Sci Rep. 2024 Oct 10;14(1):23780. doi: 10.1038/s41598-024-75168-9.

ABSTRACT

Rapid and accurate diagnosis of acute myeloid leukemia (AML) remains a significant challenge, particularly in the context of myelodysplastic syndrome (MDS) or MDS/myeloproliferative neoplasm with NPM1 mutations. This study introduces an innovative approach using holotomography (HT), a 3D label-free quantitative phase imaging technique, to detect NPM1 mutations. We analyzed a dataset of 2073 HT myeloblast images from 48 individuals, including both NPM1 wild-type and mutated samples, to distinguish subcellular morphological changes associated with NPM1 mutations. Employing a convolutional neural network, we analyzed 3D cell morphology, focusing on refractive index distributions. The machine learning model showed high accuracy, with an area under the receiver operating characteristic curve of 0.9375 and a validation accuracy of 76.0%. Our findings reveal distinct morphological differences between the NPM1 wild-type and mutation at the subcellular level. This study demonstrates the potential of HT combined with deep learning for early, efficient, and cost-effective diagnosis of AML, offering a promising alternative to traditional stepwise genetic testing methods and providing additional assistance in morphological myeloblast discrimination. This approach may revolutionize the diagnostic process in leukemia, facilitating early detection and potentially reducing the reliance on extensive genetic testing.

PMID:39390137 | DOI:10.1038/s41598-024-75168-9

Sci Rep. 2024 Oct 10;14(1):23732. doi: 10.1038/s41598-024-75445-7.

ABSTRACT

We proposed a deep learning method using a convolutional neural network on time-series (TS) images to detect and differentiate affected body parts in people with Parkinson's disease (PD) and freezing of gait (FOG) during 360° turning tasks. The 360° turning task was performed by 90 participants (60 people with PD [30 freezers and 30 nonfreezers] and 30 age-matched older adults (controls) at their preferred speed. The position and acceleration underwent preprocessing. The analysis was expanded from temporal to visual data using TS imaging methods. According to the PD vs. controls classification, the right lower third of the lateral shank (RTIB) on the least affected side (LAS) and the right calcaneus (RHEE) on the LAS were the most relevant body segments in the position and acceleration TS images. The RHEE marker exhibited the highest accuracy in the acceleration TS images. The identified markers for the classification of freezers vs. nonfreezers vs. controls were the left lateral humeral epicondyle (LELB) on the more affected side and the left posterior superior iliac spine (LPSI). The LPSI marker in the acceleration TS images displayed the highest accuracy. This approach could be a useful supplementary tool for determining PD severity and FOG.

PMID:39390087 | DOI:10.1038/s41598-024-75445-7

Sci Rep. 2024 Oct 10;14(1):23669. doi: 10.1038/s41598-024-73674-4.

ABSTRACT

Currently, image recognition based on deep neural networks has become the mainstream direction of research; therefore, significant progress has been made in its application in the field of tea detection. Many deep models exhibit high recognition rates in tea leaves detection. However, deploying these models directly on tea-picking equipment in natural environments is impractical; the extremely high parameters and computational complexity of these models make it challenging to perform real-time tea leaves detection. Meanwhile, lightweight models struggle to achieve competitive detection accuracy; therefore, this paper addresses the issue of computational resource constraints in remote mountain areas and proposes Reconstructed Feature and Dual Distillation (RFDD) to enhance the detection capability of lightweight models for tea leaves. In our method, the Reconstructed Feature selectively masks the feature of the student model based on the spatial attention map of the teacher model; it utilizes a generation block to force the student model to generate the teacher's full feature. The Dual Distillation comprises Decoupled Distillation and Global Distillation. Decoupled Distillation divides the reconstructed feature into foreground and background features based on the Ground-Truth. This compels the student model to allocate different attention to foreground and background, focusing on their critical pixels and channels. However, Decoupled Distillation leads to the loss of relation knowledge between foreground and background pixels. Therefore, we further perform Global Distillation to extract this lost knowledge. Since RFDD only requires loss calculation on feature map, it can be easily applied to various detectors. We conducted experiments on detectors with different frameworks, using a tea dataset collected at the Huangshan Houkui Tea Plantation. The experimental results indicate that, under the guidance of RFDD, the student detectors have achieved performance improvements to varying degrees. For instance, a one-stage detector like RetinaNet (ResNet-50) experienced a 3.14% increase in Average Precision (AP) after RFDD guidance. Similarly, a two-stage model like Faster RCNN (ResNet-50) obtained a 3.53% improvement in AP. This offers promising prospects for lightweight models to efficiently perform real-time tea leaves detection tasks.

PMID:39390063 | DOI:10.1038/s41598-024-73674-4

J Chem Theory Comput. 2024 Oct 10. doi: 10.1021/acs.jctc.4c00816. Online ahead of print.

ABSTRACT

Resolving the intricate details of biological phenomena at the molecular level is fundamentally limited by both length- and time scales that can be probed experimentally. Molecular dynamics (MD) simulations at various scales are powerful tools frequently employed to offer valuable biological insights beyond experimental resolution. However, while it is relatively simple to observe long-lived, stable configurations of, for example, proteins, at the required spatial resolution, simulating the more interesting rare transitions between such states often takes orders of magnitude longer than what is feasible even on the largest supercomputers available today. One common aspect of this challenge is pathway discovery, where the start and end states of a scientific phenomenon are known or can be approximated, but the mechanistic details in between are unknown. Here, we propose a representation-learning-based solution that uses interpolation and extrapolation in an abstract representation space to synthesize potential transition states, which are automatically validated using MD simulations. The new simulations of the synthesized transition states are subsequently incorporated into the representation learning, leading to an iterative framework for targeted path sampling. Our approach is demonstrated by recovering the transition of a RAS-RAF protein domain (CRD) from membrane-free to interacting with the membrane using coarse-grain MD simulations.

PMID:39388723 | DOI:10.1021/acs.jctc.4c00816

Gigascience. 2024 Jan 2;13:giae062. doi: 10.1093/gigascience/giae062.

ABSTRACT

BACKGROUND: Due to a constant evolutionary arms race, archaea and bacteria have evolved an abundance and diversity of immune responses to protect themselves against phages. Since the discovery and application of CRISPR-Cas adaptive immune systems, numerous novel candidates for immune systems have been identified. Previous approaches to identifying these new immune systems rely on hidden Markov model (HMM)-based homolog searches or use labor-intensive and costly wet-lab experiments. To aid in finding and classifying immune systems genomes, we use machine learning to classify already known immune system proteins and discover potential candidates in the genome. Neural networks have shown promising results in classifying and predicting protein functionality in recent years. However, these methods often operate under the closed-world assumption, where it is presumed that all potential outcomes or classes are already known and included in the training dataset. This assumption does not always hold true in real-world scenarios, such as in genomics, where new samples can emerge that were not previously accounted for in the training phase.

RESULTS: In this work, we explore neural networks for immune protein classification, deal with different methods for rejecting unrelated proteins in a genome-wide search, and establish a benchmark. Then, we optimize our approach for accuracy. Based on this, we develop an algorithm called Deepdefense to predict immune cassette classes based on a genome. This design facilitates the differentiation between immune system-related and unrelated proteins by analyzing variations in model-predicted confidence values, aiding in the identification of both known and potentially novel immune system proteins. Finally, we test our approach for detecting immune systems in the genome against an HMM-based method.

CONCLUSIONS: Deepdefense can automatically detect genes and define cassette annotations and classifications using 2 model classifications. This is achieved by creating an optimized deep learning model to annotate immune systems, in combination with calibration methods, and a second model to enable the scanning of an entire genome.

PMID:39388605 | DOI:10.1093/gigascience/giae062

PLoS One. 2024 Oct 10;19(10):e0311720. doi: 10.1371/journal.pone.0311720. eCollection 2024.

ABSTRACT

The malicious use of deepfake videos seriously affects information security and brings great harm to society. Currently, deepfake videos are mainly generated based on deep learning methods, which are difficult to be recognized by the naked eye, therefore, it is of great significance to study accurate and efficient deepfake video detection techniques. Most of the existing detection methods focus on analyzing the discriminative information in a specific feature domain for classification from a local or global perspective. Such detection methods based on a single type feature have certain limitations in practical applications. In this paper, we propose a deepfake detection method with the ability to comprehensively analyze the forgery face features, which integrates features in the space domain, noise domain, and frequency domain, and uses the Inception Transformer to learn the mix of global and local information dynamically. We evaluate the proposed method on the DFDC, Celeb-DF, and FaceForensic++ benchmark datasets. Extensive experiments verify the effectiveness and good generalization of the proposed method. Compared with the optimal model, the proposed method with a small number of parameters does not use pre-training, distillation, or assembly, but still achieves competitive performance. The ablation experiments evaluate the role of each component.

PMID:39388418 | DOI:10.1371/journal.pone.0311720

PLoS One. 2024 Oct 10;19(10):e0311161. doi: 10.1371/journal.pone.0311161. eCollection 2024.

ABSTRACT

People use search engines for various topics and items, from daily essentials to more aspirational and specialized objects. Therefore, search engines have taken over as people's preferred resource. The "How To" prefix has become familiar and widely used in various search styles to find solutions to particular problems. This search allows people to find sequential instructions by providing detailed guidelines to accomplish specific tasks. Categorizing instructional text is also essential for task-oriented learning and creating knowledge bases. This study uses the "How To" articles to determine the multi-label instruction category. We have brought this work with a dataset comprising 11,121 observations from wikiHow, where each record has multiple categories. To find out the multi-label category meticulously, we employ some transformer-based deep neural architectures, such as Generalized Autoregressive Pretraining for Language Understanding (XLNet), Bidirectional Encoder Representation from Transformers (BERT), etc. In our multi-label instruction classification process, we have reckoned our proposed architectures using accuracy and macro f1-score as the performance metrics. This thorough evaluation showed us much about our strategy's strengths and drawbacks. Specifically, our implementation of the XLNet architecture has demonstrated unprecedented performance, achieving an accuracy of 97.30% and micro and macro average scores of 89.02% and 93%, a noteworthy accomplishment in multi-label classification. This high level of accuracy and macro average score is a testament to the effectiveness of the XLNet architecture in our proposed 'InstructNet' approach. By employing a multi-level strategy in our evaluation process, we have gained a more comprehensive knowledge of the effectiveness of our proposed architectures and identified areas for forthcoming improvement and refinement.

PMID:39388407 | DOI:10.1371/journal.pone.0311161

PLoS One. 2024 Oct 10;19(10):e0309172. doi: 10.1371/journal.pone.0309172. eCollection 2024.

ABSTRACT

Automatic pavement disease detection aims to address the inefficiency in practical detection. However, traditional methods heavily rely on low-level image analysis, handcrafted features, and classical classifiers, leading to limited effectiveness and poor generalization in complex scenarios. Although significant progress has been made with deep learning methods, challenges persist in handling high-resolution images and diverse disease types. Therefore, this paper proposes a novel approach based on the lightweight Transformer Patch Labeling Network (LTPLN) to enhance the efficiency of automatic pavement disease detection and overcome the limitations of existing methods. Firstly, the input images undergo histogram equalization preprocessing to enhance image quality. Subsequently, the images are evenly partitioned into small patch blocks, serving as inputs to the enhanced Transformer model. This enhancement strategy involves integrating feature map labels at each layer of the model to reduce computational complexity and enhance model lightweightness. Furthermore, a depthwise separable convolution module is introduced into the Transformer architecture to introduce convolutional bias and reduce the model's dependence on large amounts of data. Finally, an iterative training process utilizing the label distillation strategy based on expectation maximization is employed to update the labels of patch blocks and roughly locate the positions of pavement diseases under weak supervision. Experimental results demonstrate that compared to the baseline model, the proposed enhanced model achieves a reduction of 2.5G Flops computational complexity and a 16% speed improvement on a private pavement disease dataset, with only a 1.2 percentage point decrease in AUC accuracy. Moreover, compared to other mainstream image classification models, this model exhibits more balanced performance on a public dataset, with improved accuracy and speed that better align with the practical requirements of pavement inspection. These findings highlight the significant performance advantages of the LTPLN model in automatic pavement disease detection tasks, making it more efficiently applicable in real-world scenarios.

PMID:39388401 | DOI:10.1371/journal.pone.0309172

IEEE Trans Neural Netw Learn Syst. 2024 Oct 10;PP. doi: 10.1109/TNNLS.2024.3460190. Online ahead of print.

ABSTRACT

Spectral super-resolution (SSR) aims to restore a hyperspectral image (HSI) from a single RGB image, in which deep learning has shown impressive performance. However, the majority of the existing deep-learning-based SSR methods inadequately address the modeling of spatial-spectral features in HSI. That is to say, they only sufficiently capture either the spatial correlations or the spectral self-similarity, which results in a loss of discriminative spatial-spectral features and hence limits the fidelity of the reconstructed HSI. To solve this issue, we propose a novel SSR network dubbed multistage spatial-spectral fusion network (MSFN). From the perspective of network design, we build a multistage Unet-like architecture that differentially captures the multiscale features of HSI both spatialwisely and spectralwisely. It consists of two types of the self-attention mechanism, which enables the proposed network to achieve global modeling of HSI comprehensively. From the perspective of feature alignment, we innovatively design the spatial fusion module (SpatialFM) and spectral fusion module (SpectralFM), aiming to preserve the comprehensively captured spatial correlations and spectral self-similarity. In this manner, the multiscale features can be better fused and the accuracy of reconstructed HSI can be significantly enhanced. Quantitative and qualitative experiments on the two largest SSR datasets (i.e., NTIRE2022 and NTIRE2020) demonstrate that our MSFN outperforms the state-of-the-art SSR methods. The code implementation will be uploaded at https://github.com/Matsuri247/MSFN-for-Spectral-Super-Resolution.

PMID:39388330 | DOI:10.1109/TNNLS.2024.3460190

IEEE Trans Med Imaging. 2024 Oct 10;PP. doi: 10.1109/TMI.2024.3477555. Online ahead of print.

ABSTRACT

Prevalent studies on deep learning-based 3D medical image segmentation capture the continuous variation across 2D slices mainly via convolution, Transformer, inter-slice interaction, and time series models. In this work, via modeling this variation by an ordinary differential equation (ODE), we propose a cross instance query-guided Transformer architecture (CQformer) that leverages features from preceding slices to improve the segmentation performance of subsequent slices. Its key components include a cross-attention mechanism in an ODE formulation, which bridges the features of contiguous 2D slices of the 3D volumetric data. In addition, a regression head is employed to shorten the gap between the bottleneck and the prediction layer. Extensive experiments on 7 datasets with various modalities (CT, MRI) and tasks (organ, tissue, and lesion) demonstrate that CQformer outperforms previous state-of-the-art segmentation algorithms on 6 datasets by 0.44%-2.45%, and achieves the second highest performance of 88.30% on the BTCV dataset. The code will be publicly available after acceptance.

PMID:39388328 | DOI:10.1109/TMI.2024.3477555

IEEE Trans Pattern Anal Mach Intell. 2024 Oct 10;PP. doi: 10.1109/TPAMI.2024.3467709. Online ahead of print.

ABSTRACT

Previous deep learning-based event denoising methods mostly suffer from poor interpretability and difficulty in real-time processing due to their complex architecture designs. In this paper, we propose window-based event denoising, which simultaneously deals with a stack of events while existing element-based denoising focuses on one event each time. Besides, we give the theoretical analysis based on probability distributions in both temporal and spatial domains to improve interpretability. In temporal domain, we use timestamp deviations between processing events and central event to judge the temporal correlation and filter out temporal-irrelevant events. In spatial domain, we choose maximum a posteriori (MAP) to discriminate real-world event and noise and use the learned convolutional sparse coding to optimize the objective function. Based on the theoretical analysis, we build Temporal Window (TW) module and Soft Spatial Feature Embedding (SSFE) module to process temporal and spatial information separately, and construct a novel multi-scale window-based event denoising network, named WedNet. The high denoising accuracy and fast running speed of our WedNet enables us to achieve real-time denoising in complex scenes. Extensive experimental results verify the effectiveness and robustness of our WedNet. Our algorithm can remove event noise effectively and efficiently and improve the performance of downstream tasks.

PMID:39388326 | DOI:10.1109/TPAMI.2024.3467709

Elife. 2024 Oct 10;13:RP98300. doi: 10.7554/eLife.98300.

ABSTRACT

Variant calling is fundamental in bacterial genomics, underpinning the identification of disease transmission clusters, the construction of phylogenetic trees, and antimicrobial resistance detection. This study presents a comprehensive benchmarking of variant calling accuracy in bacterial genomes using Oxford Nanopore Technologies (ONT) sequencing data. We evaluated three ONT basecalling models and both simplex (single-strand) and duplex (dual-strand) read types across 14 diverse bacterial species. Our findings reveal that deep learning-based variant callers, particularly Clair3 and DeepVariant, significantly outperform traditional methods and even exceed the accuracy of Illumina sequencing, especially when applied to ONT's super-high accuracy model. ONT's superior performance is attributed to its ability to overcome Illumina's errors, which often arise from difficulties in aligning reads in repetitive and variant-dense genomic regions. Moreover, the use of high-performing variant callers with ONT's super-high accuracy data mitigates ONT's traditional errors in homopolymers. We also investigated the impact of read depth on variant calling, demonstrating that 10× depth of ONT super-accuracy data can achieve precision and recall comparable to, or better than, full-depth Illumina sequencing. These results underscore the potential of ONT sequencing, combined with advanced variant calling algorithms, to replace traditional short-read sequencing methods in bacterial genomics, particularly in resource-limited settings.

PMID:39388235 | DOI:10.7554/eLife.98300

Elife. 2024 Oct 10;13:RP95160. doi: 10.7554/eLife.95160.

ABSTRACT

Errors in stimulus estimation reveal how stimulus representation changes during cognitive processes. Repulsive bias and minimum variance observed near cardinal axes are well-known error patterns typically associated with visual orientation perception. Recent experiments suggest that these errors continuously evolve during working memory, posing a challenge that neither static sensory models nor traditional memory models can address. Here, we demonstrate that these evolving errors, maintaining characteristic shapes, require network interaction between two distinct modules. Each module fulfills efficient sensory encoding and memory maintenance, which cannot be achieved simultaneously in a single-module network. The sensory module exhibits heterogeneous tuning with strong inhibitory modulation reflecting natural orientation statistics. While the memory module, operating alone, supports homogeneous representation via continuous attractor dynamics, the fully connected network forms discrete attractors with moderate drift speed and nonuniform diffusion processes. Together, our work underscores the significance of sensory-memory interaction in continuously shaping stimulus representation during working memory.

PMID:39388221 | DOI:10.7554/eLife.95160

J Appl Crystallogr. 2024 Aug 19;57(Pt 5):1323-1335. doi: 10.1107/S1600576724006897. eCollection 2024 Oct 1.

ABSTRACT

Ptychography is a powerful computational imaging technique with microscopic imaging capability and adaptability to various specimens. To obtain an imaging result, it requires a phase-retrieval algorithm whose performance directly determines the imaging quality. Recently, deep neural network (DNN)-based phase retrieval has been proposed to improve the imaging quality from the ordinary model-based iterative algorithms. However, the DNN-based methods have some limitations because of the sensitivity to changes in experimental conditions and the difficulty of collecting enough measured specimen images for training the DNN. To overcome these limitations, a ptychographic phase-retrieval algorithm that combines model-based and DNN-based approaches is proposed. This method exploits a DNN-based denoiser to assist an iterative algorithm like ePIE in finding better reconstruction images. This combination of DNN and iterative algorithms allows the measurement model to be explicitly incorporated into the DNN-based approach, improving its robustness to changes in experimental conditions. Furthermore, to circumvent the difficulty of collecting the training data, it is proposed that the DNN-based denoiser be trained without using actual measured specimen images but using a formula-driven supervised approach that systemically generates synthetic images. In experiments using simulation based on a hard X-ray ptychographic measurement system, the imaging capability of the proposed method was evaluated by comparing it with ePIE and rPIE. These results demonstrated that the proposed method was able to reconstruct higher-spatial-resolution images with half the number of iterations required by ePIE and rPIE, even for data with low illumination intensity. Also, the proposed method was shown to be robust to its hyperparameters. In addition, the proposed method was applied to ptychographic datasets of a Simens star chart and ink toner particles measured at SPring-8 BL24XU, which confirmed that it can successfully reconstruct images from measurement scans with a lower overlap ratio of the illumination regions than is required by ePIE and rPIE.

PMID:39387085 | PMC:PMC11460392 | DOI:10.1107/S1600576724006897

Proc (IEEE Int Conf Healthc Inform). 2024 Jun;2024:177-182. doi: 10.1109/ichi61247.2024.00030. Epub 2024 Aug 22.

ABSTRACT

The imputation of missing values (IMV) in electronic health records tabular data is crucial to enable machine learning for patient-specific predictive modeling. While IMV methods are developed in biostatistics and recently in machine learning, deep learning-based solutions have shown limited success in learning tabular data. This paper proposes a novel attention-based missing value imputation framework that learns to reconstruct data with missing values leveraging between-feature (self-attention) or between-sample attentions. We adopt data manipulation methods used in contrastive learning to improve the generalization of the trained imputation model. The proposed self-attention imputation method outperforms state-of-the-art statistical and machine learning-based (decision-tree) imputation methods, reducing the normalized root mean squared error by 18.4% to 74.7% on five tabular data sets and 52.6% to 82.6% on two electronic health records data sets. The proposed attention-based missing value imputation method shows superior performance across a wide range of missingness (10% to 50%) when the values are missing completely at random.

PMID:39387063 | PMC:PMC11463999 | DOI:10.1109/ichi61247.2024.00030

Health Data Sci. 2024 Oct 9;4:0182. doi: 10.34133/hds.0182. eCollection 2024.

ABSTRACT

Importance: Heart sound auscultation is a routinely used physical examination in clinical practice to identify potential cardiac abnormalities. However, accurate interpretation of heart sounds requires specialized training and experience, which limits its generalizability. Deep learning, a subset of machine learning, involves training artificial neural networks to learn from large datasets and perform complex tasks with intricate patterns. Over the past decade, deep learning has been successfully applied to heart sound analysis, achieving remarkable results and accumulating substantial heart sound data for model training. Although several reviews have summarized deep learning algorithms for heart sound analysis, there is a lack of comprehensive summaries regarding the available heart sound data and the clinical applications. Highlights: This review will compile the commonly used heart sound datasets, introduce the fundamentals and state-of-the-art techniques in heart sound analysis and deep learning, and summarize the current applications of deep learning for heart sound analysis, along with their limitations and areas for future improvement. Conclusions: The integration of deep learning into heart sound analysis represents a significant advancement in clinical practice. The growing availability of heart sound datasets and the continuous development of deep learning techniques contribute to the improvement and broader clinical adoption of these models. However, ongoing research is needed to address existing challenges and refine these technologies for broader clinical use.

PMID:39387057 | PMC:PMC11461928 | DOI:10.34133/hds.0182

Heliyon. 2024 Sep 19;10(19):e38104. doi: 10.1016/j.heliyon.2024.e38104. eCollection 2024 Oct 15.

ABSTRACT

Significant strides in deep learning for image recognition have expanded the potential of visual data in assessing damage to reinforced concrete (RC) structures. Our study proposes an automated technique, merging convolutional neural networks (CNNs) and fully convolutional networks (FCNs), to detect, classify, and segment building damage. These deep networks extract RC damage-related features from high-resolution smartphone images (3264 × 2448 pixels), categorized into two groups: damage (exposed reinforcement and spalled concrete) and undamaged area. With a labeled dataset of 2000 images, fine-tuning of network architecture and hyperparameters ensures effective training and testing. Remarkably, we achieve 98.75 % accuracy in damage classification and 95.98 % in segmentation, without overfitting. Both CNNs and FCNs play crucial roles in extracting features, showcasing the adaptability of deep learning. Our promising results validate the potential of these techniques for inspectors, providing an effective means to assess the severity of identified damage in image-based evaluations.

PMID:39386784 | PMC:PMC11462242 | DOI:10.1016/j.heliyon.2024.e38104

bioRxiv [Preprint]. 2024 Sep 26:2024.09.25.614955. doi: 10.1101/2024.09.25.614955.

ABSTRACT

We present an easy-to-use, nonlinear filter for effective background identification in fluorescence microscopy images with dense and low-contrast foreground. The pixel-wise filtering is based on comparison of the pixel intensity with the mean intensity of pixels in its local neighborhood. The pixel is given a background or foreground label depending on whether its intensity is less than or greater than the mean respectively. Multiple labels are generated for the same pixel by computing mean expression values by varying neighborhood size. These labels are accumulated to decide the final pixel label. We demonstrate that the performance of our filter favorably compares with state-of-the-art image processing, machine learning, and deep learning methods. We present three use cases that demonstrate its effectiveness, and also show how it can be used in multiplexed fluorescence imaging contexts and as a denoising step in image segmentation. A fast implementation of the filter is available in Python 3 on GitHub .

PMID:39386682 | PMC:PMC11463662 | DOI:10.1101/2024.09.25.614955

bioRxiv [Preprint]. 2024 Sep 25:2024.09.23.614539. doi: 10.1101/2024.09.23.614539.

ABSTRACT

Recent advances in spatial transcriptomics technologies have led to a growing number of diverse datasets, offering unprecedented opportunities to explore tissue organizations and functions within spatial contexts. However, it remains a significant challenge to effectively integrate and interpret these data, often originating from different samples, technologies, and developmental stages. In this paper, we present INSPIRE, a deep learning method for integrative analyses of multiple spatial transcriptomics datasets to address this challenge. With designs of graph neural networks and an adversarial learning mechanism, INSPIRE enables spatially informed and adaptable integration of data from varying sources. By incorporating non-negative matrix factorization, INSPIRE uncovers interpretable spatial factors with corresponding gene programs, revealing tissue architectures, cell type distributions and biological processes. We demonstrate the capabilities of INSPIRE by applying it to human cortex slices from different samples, mouse brain slices with complementary views, mouse hippocampus and embryo slices generated through different technologies, and spatiotemporal organogenesis atlases containing half a million spatial spots. INSPIRE shows superior performance in identifying detailed biological signals, effectively borrowing information across distinct profiling technologies, and elucidating dynamical changes during embryonic development. Furthermore, we utilize INSPIRE to build 3D models of tissues and whole organisms from multiple slices, demonstrating its power and versatility.

PMID:39386646 | PMC:PMC11463460 | DOI:10.1101/2024.09.23.614539

bioRxiv [Preprint]. 2024 Sep 25:2024.09.25.614868. doi: 10.1101/2024.09.25.614868.

ABSTRACT

Modeling the conformational heterogeneity of protein-small molecule systems is an outstanding challenge. We reasoned that while residue level descriptions of biomolecules are efficient for de novo structure prediction, for probing heterogeneity of interactions with small molecules in the folded state an entirely atomic level description could have advantages in speed and generality. We developed a graph neural network called ChemNet trained to recapitulate correct atomic positions from partially corrupted input structures from the Cambridge Structural Database and the Protein Data Bank; the nodes of the graph are the atoms in the system. ChemNet accurately generates structures of diverse organic small molecules given knowledge of their atom composition and bonding, and given a description of the larger protein context, and builds up structures of small molecules and protein side chains for protein-small molecule docking. Because ChemNet is rapid and stochastic, ensembles of predictions can be readily generated to map conformational heterogeneity. In enzyme design efforts described here and elsewhere, we find that using ChemNet to assess the accuracy and pre-organization of the designed active sites results in higher success rates and higher activities; we obtain a preorganized retroaldolase with a k cat / K M of 11000 M -1 min - 1 , considerably higher than any pre-deep learning design for this reaction. We anticipate that ChemNet will be widely useful for rapidly generating conformational ensembles of small molecule and small molecule-protein systems, and for designing higher activity preorganized enzymes.

PMID:39386615 | PMC:PMC11463446 | DOI:10.1101/2024.09.25.614868