Aim Splicing factor proline and glutamine rich (SFPQ) is an RNA–DNA binding protein that is dysregulated in Alzheimer's disease and frontotemporal dementia. Dysregulation of SFPQ, specifically increased intron retention and nuclear depletion, has been linked to several genetic subtypes of amyotrophic lateral sclerosis (ALS), suggesting that SFPQ pathology may be a common feature of this heterogeneous disease. Our study aimed to investigate this hypothesis by providing the first comprehensive assessment of SFPQ pathology in large ALS case–control cohorts. Methods We examined SFPQ at the RNA, protein and DNA levels. SFPQ RNA expression and intron retention were examined using RNA‐sequencing and quantitative PCR. SFPQ protein expression was assessed by immunoblotting and immunofluorescent staining. At the DNA level, SFPQ was examined for genetic variation novel to ALS patients. Results At the RNA level, retention of SFPQ intron nine was significantly increased in ALS patients' motor cortex. In addition, SFPQ RNA expression was significantly reduced in the central nervous system, but not blood, of patients. At the protein level, neither nuclear depletion nor reduced expression of SFPQ was found to be a consistent feature of spinal motor neurons. However, SFPQ‐positive ubiquitinated protein aggregates were observed in patients' spinal motor neurons. At the DNA level, our genetic screen identified two novel and two rare SFPQ sequence variants not previously reported in the literature. Conclusions Our findings confirm dysregulation of SFPQ as a pathological feature of the central nervous system of ALS patients and indicate that investigation of the functional consequences of this pathology will provide insight into ALS biology.

A major gap between few-shot and many-shot learning is the data distribution empirically observed by the model during training. In few-shot learning, the learned model can easily become over-fitted based on the biased distribution formed by only a few training examples, while the ground-truth data distribution is more accurately uncovered in many-shot learning to learn a well-generalized model. In this paper, we propose to calibrate the distribution of these few-sample classes to be more unbiased to alleviate such an over-fitting problem. The distribution calibration is achieved by transferring statistics from the classes with sufficient examples to those few-sample classes. After calibration, an adequate number of examples can be sampled from the calibrated distribution to expand the inputs to the classifier. Specifically, we assume every dimension in the feature representation follows a Gaussian distribution so that the mean and the variance of the distribution can borrow from that of similar classes whose statistics are better estimated with an adequate number of samples. Extensive experiments on three datasets, miniImageNet, tieredImageNet, and CUB, show that a simple linear classifier trained using the features sampled from our calibrated distribution can outperform the state-of-the-art accuracy by a large margin. Besides the favorable performance, the proposed method also exhibits high flexibility by showing consistent accuracy improvement when it is built on top of any off-the-shelf pretrained feature extractors and classification models without extra learnable parameters. The visualization of these generated features demonstrates that our calibrated distribution is an accurate estimation thus the generalization ability gain is convincing. We also establish a generalization error bound for the proposed distribution-calibration-based few-shot learning, which consists of the distribution assumption error, the distribution approximation error, and the estimation error. This generalization error bound theoretically justifies the effectiveness of the proposed method.

Background: Splicing factor proline and glutamine rich (SFPQ, also known as polypyrimidine tract-binding protein-associated-splicing factor, PSF) is a RNA-DNA binding protein with roles in key cellular pathways such as DNA transcription and repair, RNA processing and paraspeckle formation. Dysregulation of SFPQ is emerging as a common pathological feature of multiple neurodegenerative diseases including amyotrophic lateral sclerosis (ALS). Increased retention of SFPQ intron nine and nuclear loss of the protein have been linked to multiple genetic subtypes of ALS. Consequently, SFPQ dysregulation has been hypothesised to be a common pathological feature of this highly heterogeneous disease. Methods: This study provides a comprehensive assessment of SFPQ pathology in large ALS patient cohorts. SFPQ gene expression and intron nine retention were examined in multiple neuroanatomical regions and blood from ALS patients and control individuals using RNA sequencing (RNA-Seq) and quantitative PCR (RT-qPCR). SFPQ protein levels were assessed by immunoblotting of patient and control motor cortex and SFPQ expression pattern was examined by immunofluorescent staining of patient and control spinal cord sections. Finally, whole-genome sequencing data from a large cohort of sporadic ALS patients was analysed for genetic variation in SFPQ. Results: SFPQ intron nine retention was significantly increased in ALS patient motor cortex. Total SFPQ mRNA expression was significantly downregulated in ALS patient motor cortex but not ALS patient blood, indicating tissue specific SFPQ dysregulation. At the protein level, nuclear expression of SFPQ in both control and patient spinal motor neurons was highly variable and nuclear depletion of SFPQ was not a consistent feature in our ALS cohort. However, we did observe SFPQ-positive cytoplasmic ubiquitinated protein aggregates in ALS spinal motor neurons. In addition, our genetic screen of ALS patients identified two novel, and two rare sequence variants in SFPQ not previously reported in ALS. Conclusions: This study shows that dysregulation of SFPQ is a feature of ALS patient central nervous system tissue. These findings confirm SFPQ pathology as a feature of ALS and indicate that investigations into the functional consequences of this pathology will provide insight into the biology of ALS.

A major gap between few-shot and many-shot learning is the data distribution empirically observed by the model during training. In few-shot learning, the learned model can easily become over-fitted based on the biased distribution formed by only a few training examples, while the ground-truth data distribution is more accurately uncovered in many-shot learning to learn a well-generalized model. In this paper, we propose to calibrate the distribution of these few-sample classes to be more unbiased to alleviate such an over-fitting problem. The distribution calibration is achieved by transferring statistics from the classes with sufficient examples to those few-sample classes. After calibration, an adequate number of examples can be sampled from the calibrated distribution to expand the inputs to the classifier. Extensive experiments on three datasets, miniImageNet, tieredImageNet, and CUB, show that a simple linear classifier trained using the features sampled from our calibrated distribution can outperform the state-of-the-art accuracy by a large margin. We also establish a generalization error bound for the proposed distribution-calibration-based few-shot learning, which consists of the distribution assumption error, the distribution approximation error, and the estimation error. This generalization error bound theoretically justifies the effectiveness of the proposed method.

ObjectiveSince the first report of CHCHD10 gene mutations in amyotrophiclateral sclerosis (ALS)/frontotemporaldementia (FTD) patients, genetic variation in CHCHD10 has been inconsistently linked to disease. A pathological assessment of the CHCHD10 protein in patient neuronal tissue also remains to be reported. We sought to characterise the genetic and pathological contribution of CHCHD10 to ALS/FTD in Australia.MethodsWhole-exome and whole-genome sequencing data from 81 familial and 635 sporadic ALS, and 108 sporadic FTD cases, were assessed for genetic variation in CHCHD10. CHCHD10 protein expression was characterised by immunohistochemistry, immunofluorescence and western blotting in control, ALS and/or FTD postmortem tissues and further in a transgenic mouse model of TAR DNA-binding protein 43 (TDP-43) pathology.ResultsNo causal, novel or disease-associated variants in CHCHD10 were identified in Australian ALS and/or FTD patients. In human brain and spinal cord tissues, CHCHD10 was specifically expressed in neurons. A significant decrease in CHCHD10 protein level was observed in ALS patient spinal cord and FTD patient frontal cortex. In a TDP-43 mouse model with a regulatable nuclear localisation signal (rNLS TDP-43 mouse), CHCHD10 protein levels were unaltered at disease onset and early in disease, but were significantly decreased in cortex in mid-stage disease.ConclusionsGenetic variation in CHCHD10 is not a common cause of ALS/FTD in Australia. However, we showed that in humans, CHCHD10 may play a neuron-specific role and a loss of CHCHD10 function may be linked to ALS and/or FTD. Our data from the rNLS TDP-43 transgenic mice suggest that a decrease in CHCHD10 levels is a late event in aberrant TDP-43-induced ALS/FTD pathogenesis.

Label noise is ubiquitous in the era of big data. Deep learning algorithms can easily t the noise and thus cannot generalize well without properly modeling the noise. In this paper, we propose a new perspective on dealing with label noise called "Class2Simi". Speci cally, we transform the training examples with noisy class labels into pairs of examples with noisy similarity labels and propose a deep learning framework to learn robust classi ers directly with the noisy similarity labels. Note that a class label shows the class that an instance belongs to; while a similarity label indicates whether or not two instances belong to the same class. It is worthwhile to perform the transformation: We prove that the noise rate for the noisy similarity labels is lower than that of the noisy class labels, because similarity labels themselves are robust to noise. For example, given two instances, even if both of their class labels are incorrect, their similarity label could be correct. Due to the lower noise rate, Class2Simi achieves remarkably be er classi cation accuracy than its baselines that directly deals with the noisy class labels.

No abstract

A similarity label indicates whether two instances belong to the same class while a class label shows the class of the instance. Without class labels, a multi-class classi er could be learned from similarity-labeled pairwise data by meta classi cation learning [Hsu et al., 2019]. However, since the similarity label is less informative than the class label, it is more likely to be noisy. Deep neural networks can easily remember noisy data, leading to over ing in classi cation. In this paper, we propose a method for learning from only noisy-similarity-labeled data. Speci cally, to model the noise, we employ a noise transition matrix to bridge the class-posterior probability between clean and noisy data. We further estimate the transition matrix from only noisy data and build a novel learning system to learn a classi er which can assign noisefree class labels for instances. Moreover, we theoretically justify how our proposed method generalizes for learning classi ers. Experimental results demonstrate the superiority of the proposed method over the state-of-the-art method on benchmarksimulated and real-world noisy-label datasets. * Equal contributions.