Multistage feature fusion knowledge distillation.

Generally, the recognition performance of lightweight models is often lower than that of large models. Knowledge distillation, by teaching a student model using a teacher model, can further enhance the recognition accuracy of lightweight models. In this paper, we approach knowledge distillation from the perspective of intermediate feature-level knowledge distillation. We combine a cross-stage feature fusion symmetric framework, an attention mechanism to enhance the fused features, and a contrastive loss function for teacher and student models at the same stage to comprehensively implement a multistage feature fusion knowledge distillation method. This approach addresses the problem of significant differences in the intermediate feature distributions between teacher and student models, making it difficult to effectively learn implicit knowledge and thus improving the recognition accuracy of the student model. Compared to existing knowledge distillation methods, our method performs at a superior level. On the CIFAR100 dataset, it boosts the recognition accuracy of ResNet20 from 69.06% to 71.34%, and on the TinyImagenet dataset, it increases the recognition accuracy of ResNet18 from 66.54% to 68.03%, demonstrating the effectiveness and generalizability of our approach. Furthermore, there is room for further optimization of the overall distillation structure and feature extraction methods in this approach, which requires further research and exploration.

Combining convolutional neural networks and self-attention for fundus diseases identification.

Early detection of lesions is of great significance for treating fundus diseases. Fundus photography is an effective and convenient screening technique by which common fundus diseases can be detected. In this study, we use color fundus images to distinguish among multiple fundus diseases. Existing research on fundus disease classification has achieved some success through deep learning techniques, but there is still much room for improvement in model evaluation metrics using only deep convolutional neural network (CNN) architectures with limited global modeling ability; the simultaneous diagnosis of multiple fundus diseases still faces great challenges. Therefore, given that the self-attention (SA) model with a global receptive field may have robust global-level feature modeling ability, we propose a multistage fundus image classification model MBSaNet which combines CNN and SA mechanism. The convolution block extracts the local information of the fundus image, and the SA module further captures the complex relationships between different spatial positions, thereby directly detecting one or more fundus diseases in retinal fundus image. In the initial stage of feature extraction, we propose a multiscale feature fusion stem, which uses convolutional kernels of different scales to extract low-level features of the input image and fuse them to improve recognition accuracy. The training and testing were performed based on the ODIR-5k dataset. The experimental results show that MBSaNet achieves state-of-the-art performance with fewer parameters. The wide range of diseases and different fundus image collection conditions confirmed the applicability of MBSaNet.

Multiplex epigenome editing of MECP2 to rescue Rett syndrome neurons.

Rett syndrome (RTT) is an X-linked neurodevelopmental disorder caused by loss-of-function heterozygous mutations of methyl CpG-binding protein 2 (MECP2) on the X chromosome in young females. Reactivation of the silent wild-type MECP2 allele from the inactive X chromosome (Xi) represents a promising therapeutic opportunity for female patients with RTT. Here, we applied a multiplex epigenome editing approach to reactivate MECP2 from Xi in RTT human embryonic stem cells (hESCs) and derived neurons. Demethylation of the MECP2 promoter by dCas9-Tet1 with target single-guide RNA reactivated MECP2 from Xi in RTT hESCs without detectable off-target effects at the transcriptional level. Neurons derived from methylation-edited RTT hESCs maintained MECP2 reactivation and reversed the smaller soma size and electrophysiological abnormalities, two hallmarks of RTT. In RTT neurons, insulation of the methylation-edited MECP2 locus by dCpf1-CTCF (a catalytically dead Cpf1 fused with CCCTC-binding factor) with target CRISPR RNA enhanced MECP2 reactivation and rescued RTT-related neuronal defects, providing a proof-of-concept study for epigenome editing to treat RTT and potentially other dominant X-linked diseases.

Cervical Cell Image Classification-Based Knowledge Distillation.

Current deep-learning-based cervical cell classification methods suffer from parameter redundancy and poor model generalization performance, which creates challenges for the intelligent classification of cervical cytology smear images. In this paper, we establish a method for such classification that combines transfer learning and knowledge distillation. This new method not only transfers common features between different source domain data, but also realizes model-to-model knowledge transfer using the unnormalized probability output between models as knowledge. A multi-exit classification network is then introduced as the student network, where a global context module is embedded in each exit branch. A self-distillation method is then proposed to fuse contextual information; deep classifiers in the student network guide shallow classifiers to learn, and multiple classifier outputs are fused using an average integration strategy to form a classifier with strong generalization performance. The experimental results show that the developed method achieves good results using the SIPaKMeD dataset. The accuracy, sensitivity, specificity, and F-measure of the five classifications are 98.52%, 98.53%, 98.68%, 98.59%, respectively. The effectiveness of the method is further verified on a natural image dataset.

In situ cell-type-specific cell-surface proteomic profiling in mice.

Cell-surface proteins (CSPs) mediate intercellular communication throughout the lives of multicellular organisms. However, there are no generalizable methods for quantitative CSP profiling in specific cell types in vertebrate tissues. Here, we present in situ cell-surface proteome extraction by extracellular labeling (iPEEL), a proximity labeling method in mice that enables spatiotemporally precise labeling of cell-surface proteomes in a cell-type-specific environment in native tissues for discovery proteomics. Applying iPEEL to developing and mature cerebellar Purkinje cells revealed differential enrichment in CSPs with post-translational protein processing and synaptic functions in the developing and mature cell-surface proteomes, respectively. A proteome-instructed in vivo loss-of-function screen identified a critical, multifaceted role for Armh4 in Purkinje cell dendrite morphogenesis. Armh4 overexpression also disrupts dendrite morphogenesis; this effect requires its conserved cytoplasmic domain and is augmented by disrupting its endocytosis. Our results highlight the utility of CSP profiling in native mammalian tissues for identifying regulators of cell-surface signaling.

Cervical Cell/Clumps Detection in Cytology Images Using Transfer Learning.

Cervical cancer is one of the most common and deadliest cancers among women and poses a serious health risk. Automated screening and diagnosis of cervical cancer will help improve the accuracy of cervical cell screening. In recent years, there have been many studies conducted using deep learning methods for automatic cervical cancer screening and diagnosis. Deep-learning-based Convolutional Neural Network (CNN) models require large amounts of data for training, but large cervical cell datasets with annotations are difficult to obtain. Some studies have used transfer learning approaches to handle this problem. However, such studies used the same transfer learning method that is the backbone network initialization by the ImageNet pre-trained model in two different types of tasks, the detection and classification of cervical cell/clumps. Considering the differences between detection and classification tasks, this study proposes the use of COCO pre-trained models when using deep learning methods for cervical cell/clumps detection tasks to better handle limited data set problem at training time. To further improve the model detection performance, based on transfer learning, we conducted multi-scale training according to the actual situation of the dataset. Considering the effect of bounding box loss on the precision of cervical cell/clumps detection, we analyzed the effects of different bounding box losses on the detection performance of the model and demonstrated that using a loss function consistent with the type of pre-trained model can help improve the model performance. We analyzed the effect of mean and std of different datasets on the performance of the model. It was demonstrated that the detection performance was optimal when using the mean and std of the cervical cell dataset used in the current study. Ultimately, based on backbone Resnet50, the mean Average Precision (mAP) of the network model is 61.6% and Average Recall (AR) is 87.7%. Compared to the current values of 48.8% and 64.0% in the used dataset, the model detection performance is significantly improved by 12.8% and 23.7%, respectively.

Transcription factor Acj6 controls dendrite targeting via a combinatorial cell-surface code.

Transcription factors specify the fate and connectivity of developing neurons. We investigate how a lineage-specific transcription factor, Acj6, controls the precise dendrite targeting of Drosophila olfactory projection neurons (PNs) by regulating the expression of cell-surface proteins. Quantitative cell-surface proteomic profiling of wild-type and acj6 mutant PNs in intact developing brains, and a proteome-informed genetic screen identified PN surface proteins that execute Acj6-regulated wiring decisions. These include canonical cell adhesion molecules and proteins previously not associated with wiring, such as Piezo, whose mechanosensitive ion channel activity is dispensable for its function in PN dendrite targeting. Comprehensive genetic analyses revealed that Acj6 employs unique sets of cell-surface proteins in different PN types for dendrite targeting. Combined expression of Acj6 wiring executors rescued acj6 mutant phenotypes with higher efficacy and breadth than expression of individual executors. Thus, Acj6 controls wiring specificity of different neuron types by specifying distinct combinatorial expression of cell-surface executors.

Reciprocal repulsions instruct the precise assembly of parallel hippocampal networks.

Mammalian medial and lateral hippocampal networks preferentially process spatial- and object-related information, respectively. However, the mechanisms underlying the assembly of such parallel networks during development remain largely unknown. Our study shows that, in mice, complementary expression of cell surface molecules teneurin-3 (Ten3) and latrophilin-2 (Lphn2) in the medial and lateral hippocampal networks, respectively, guides the precise assembly of CA1-to-subiculum connections in both networks. In the medial network, Ten3-expressing (Ten3+) CA1 axons are repelled by target-derived Lphn2, revealing that Lphn2- and Ten3-mediated heterophilic repulsion and Ten3-mediated homophilic attraction cooperate to control precise target selection of CA1 axons. In the lateral network, Lphn2-expressing (Lphn2+) CA1 axons are confined to Lphn2+ targets via repulsion from Ten3+ targets. Our findings demonstrate that assembly of parallel hippocampal networks follows a "Ten3→Ten3, Lphn2→Lphn2" rule instructed by reciprocal repulsions.

Temporal evolution of single-cell transcriptomes of Drosophila olfactory projection neurons.

Neurons undergo substantial morphological and functional changes during development to form precise synaptic connections and acquire specific physiological properties. What are the underlying transcriptomic bases? Here, we obtained the single-cell transcriptomes of Drosophila olfactory projection neurons (PNs) at four developmental stages. We decoded the identity of 21 transcriptomic clusters corresponding to 20 PN types and developed methods to match transcriptomic clusters representing the same PN type across development. We discovered that PN transcriptomes reflect unique biological processes unfolding at each stage-neurite growth and pruning during metamorphosis at an early pupal stage; peaked transcriptomic diversity during olfactory circuit assembly at mid-pupal stages; and neuronal signaling in adults. At early developmental stages, PN types with adjacent birth order share similar transcriptomes. Together, our work reveals principles of cellular diversity during brain development and provides a resource for future studies of neural development in PNs and other neuronal types.

Single-Cell Transcriptomes Reveal Diverse Regulatory Strategies for Olfactory Receptor Expression and Axon Targeting.

The regulatory mechanisms by which neurons coordinate their physiology and connectivity are not well understood. The Drosophila olfactory receptor neurons (ORNs) provide an excellent system to investigate this question. Each ORN type expresses a unique olfactory receptor, or a combination thereof, and sends their axons to a stereotyped glomerulus. Using single-cell RNA sequencing, we identified 33 transcriptomic clusters for ORNs and mapped 20 to their glomerular types, demonstrating that transcriptomic clusters correspond well with anatomically and physiologically defined ORN types. Each ORN type expresses hundreds of transcription factors. Transcriptome-instructed genetic analyses revealed that (1) one broadly expressed transcription factor (Acj6) only regulates olfactory receptor expression in one ORN type and only wiring specificity in another type, (2) one type-restricted transcription factor (Forkhead) only regulates receptor expression, and (3) another type-restricted transcription factor (Unplugged) regulates both events. Thus, ORNs utilize diverse strategies and complex regulatory networks to coordinate their physiology and connectivity.

Cell-Surface Proteomic Profiling in the Fly Brain Uncovers Wiring Regulators.

Molecular interactions at the cellular interface mediate organized assembly of single cells into tissues and, thus, govern the development and physiology of multicellular organisms. Here, we developed a cell-type-specific, spatiotemporally resolved approach to profile cell-surface proteomes in intact tissues. Quantitative profiling of cell-surface proteomes of Drosophila olfactory projection neurons (PNs) in pupae and adults revealed global downregulation of wiring molecules and upregulation of synaptic molecules in the transition from developing to mature PNs. A proteome-instructed in vivo screen identified 20 cell-surface molecules regulating neural circuit assembly, many of which belong to evolutionarily conserved protein families not previously linked to neural development. Genetic analysis further revealed that the lipoprotein receptor LRP1 cell-autonomously controls PN dendrite targeting, contributing to the formation of a precise olfactory map. These findings highlight the power of temporally resolved in situ cell-surface proteomic profiling in discovering regulators of brain wiring.

Transsynaptic Fish-lips signaling prevents misconnections between nonsynaptic partner olfactory neurons.

Our understanding of the mechanisms of neural circuit assembly is far from complete. Identification of wiring molecules with novel mechanisms of action will provide insights into how complex and heterogeneous neural circuits assemble during development. In the Drosophila olfactory system, 50 classes of olfactory receptor neurons (ORNs) make precise synaptic connections with 50 classes of partner projection neurons (PNs). Here, we performed an RNA interference screen for cell surface molecules and identified the leucine-rich repeat-containing transmembrane protein known as Fish-lips (Fili) as a novel wiring molecule in the assembly of the Drosophila olfactory circuit. Fili contributes to the precise axon and dendrite targeting of a small subset of ORN and PN classes, respectively. Cell-type-specific expression and genetic analyses suggest that Fili sends a transsynaptic repulsive signal to neurites of nonpartner classes that prevents their targeting to inappropriate glomeruli in the antennal lobe.

Stepwise wiring of the Drosophila olfactory map requires specific Plexin B levels.

The precise assembly of a neural circuit involves many consecutive steps. The conflict between a limited number of wiring molecules and the complexity of the neural network impels each molecule to execute multiple functions at different steps. Here, we examined the cell-type specific distribution of endogenous levels of axon guidance receptor Plexin B (PlexB) in the developing antennal lobe, the first olfactory processing center in Drosophila. We found that different classes of olfactory receptor neurons (ORNs) express PlexB at different levels in two wiring steps - axonal trajectory choice and subsequent target selection. In line with its temporally distinct patterns, the proper levels of PlexB control both steps in succession. Genetic interactions further revealed that the effect of high-level PlexB is antagonized by its canonical partner Sema2b. Thus, PlexB plays a multifaceted role in instructing the assembly of the Drosophila olfactory circuit through temporally-regulated expression patterns and expression level-dependent effects.

Rescue of Fragile X Syndrome Neurons by DNA Methylation Editing of the FMR1 Gene.

Fragile X syndrome (FXS), the most common genetic form of intellectual disability in males, is caused by silencing of the FMR1 gene associated with hypermethylation of the CGG expansion mutation in the 5' UTR of FMR1 in FXS patients. Here, we applied recently developed DNA methylation editing tools to reverse this hypermethylation event. Targeted demethylation of the CGG expansion by dCas9-Tet1/single guide RNA (sgRNA) switched the heterochromatin status of the upstream FMR1 promoter to an active chromatin state, restoring a persistent expression of FMR1 in FXS iPSCs. Neurons derived from methylation-edited FXS iPSCs rescued the electrophysiological abnormalities and restored a wild-type phenotype upon the mutant neurons. FMR1 expression in edited neurons was maintained in vivo after engrafting into the mouse brain. Finally, demethylation of the CGG repeats in post-mitotic FXS neurons also reactivated FMR1. Our data establish that demethylation of the CGG expansion is sufficient for FMR1 reactivation, suggesting potential therapeutic strategies for FXS.

Corrigendum: RNA m6A methylation regulates the ultraviolet-induced DNA damage response.

This corrects the article DOI: 10.1038/nature21671.

RNA m6A methylation regulates the ultraviolet-induced DNA damage response.

Cell proliferation and survival require the faithful maintenance and propagation of genetic information, which are threatened by the ubiquitous sources of DNA damage present intracellularly and in the external environment. A system of DNA repair, called the DNA damage response, detects and repairs damaged DNA and prevents cell division until the repair is complete. Here we report that methylation at the 6 position of adenosine (m6A) in RNA is rapidly (within 2 min) and transiently induced at DNA damage sites in response to ultraviolet irradiation. This modification occurs on numerous poly(A)+ transcripts and is regulated by the methyltransferase METTL3 (methyltransferase-like 3) and the demethylase FTO (fat mass and obesity-associated protein). In the absence of METTL3 catalytic activity, cells showed delayed repair of ultraviolet-induced cyclobutane pyrimidine adducts and elevated sensitivity to ultraviolet, demonstrating the importance of m6A in the ultraviolet-responsive DNA damage response. Multiple DNA polymerases are involved in the ultraviolet response, some of which resynthesize DNA after the lesion has been excised by the nucleotide excision repair pathway, while others participate in trans-lesion synthesis to allow replication past damaged lesions in S phase. DNA polymerase κ (Pol κ), which has been implicated in both nucleotide excision repair and trans-lesion synthesis, required the catalytic activity of METTL3 for immediate localization to ultraviolet-induced DNA damage sites. Importantly, Pol κ overexpression qualitatively suppressed the cyclobutane pyrimidine removal defect associated with METTL3 loss. Thus, we have uncovered a novel function for RNA m6A modification in the ultraviolet-induced DNA damage response, and our findings collectively support a model in which m6A RNA serves as a beacon for the selective, rapid recruitment of Pol κ to damage sites to facilitate repair and cell survival.