Search

Motivation The annotation of cell types from single-cell transcriptomics is essential for understanding the biological identity and functionality of cellular populations. Although manual annotation remains the gold standard, the advent of automatic pipelines has become crucial for scalable, unbiased, and cost-effective annotations. Nonetheless, the effectiveness of these automatic methods, particularly those employing deep learning, significantly depends on the architecture of the classifier and the quality and diversity of the training datasets. Results To address these limitations, we present a Pruning-enabled Gene-Cell Net (PredGCN) incorporating a Coupled Gene-Cell Net (CGCN) to enable representation learning and information storage. PredGCN integrates a Gene Splicing Net (GSN) and a Cell Stratification Net (CSN), employing a pruning operation (PrO) to dynamically tackle the complexity of heterogeneous cell identification. Among them, GSN leverages multiple statistical and hypothesis-driven feature extraction methods to selectively assemble genes with specificity for scRNA-seq data while CSN unifies elements based on diverse region demarcation principles, exploiting the representations from GSN and precise identification from different regional homogeneity perspectives. Furthermore, we develop a multi-objective Pareto pruning operation (Pareto PrO) to expand the dynamic capabilities of CGCN, optimizing the sub-network structure for accurate cell type annotation. Multiple comparison experiments on real scRNA-seq datasets from various species have demonstrated that PredGCN surpasses existing state-of-the-art methods, including its scalability to cross-species datasets. Moreover, PredGCN can uncover unknown cell types and provide functional genomic analysis by quantifying the influence of genes on cell clusters, bringing new insights into cell type identification and characterizing scRNA-seq data from different perspectives. Availability and implementation The source code is available at https://github.com/IrisQi7/PredGCN and test data is available at https://figshare.com/articles/dataset/PredGCN/25251163. [ABSTRACT FROM AUTHOR]

Antarctic sea ice concentration (SIC) prediction at seasonal scale has been documented, but a gap remains at subseasonal scale (1–8 weeks) due to limited understanding of ice‐related physical mechanisms. To overcome this limitation, we developed a deep learning model named Sea Ice Prediction Network (SIPNet) that can predict SIC without the need to account for complex physical processes. Compared to mainstream dynamical models like European Centre for Medium‐Range Weather Forecasts, National Centers for Environmental Prediction, and Seamless System for Prediction and Earth System Research developed at Geophysical Fluid Dynamics Laboratory, as well as a relatively advanced statistical model like the linear Markov model, SIPNet outperforms them all, effectively filling the gap in subseasonal Antarctic SIC prediction capability. SIPNet results indicate that autumn SIC variability contributes the most to sea ice predictability, whereas spring contributes the least. In addition, the Weddell Sea displays the highest sea ice predictability, while predictability is low in the West Pacific. SIPNet can also capture the signal of ENSO and SAM on sea ice. Plain Language Summary: Antarctic sea ice has changed significantly since 2016, leading to a higher demand for sea ice forecasts. However, forecasting Antarctic sea ice has not received enough attention. Limited observations and lack of understanding of ice‐related physical mechanisms result in significant errors in sea ice predictions in dynamical models, particularly at the subseasonal timescale. We utilized a deep‐learning model to predict Antarctic sea ice at this timescale to fill this gap. Results showed that the deep‐learning model performed skillfully 1–8 weeks in advance, with the Weddell Sea being the best‐predicted region and the West Pacific being the worst. Furthermore, our study found that the model significantly outperformed mainstream dynamic models and a conventional statistical model. These findings build a foundation for developing more advanced prediction models at high resolutions for operational applications. Key Points: A deep‐learning model outperforms dynamic models, filling a subseasonal Antarctic sea ice prediction gap by bypassing physical mechanismsUnlike a Markov model that predicts time series of fixed spatial patterns, our model can extract ice synchronized spatiotemporal evolutionsThe deep‐learning model captures climate signals in sea ice, delivering the best prediction in fall season and the Weddell Sea region [ABSTRACT FROM AUTHOR]

This paper introduces versatile filters to construct efficient convolutional neural networks that are widely used in various visual recognition tasks. Considering the demands of efficient deep learning techniques running on cost-effective hardware, a number of methods have been developed to learn compact neural networks. Most of these works aim to slim down filters in different ways, e.g., investigating small, sparse or quantized filters. In contrast, we treat filters from an additive perspective. A series of secondary filters can be derived from a primary filter with the help of binary masks. These secondary filters all inherit in the primary filter without occupying more storage, but once been unfolded in computation they could significantly enhance the capability of the filter by integrating information extracted from different receptive fields. Besides spatial versatile filters, we additionally investigate versatile filters from the channel perspective. Binary masks can be further customized for different primary filters under orthogonal constraints. We conduct theoretical analysis on network complexity and an efficient convolution scheme is introduced. Experimental results on benchmark datasets and neural networks demonstrate that our versatile filters are able to achieve comparable accuracy as that of original filters, but require less memory and computation cost. [ABSTRACT FROM AUTHOR]

Haploinsufficiency, wherein a single allele is not enough to maintain normal functions, can lead to many diseases including cancers and neurodevelopmental disorders. Recently, computational methods for identifying haploinsufficiency have been developed. However, most of those computational methods suffer from study bias, experimental noise and instability, resulting in unsatisfactory identification of haploinsufficient genes. To address those challenges, we propose a deep forest model, called HaForest, to identify haploinsufficient genes. The multiscale scanning is proposed to extract local contextual representations from input features under Linear Discriminant Analysis. After that, the cascade forest structure is applied to obtain the concatenated features directly by integrating decision-tree-based forests. Meanwhile, to exploit the complex dependency structure among haploinsufficient genes, the LightGBM library is embedded into HaForest to reveal the highly expressive features. To validate the effectiveness of our method, we compared it to several computational methods and four deep learning algorithms on five epigenomic data sets. The results reveal that HaForest achieves superior performance over the other algorithms, demonstrating its unique and complementary performance in identifying haploinsufficient genes. The standalone tool is available at https://github.com/yangyn533/HaForest. [ABSTRACT FROM AUTHOR]

The identification of hidden responders is often an essential challenge in precision oncology. A recent attempt based on machine learning has been proposed for classifying aberrant pathway activity from multiomic cancer data. However, we note several critical limitations there, such as high-dimensionality, data sparsity and model performance. Given the central importance and broad impact of precision oncology, we propose nature-inspired deep Ras activation pan-cancer (NatDRAP), a deep neural network (DNN) model, to address those restrictions for the identification of hidden responders. In this study, we develop the nature-inspired deep learning model that integrates bulk RNA sequencing, copy number and mutation data from PanCanAltas to detect pan-cancer Ras pathway activation. In NatDRAP, we propose to synergize the nature-inspired artificial bee colony algorithm with different gradient-based optimizers in one framework for optimizing DNNs in a collaborative manner. Multiple experiments were conducted on 33 different cancer types across PanCanAtlas. The experimental results demonstrate that the proposed NatDRAP can provide superior performance over other benchmark methods with strong robustness towards diagnosing RAS aberrant pathway activity across different cancer types. In addition, gene ontology enrichment and pathological analysis are conducted to reveal novel insights into the RAS aberrant pathway activity identification and characterization. NatDRAP is written in Python and available at https://github.com/lixt314/NatDRAP1. [ABSTRACT FROM AUTHOR]