogeneity and adaptation. ### Competing Interest Statement The authors have declared no competing interest.

Epithelial-mesenchymal plasticity plays a significant role in various biological processes including tumour progression and chemoresistance. However, the expression programmes underlying the epithelial-mesenchymal transition (EMT) in cancer are diverse, and accurately defining the EMT status of tumour cells remains a challenging task. In this study, we employed a pre-trained single-cell large language model (LLM) to develop an EMT-language model (EMT-LM) that allows us to capture discrete states within the EMT continuum in single cell cancer data. In capturing EMT states, we achieved an average Area Under the Receiver Operating Characteristic curve (AUROC) of 90% across multiple cancer types. We propose a new metric, ADESI, to aid the biological interpretability of our model, and derive EMT signatures liked with energy metabolism and motility reprogramming underlying these state switches. We further employ our model to explore the emergence of EMT states in spatial transcriptomics data, uncovering hybrid EMT niches with contrasting potential for antitumour immunity or immune evasion. Our study provides a proof of concept that LLMs can be applied to characterise cell states in single cell data, and proposes a generalisable framework to predict EMT in single cell RNA-seq that can be adapted and expanded to characterise other cellular states. ### Competing Interest Statement The authors have declared no competing interest.

Hematoxylin and eosin (H&E) stained slides are widely used in disease diagnosis. Remarkable advances in deep learning have made it possible to detect complex molecular patterns in these histopathology slides, suggesting automated approaches could help inform pathologists' decisions. Multiple instance learning (MIL) algorithms have shown promise in this context, outperforming transfer learning (TL) methods for various tasks, but their implementation and usage remains complex. We introduce HistoMIL, a Python package designed to streamline the implementation, training and inference process of MIL-based algorithms for computational pathologists and biomedical researchers. It integrates a self-supervised learning module for feature encoding, and a full pipeline encompassing TL and three MIL algorithms: ABMIL, DSMIL, and TransMIL. The PyTorch Lightning framework enables effortless customization and algorithm implementation. We illustrate HistoMIL's capabilities by building predictive models for 2,487 cancer hallmark genes on breast cancer histology slides, achieving AUROC performances of up to 85%.

In this paper, we study the effective field theory (EFT) of dark energy (DE) for the k -essence model beyond linear order. Using particle-mesh N -body simulations that consistently solve the DE evolution on a grid, we find that the next-to-leading order in the EFT expansion, which comprises the terms of the equations of motion that are quadratic in the field variables, gives rise to a generic instability in the regime of low speed of sound (high Mach number). We rule out the possibility of a numerical artefact by considering simplified cases in spherically and plane symmetric situations analytically. If the speed of sound vanishes exactly, the non-linear instability makes the evolution singular in finite time, signalling a breakdown of the EFT framework. The case of finite (but small) speed of sound is subtle, and the local singularity could be replaced by some other type of behaviour with strong non-linearities. While an ultraviolet completion may cure the problem in principle, there is no reason why this should be the case in general. As a result, for a large range of the effective speed of sound c s , a linear treatment is not adequate.

BACKGROUND:Homologous recombination is a robust, broadly error-free mechanism of double-strand break repair, and deficiencies lead to PARP inhibitor sensitivity. Patients displaying homologous recombination deficiency can be identified using 'mutational signatures'. However, these patterns are difficult to reliably infer from exome sequencing. Additionally, as mutational signatures are a historical record of mutagenic processes, this limits their utility in describing the current status of a tumour. METHODS:We apply two methods for characterising homologous recombination deficiency in breast cancer to explore the features and heterogeneity associated with this phenotype. We develop a likelihood-based method which leverages small insertions and deletions for high-confidence classification of homologous recombination deficiency for exome-sequenced breast cancers. We then use multinomial elastic net regression modelling to develop a transcriptional signature of heterogeneous homologous recombination deficiency. This signature is then applied to single-cell RNA-sequenced breast cancer cohorts enabling analysis of homologous recombination deficiency heterogeneity and differential patterns of tumour microenvironment interactivity. RESULTS:We demonstrate that the inclusion of indel events, even at low levels, improves homologous recombination deficiency classification. Whilst BRCA-positive homologous recombination deficient samples display strong similarities to those harbouring BRCA1/2 defects, they appear to deviate in microenvironmental features such as hypoxic signalling. We then present a 228-gene transcriptional signature which simultaneously characterises homologous recombination deficiency and BRCA1/2-defect status, and is associated with PARP inhibitor response. Finally, we show that this signature is applicable to single-cell transcriptomics data and predict that these cells present a distinct milieu of interactions with their microenvironment compared to their homologous recombination proficient counterparts, typified by a decreased cancer cell response to TNFα signalling. CONCLUSIONS:We apply multi-scale approaches to characterise homologous recombination deficiency in breast cancer through the development of mutational and transcriptional signatures. We demonstrate how indels can improve homologous recombination deficiency classification in exome-sequenced breast cancers. Additionally, we demonstrate the heterogeneity of homologous recombination deficiency, especially in relation to BRCA1/2-defect status, and show that indications of this feature can be captured at a single-cell level, enabling further investigations into interactions between DNA repair deficient cells and their tumour microenvironment.

Community detection is a crucial task in the field of network analysis. A community is a collection of tightly connected nodes only have sporadic external connections. In many real-world networks, communities naturally overlap, and prior knowledge about them is usually unavailable, such as the number of ground-truth communities in the network. In this work, we present the QOCE (Quadratic Optimization based Clique Expansion), an overlapping community detection method that does not require any prior knowledge. QOCE follows the popular seed set expansion strategy and regards each high-quality maximal clique as the initial seed set. For seed set expansion, QOCE uses a fast short random walk to sample a subgraph from a clique seed set, then adopts a quadratic optimization to approximate the Cheeger cut on the sampled subgraph. Finally, a local minimum of conductance determines the boundary of the community. We extensively evaluate our method by comparing it with four state-of-the-art baseline algorithms on synthetic and real-world networks in various domains and scales. Empirical results demonstrate the competitive performance of our method in terms of detection accuracy and efficiency.

In this paper, we study the effective field theory (EFT) of dark energy for the k-essence model beyond linear order. Using particle-mesh N-body simulations that consistently solve the dark energy evolution on a grid, we find that the next-to-leading order in the EFT expansion, which comprises the terms of the equations of motion that are quadratic in the field variables, gives rise to a new instability in the regime of low speed of sound (high Mach number). We rule out the possibility of a numerical artefact by considering simplified cases in spherically and plane symmetric situations analytically. If the speed of sound vanishes exactly, the non-linear instability makes the evolution singular in finite time, signalling a breakdown of the EFT framework. The case of finite (but small) speed of sound is subtle, and the local singularity could be replaced by some other type of behaviour with strong non-linearities. While an ultraviolet completion may cure the problem in principle, there is no reason why this should be the case in general. As a result, for a large range of the effective speed of sound c_s, a linear treatment is not adequate.

ABSTRACT Tumour immunity is key for the prognosis and treatment of colon adenocarcinoma, but its characterisation remains cumbersome and expensive, requiring sequencing or other complex assays. Detecting tumour-infiltrating lymphocytes in haematoxylin and eosin (H&E) slides of cancer tissue would provide a cost-effective alternative to support clinicians in treatment decisions, but inter- and intra-observer variability can arise even amongst experienced pathologists. Furthermore, the compounded effect of other cells in the tumour microenvironment is challenging to quantify but could yield useful additional biomarkers. We combined RNA sequencing, digital pathology and deep learning through the InceptionV3 architecture to develop a fully automated computer vision model that detects prognostic tumour immunity levels in H&E slides of colon adenocarcinoma with an area under the curve (AUC) of 82%. Amongst tumour infiltrating T cell subsets, we demonstrate that CD8+ effector memory T cell patterns are most recognisable algorithmically with an average AUC of 83%. We subsequently applied nuclear segmentation and classification via HoVer-Net to derive complex cell-cell interaction graphs, which we queried efficiently through a bespoke Neo4J graph database. This uncovered stromal barriers and lymphocyte triplets that could act as structural hallmarks of low immunity tumours with poor prognosis. Our integrated deep learning and graph-based workflow provides evidence for the feasibility of automated detection of complex immune cytotoxicity patterns within H&E-stained colon cancer slides, which could inform new cellular biomarkers and support treatment management of this disease in the future.

In this letter we introduce the non-linear partial differential equation (PDE) $\partial^2_{\tau} \pi \propto (\vec\nabla \pi)^2$ showing a new type of instability. Such equations appear in the effective field theory (EFT) of dark energy for the $k$-essence model as well as in many other theories based on the EFT formalism. We demonstrate the occurrence of instability in the cosmological context using a relativistic $N$-body code, and we study it mathematically in 3+1 dimensions within spherical symmetry. We show that this term dominates for the low speed of sound limit where some important linear terms are suppressed.

Laparoscopic surgery, as a representative minimally invasive surgery (MIS), is an active research area of clinical practice. Automatic surgical phase recognition of laparoscopic videos is a vital task with the potential to improve surgeons' efficiency and has gradually become an integral part of computer-assisted intervention systems in MIS. However, the performance of most methods currently employed for surgical phase recognition is deteriorated by optimization difficulties and inefficient computation, which hinders their large-scale practical implementation. This study proposes an efficient and novel surgical phase recognition method using an attention-based spatial-temporal neural network consisting of a spatial model and a temporal model for accurate recognition by end-to-end training. The former subtly incorporates the attention mechanism to enhance the model's ability to focus on the key regions in video frames and efficiently capture more informative visual features. In the temporal model, we employ independently recurrent long short-term memory (IndyLSTM) and non-local block to extract long-term temporal information of video frames. We evaluated the performance of our method on the publicly available Cholec80 dataset. Our attention-based spatial-temporal neural network purely produces the phase predictions without any post-processing strategies, achieving excellent recognition performance and outperforming other state-of-the-art phase recognition methods.

Despite the high performances achieved using deep learning techniques in biometric systems, the inability to rationalise the decisions reached by such approaches is a significant drawback for the usability and security requirements of many applications. For Facial Biometric Presentation Attack Detection (PAD), deep learning approaches can provide good classification results but cannot answer the questions such as "Why did the system make this decision"? To overcome this limitation, an explainable deep neural architecture for Facial Biometric Presentation Attack Detection is introduced in this paper. Both visual and verbal explanations are produced using the saliency maps from a Grad-CAM approach and the gradient from a Long-Short-Term-Memory (LSTM) network with a modified gate function. These explanations have also been used in the proposed framework as additional information to further improve the classification performance. The proposed framework utilises both spatial and temporal information to help the model focus on anomalous visual characteristics that indicate spoofing attacks. The performance of the proposed approach is evaluated using the CASIA-FA, Replay Attack, MSU-MFSD, and HKBU MARs datasets and indicates the effectiveness of the proposed method for improving performance and producing usable explanations.

Surgical tool detection is a key technology in computer-assisted surgery, and can help surgeons to obtain more comprehensive visual information. Currently, a data shortage problem still exists in surgical tool detection. In addition, some surgical tool detection methods may not strike a good balance between detection accuracy and speed. Given the above problems, in this study a new Cholec80-tool6 dataset was manually annotated, which provided a better validation platform for surgical tool detection methods. We propose an enhanced feature-fusion network (EFFNet) for real-time surgical tool detection. FENet20 is the backbone of the network and performs feature extraction more effectively. EFFNet is the feature-fusion part and performs two rounds of feature fusion to enhance the utilization of low-level and high-level feature information. The latter part of the network contains the weight fusion and predictor responsible for the output of the prediction results. The performance of the proposed method was tested using the ATLAS Dione and Cholec80-tool6 datasets, yielding mean average precision values of 97.0% and 95.0% with 21.6 frames per second, respectively. Its speed met the real-time standard and its accuracy outperformed that of other detection methods.

Surgical tool detection and automatic operation skill assessment (AOSA) have important and extensive application scenarios in minimally invasive surgery (MIS). However, most of the deep learning methods currently used for surgical tool detection cannot achieve a good balance between speed and accuracy. We propose a new real-time detection algorithm for MIS tools, which called depth-wise separable convolutional network with convolutional long short-term memory (DSCNet-CLSTM). The network combines the advantages of the one-stage multi-scale feature maps’ concept in the state-of-art detection methods and the convolutional variant of the LSTM. This combination makes full use of the complementary information of spatial and temporal features learned from the laparoscopic video frames. In addition, we have established AOSA system. By processing the output information of the MIS tool detection algorithm, we can obtain the operation tool usage information in the laparoscopic video. Then, this information is taken as the dataset of AOSA system, which makes it possible to realize AOSA based on convolutional neural network (CNN). The proposed method achieved the mAP values of 100, 90.07 and 89.96% at a speed of 50.0 fps for the Endovis Challenge, ATLAS Dione, and Cholec80-locations datasets, respectively. The AOSA system obtained the mean squared error of 2.281, 0.987, 0.069, and 0.009 for the action timeline, heat map, motion trajectory, and all, respectively. The experimental results prove that the framework can be efficiently trained in an end-to-end manner and improves the algorithm’s detection accuracy and speed. Finally, we verify the feasibility of the designed AOSA system through experiments.

Robot-assisted surgery (RAS) is a type of minimally invasive surgery which is completely different from the traditional surgery. RAS reduces surgeon's fatigue and the number of doctors participating in surgery. At the same time, it causes less pain and has a faster recovery rate. Real-time surgical tools detection is important for computer-assisted surgery because the prerequisite for controlling surgical tools is to know the location of surgical tools. In order to achieve comparable performance, most Convolutional Neural Network (CNN) employed for detecting surgical tools generate a huge number of feature maps from expensive operation, which results in redundant computation and long inference time. In this paper, we propose an efficient and novel CNN architecture which generate ghost feature maps cheaply based on intrinsic feature maps. The proposed detector is more efficient and simpler than the state-of-the-art detectors. We believe the proposed method is the first to generate ghost feature maps for detecting surgical tools. Experimental results show that the proposed method achieves 91.6% mAP on the Cholec80-locations dataset and 100% mAP on the Endovis Challenge dataset with the detection speed of 38.5 fps, and realizes real-time and accurate surgical tools detection in the Laparoscopic surgery video.

Learned lossy image compression has demonstrated impressive progress via end-to-end neural network training. However, this end-to-end training belies the fact that lossy compression is inherently not differentiable, due to the necessity of quantisation. To overcome this difficulty in training, researchers have used various approximations to the quantisation step. However, little work has studied the mechanism of quantisation approximation itself. We address this issue, identifying three gaps arising in the quantisation approximation problem. These gaps are visualised, and show the effect of applying different quantisation approximation methods. Following this analysis, we propose a Soft-STE quantisation approximation method, which closes these gaps and demonstrates better performance than other quantisation approaches on the Kodak dataset.

To enhance surgeons' efficiency and safety of patients, minimally invasive surgery (MIS) is widely used in a variety of clinical surgeries. Real-time surgical tool detection plays an important role in MIS. However, most methods of surgical tool detection may not achieve a good trade-off between detection speed and accuracy. We propose a real-time attention-guided convolutional neural network (CNN) for frame-by-frame detection of surgical tools in MIS videos, which comprises a coarse (CDM) and a refined (RDM) detection modules. The CDM is used to coarsely regress the parameters of locations to get the refined anchors and perform binary classification, which determines whether the anchor is a tool or background. The RDM subtly incorporates the attention mechanism to generate accurate detection results utilizing the refined anchors from CDM. Finally, a light-head module for more efficient surgical tool detection is proposed. The proposed method is compared to eight state-of-the-art detection algorithms using two public (EndoVis Challenge and ATLAS Dione) datasets and a new dataset we introduced (Cholec80-locations), which extends the Cholec80 dataset with spatial annotations of surgical tools. Our approach runs in real-time at 55.5 FPS and achieves 100, 94.05, and 91.65% mAP for the above three datasets, respectively. Our method achieves accurate, fast, and robust detection results by end-to-end training in MIS videos. The results demonstrate the effectiveness and superiority of our method over the eight state-of-the-art methods.

深度学习理论在微创手术视频分析中的应用日趋广泛,在微创手术工具检测与跟踪、微创手术工具存在检测和微创手术流程识别等领域已取得令人瞩目的成果.从长远来看,对微创手术视频内容进行细致分析,不但可以自动识别正在进行的微创手术任务,而且可以用来提醒临床医生注意可能出现的并发症.近年来,随着技术的不断发展,深度学习在微创手术视频分析中的应用已取得很大的进展.首先系统阐述微创手术视频分析的意义、难点和相关技术内容,重点介绍深度学习算法的优势;然后总结近年来深度学习在微创手术工具检测与跟踪、微创手术工具存在检测及微创手术流程识别等领域取得的研究成果,在微创手术视频分析的不同领域基于算法特点进行分类总结,并对不同算法进行比较评价;最后,对微创手术视频分析未来的发展方向进行展望.

To prove the reciprocity of an atmospheric turbulent channel in bidirectional optical transmission systems, we propose a method for measuring the correlation between the fading of instantaneously received signals and establish a mathematical model for analyzing the measurement data. Experiments of bidirectional optical transmission measurements were carried out between two tall buildings separated by 883 m. According to the measured speckle image data, we verified the instantaneous-fading correlation of the channel and analyzed the effect of the normalized received signal fluctuation variance on the correlation coefficient in practical scenarios. It was shown that most of the instantaneous-fading correlation coefficients of optical channels in the two counter-directions were above 0. 85 and even up to 0. 95, which proves that reciprocity can be well maintained for a bidirectional turbulent optical channel. With an increasing fluctuation variance of the normalized received optical signal, the correlation coefficient is slightly descending.

Spatio-temporal information is valuable as a discriminative cue for presentation attack detection, where the temporal texture changes and fine-grained motions (such as eye blinking) can be indicative of some types of spoofing attacks. In this paper, we propose a novel spatio-temporal feature, based on motion history, which can offer an efficient way to encapsulate temporal texture changes. Patterns of motion history are used as primary features followed by secondary feature extraction using Local Binary Patterns and Convolutional Neural Networks, and evaluated using the Replay Attack and CASIA-FASD datasets, demonstrating the effectiveness of the proposed approach.

We propose a Locally-Biased Spectral Approximation (LBSA) approach for identifying all latent members of a local community from very few seed members. To reduce the computation complexity, we first apply a fast random walk, personalized PageRank and heat kernel diffusion to sample a comparatively small subgraph covering almost all potential community members around the seeds. Then starting from a normalized indicator vector of the seeds and by a few steps of either Lanczos iteration or power iteration on the sampled subgraph, a local eigenvector is gained for approximating the eigenvector of the transition matrix with the largest eigenvalue. Elements of this local eigenvector is a relaxed indicator for the affiliation probability of the corresponding nodes to the target community. We conduct extensive experiments on real-world datasets in various domains as well as synthetic datasets. Results show that the proposed method outperforms state-of-the-art local community detection algorithms. To the best of our knowledge, this is the first work to adapt the Lanczos method for local community detection, which is natural and potentially effective. Also, we did the first attempt of using heat kernel as a sampling method instead of detecting communities directly, which is proved empirically to be very efficient and effective.