Search

Deep learning models have been very successful in the application of machine learning methods, often out-performing classical statistical models such as linear regression models or generalized linear models. On the other hand, deep learning models are often criticized for not being explainable nor allowing for variable selection. There are two different ways of dealing with this problem, either we use post-hoc model interpretability methods or we design specific deep learning architectures that allow for (more) easy interpretation and explanation. This paper builds on our previous work on the LocalGLMnet architecture that gives an interpretable deep learning architecture. In the present paper, we show how group LASSO regularization (and other regularization schemes) can be implemented within the LocalGLMnet architecture so that we receive feature sparsity for variable selection. We benchmark our approach with the recently developed LassoNet of Lemhadri et al.

We consider the following question: given information on individual policyholder characteristics, how can we ensure that insurance prices do not discriminate with respect to protected characteristics, such as gender? We address the issues of direct and indirect discrimination, the latter resulting from implicit learning of protected characteristics from nonprotected ones. We provide rigorous mathematical definitions for direct and indirect discrimination, and we introduce a simple formula for discrimination-free pricing, that avoids both direct and indirect discrimination. Our formula works in any statistical model. We demonstrate its application on a health insurance example, using a state-of-the-art generalized linear model and a neural network regression model. An important conclusion is that discrimination-free pricing in general requires collection of policyholders’ discriminatory characteristics, posing potential challenges in relation to policyholder’s privacy concerns., ASTIN Bulletin, 52 (1), ISSN:0515-0361, ISSN:1783-1350

Rapid advances in artificial intelligence (AI) and machine learning are creating products and services with the potential not only to change the environment in which actuaries operate but also to provide new opportunities within actuarial science. These advances are based on a modern approach to designing, fitting and applying neural networks, generally referred to as “Deep Learning.” This paper investigates how actuarial science may adapt and evolve in the coming years to incorporate these new techniques and methodologies. Part 1 of this paper provides background on machine learning and deep learning, as well as an heuristic for where actuaries might benefit from applying these techniques. Part 2 of the paper then surveys emerging applications of AI in actuarial science, with examples from mortality modelling, claims reserving, non-life pricing and telematics. For some of the examples, code has been provided on GitHub so that the interested reader can experiment with these techniques for themselves. Part 2 concludes with an outlook on the potential for actuaries to integrate deep learning into their activities. Finally, a supplementary appendix discusses further resources providing more in-depth background on machine learning and deep learning.

Rapid advances in artificial intelligence (AI) and machine learning are creating products and services with the potential not only to change the environment in which actuaries operate, but also to provide new opportunities within actuarial science. These advances are based on a modern approach to designing, fitting and applying neural networks, generally referred to as “Deep Learning”. This paper investigates how actuarial science may adapt and evolve in the coming years to incorporate these new techniques and methodologies. Part 1 of this paper provides background on machine learning and deep learning, as well as an heuristic for where actuaries might benefit from applying these techniques. Part 2 of the paper then surveys emerging applications of AI in actuarial science, with examples from mortality modelling, claims reserving, non-life pricing and telematics. For some of the examples, code has been provided on GitHub so that the interested reader can experiment with these techniques for themselves. Part 2 concludes with an outlook on the potential for actuaries to integrate deep learning into their activities. Finally, a supplementary appendix discusses further resources providing more in-depth background on machine learning and deep learning.

Deep learning models have gained great popularity in statistical modeling because they lead to very competitive regression models, often outperforming classical statistical models such as generalized linear models. The disadvantage of deep learning models is that their solutions are difficult to interpret and explain, and variable selection is not easily possible because deep learning models solve feature engineering and variable selection internally in a nontransparent way. Inspired by the appealing structure of generalized linear models, we propose a new network architecture that shares similar features as generalized linear models but provides superior predictive power benefiting from the art of representation learning. This new architecture allows for variable selection of tabular data and for interpretation of the calibrated deep learning model, in fact, our approach provides an additive decomposition that can be related to other model interpretability techniques., Scandinavian Actuarial Journal, 2023 (1), ISSN:0346-1238, ISSN:1651-2030

The Lee–Carter (LC) model is a basic approach to forecasting mortality rates of a single population. Although extensions of the LC model to forecasting rates for multiple populations have recently been proposed, the structure of these extended models is hard to justify and the models are often difficult to calibrate, relying on customised optimisation schemes. Based on the paradigm of representation learning, we extend the LCmodel to multiple populations using neural networks, which automatically select an optimal model structure. We fit this model to mortality rates since 1950 for all countries in the Human Mortality Database and observe that the out-of-sample forecasting performance of the model is highly competitive.

Genome sequencing has revealed an increasing number of genetic variations that are associated with neuropsychiatric disorders. Frequently, studies limit their focus to likely gene-disrupting mutations because they are relatively easy to interpret. Missense variants, instead, have often been undervalued. However, some missense variants can be informative for developing a more profound understanding of disease pathogenesis and ultimately targeted therapies. Here we present an example of this by studying a missense variant in a well-known autism spectrum disorder (ASD) causing gene SHANK3. We analyzed Shank3’s in vivo phosphorylation profile and identified S685 as one phosphorylation site where one ASD-linked variant has been reported. Detailed analysis of this variant revealed a novel function of Shank3 in recruiting Abelson interactor 1 (ABI1) and the WAVE complex to the post-synaptic density (PSD), which is critical for synapse and dendritic spine development. This function was found to be independent of Shank3’s other functions such as binding to GKAP and Homer. Introduction of this human ASD mutation into mice resulted in a small subset of phenotypes seen previously in constitutive Shank3 knockout mice, including increased allogrooming, increased social dominance, and reduced pup USV. Together, these findings demonstrate the modularity of Shank3 function in vivo. This modularity further indicates that there is more than one independent pathogenic pathway downstream of Shank3 and correcting a single downstream pathway is unlikely to be sufficient for clear clinical improvement. In addition, this study illustrates the value of deep biological analysis of select missense mutations in elucidating the pathogenesis of neuropsychiatric phenotypes.

The time-series nature of mortality rates lends itself to processing through neural networks that are specialized to deal with sequential data, such as recurrent and convolutional networks. Although appealing intuitively, a naive implementation of these networks does not lead to enhanced predictive performance. We show how the structure of the Lee Carter model can be generalized, and propose a relatively simple convolutional network model that can be interpreted as a generalization of the Lee Carter model, allowing for its components to be evaluated in familiar terms. The model produces highly accurate forecasts on the Human Mortality Database, and, without further modification, generalizes well to the United States Mortality Database.

A vast and growing literature on explaining deep learning models has emerged. This paper contributes to that literature by introducing a global gradient-based model-agnostic method, which we call Marginal Attribution by Conditioning on Quantiles (MACQ). Our approach is based on analyzing the marginal attribution of predictions (outputs) to individual features (inputs). Specifically, we consider variable importance by fixing (global) output levels, and explaining how features marginally contribute to these fixed global output levels. MACQ can be seen as a marginal attribution counterpart to approaches such as accumulated local effects, which study the sensitivities of outputs by perturbing inputs. Furthermore, MACQ allows us to separate marginal attribution of individual features from interaction effects and to visualize the 3-way relationship between marginal attribution, output level, and feature value., Data Mining and Knowledge Discovery, 36 (4), ISSN:1384-5810, ISSN:1573-756X

Spinocerebellar ataxias (SCAs) are a group of genetically heterogeneous inherited neurodegenerative disorders characterized by progressive ataxia and cerebellar degeneration. Here, we tested if Ronin (Thap11), a polyglutamine-containing protein encoded in a region on human chromosome 16q22.1 that has been genetically linked to SCA4, can be connected with SCA disease in a mouse model. We report that transgenic expression of Ronin in mouse cerebellar Purkinje cells leads not only histopathologically to detrimental loss of Purkinje cells but also phenotypically to the development of severe ataxia as early as 10-12 weeks after birth. Mechanistically, we find that Ronin is part of a protein complex in the cerebellum that is distinct from the one previously found in embryonic stem cells. Importantly, ectopically expressed Ronin raises the protein level of Ataxin-1 (Atxn1), the causative gene of the most common type of SCA, SCA1. Hence, our data provide evidence for a link between Ronin and SCAs, and also suggest that Ronin may be involved in the development of other neurodegenerative diseases.

Spinocerebellar ataxias (SCAs) are a group of genetically heterogeneous inherited neurodegenerative disorders characterized by progressive ataxia and cerebellar degeneration. Here, we used a mouse model to test a possible connection between SCA and Ronin (Thap11), a polyglutamine-containing transcriptional regulator encoded in a region of human chromosome 16q22.1 that has been genetically linked to SCA type 4. We report that transgenic expression of Ronin in mouse cerebellar Purkinje cells leads to detrimental loss of these cells and the development of severe ataxia as early as 10 weeks after birth. Mechanistically, we find that several SCA-causing genes harbor Ronin DNA-binding motifs and are transcriptionally deregulated in transgenic animals. In addition, ectopic expression of Ronin in embryonic stem cells significantly increases the protein level of Ataxin-1, the protein encoded by Atxn1, alterations of which cause SCA type 1. This increase is also seen in the cerebellum of transgenic animals, although the latter was not statistically significant. Hence, our data provide evidence for a link between Ronin and SCAs, and suggest that Ronin may be involved in the development of other neurodegenerative diseases., Summary: Ronin is a polyglutamine protein encoded in a region of human chromosome 16q22.1 linked to spinocerebellar ataxia type 4. Overexpression of Ronin in mouse cerebellar Purkinje cells leads to their loss and ataxia.

In this paper we investigate the role of the Head of the Actuarial Function (HAF) and its implementation in the South African insurance market, two years after the passing of the new Insurance Act in 2017. We compare the current role in South Africa to actuarial roles in other jurisdictions and the prior statutory actuary role based on the Long-term Insurance Act. Based on an industry survey conducted by us, we highlight the differences in the interpretation and implementation of the role in different segments of the market. We discuss areas that have been problematic when embedding the HAF role in an organisation and investigate those from a professional, organisational and regulatory perspective. Based on this analysis we propose best practices for the implementation of the HAF role including a minimally compliant interpretation for smaller insurers and outline how the role can be interpreted, clarified or changed to achieve further improvements for the South African insurance market.

Actuarial reserving techniques have evolved from the application of algorithms, like the chain-ladder method, to stochastic models of claims development, and, more recently, have been enhanced by the application of machine learning techniques. Despite this proliferation of theory and techniques, there is relatively little guidance on which reserving techniques should be applied and when. In this paper, we revisit traditional reserving techniques within the framework of supervised learning to select optimal reserving models. We show that the use of optimal techniques can lead to more accurate reserves and investigate the circumstances under which different scoring metrics should be used.

The study of mortality improvements in South Africa has been complicated by data limitations: from a population perspective, the data are incomplete and misreported whereas pooled data collected from insurance companies for industry studies often span short time periods and exhibit significant heterogeneity. Notwithstanding these issues, accurate quantification of mortality improvement is critical for actuarial valuations of life insurers and pension funds. In this study, we analyse a unique pensioner dataset, covering the years 2000-2019, through the dual lenses of traditional actuarial analysis and deep learning techniques. We report on aggregate mortality improvements, as well as how mortality improvements differ depending on gender and pension amount. We investigate whether the observed mortality improvements might be attributed to changes in the composition of the pensioner dataset by contrasting these results with those derived from a more complex model that allows for these changes over time. Finally, we show the estimated impact of the observed improvements on pension liabilities and calibrate models to provide uncertainty around the estimated mortality improvement rates, to benchmark the SAM longevity stress and provide a basis for IFRS 17 risk adjustment calculations.

α-Synuclein (α-Syn) accumulation is a pathological hallmark of Parkinson's disease. Duplications and triplications ofSNCA, the gene coding for α-Syn, cause genetic forms of the disease, which suggests that increased α-Syn dosage can drive PD. To identify the proteins that regulate α-Syn, we previously performed a screen of potentially druggable genes that led to the identification of 60 modifiers. Among them, Doublecortin-like kinase 1 (DCLK1), a microtubule binding serine threonine kinase, emerged as a promising target due to its potent effect on α-Syn and potential druggability as a neuron-expressed kinase. In this study, we explore the relationship between DCLK1 and α-Syn in human cellular and mouse models of PD. First, we show that DCLK1 regulates α-Syn levels post-transcriptionally. Second, we demonstrate that knockdown ofDclk1reduces phosphorylated species of α-Syn and α-Syn-induced neurotoxicity in the SNc in two distinct mouse models of synucleinopathy. Last, silencingDCLK1in human neurons derived from individuals withSNCAtriplications reduces phosphorylated and total α-Syn, thereby highlighting DCLK1 as a potential therapeutic target to reduce pathological α-Syn in disease.SIGNIFICANCE STATEMENTDCLK1 regulates α-Syn protein levels, andDclk1knockdown rescues α-Syn toxicity in mice. This study provides evidence for a novel function for DCLK1 in the mature brain, and for its potential as a new therapeutic target for synucleinopathies.

In this tutorial we introduce recurrent neural networks (RNNs), and we describe the two most popular RNN architectures. These are the long short-term memory (LSTM) network and gated recurrent unit (GRU) network. Their common field of application is time series modeling, and we demonstrate their use on a mortality rate prediction problem using data from the Swiss female and male populations.

In this paper we review existing modelling approaches for analysing claims experience in the presence of reporting delays, reviewing the formulation of mortality incidence models such as GLMs. We then show how these approaches have traditionally been adjusted for late reporting of claims using either the IBNR approach or the more recent EBNER approach. We then go on to introduce a new model formulation that combines a model for late reported claims with a model for mortality incidence into a single model formulation. We then illustrate the use and performance of the traditional and the combined model formulations on data from a multinational reinsurer. We show how GLMs, lasso regression, gradient boosted trees and deep learning can be applied to the new formulation to produce results of superior accuracy compared to the traditional approaches.

Deep Learning models are currently being introduced into business processes to support decision-making in insurance companies. At the same time model risk is recognized as an increasingly relevant field within the management of operational risk that tries to mitigate the risk of poor business decisions because of flawed models or inappropriate model use. In this paper we try to determine how Deep Learning models are different from established actuarial models currently in use in insurance companies and how these differences might necessitate changes in the model risk management framework. We analyse operational risk in the development and implementation of Deep Learning models using examples from pricing and mortality forecasting to illustrate specific model risks and controls to mitigate those risks. We discuss changes in model governance and the role that model risk managers could play in providing assurance on the appropriate use of Deep Learning models.

The Lee-Carter model is a basic approach to forecasting mortality rates of a single population. Although extensions of the Lee-Carter model to forecasting rates for multiple populations have recently been proposed, the structure of these extended models is hard to justify and the models are often difficult to calibrate, relying on customized optimization schemes. Based on the paradigm of representation learning, we extend the Lee-Carter model to multiple populations using neural networks, which automatically select an optimal model structure. We fit this model to mortality rates since 1950 for all countries in the Human Mortality Database and observe that the out-of-sample forecasting performance of the model is highly competitive.

The main idea of this paper is to embed a classical actuarial regression model into a neural network architecture. This nesting allows us to learn model structure beyond the classical actuarial regression model if we use as starting point of the neural network calibration exactly the classical actuarial model. Such models can be fitted efficiently which allows us to explore bootstrap methods for prediction uncertainty. As an explicit example we consider the cross-classified over-dispersed Poisson model for general insurance claims reserving. We demonstrate how this model can be improved by neural network features.

Rapid advances in Artificial Intelligence and Machine Learning are creating products and services with the potential not only to change the environment in which actuaries operate, but also to provide new opportunities within actuarial science. These advances are based on a modern approach to designing, fitting and applying neural networks, generally referred to as “Deep Learning”. This paper investigates how actuarial science may adapt and evolve in the coming years to incorporate these new techniques and methodologies. After providing some background on machine learning and deep learning, and providing a heuristic for where actuaries might benefit from applying these techniques, the paper surveys emerging applications of AI in actuarial science, with examples from mortality modelling, claims reserving, non-life pricing and telematics. For some of the examples, code has been provided on GitHub so that the interested reader can experiment with these techniques for themselves. The paper concludes with an outlook on the potential for actuaries to integrate deep learning into their activities.

Silencing of fragile X mental retardation 1 (FMR1) gene and loss of fragile X mental retardation protein (FMRP) cause fragile X syndrome (FXS), a genetic disorder characterized by intellectual disability and autistic behaviors. FMRP is an mRNA-binding protein regulating neuronal translation of target mRNAs. Abnormalities in actin-rich dendritic spines are major neuronal features in FXS, but the molecular mechanism and identity of FMRP targets mediating this phenotype remain largely unknown. Cytoplasmic FMR1-interacting protein 2 (Cyfip2) was identified as an interactor of FMRP, and its mRNA is a highly ranked FMRP target in mouse brain. Importantly, Cyfip2 is a component of WAVE regulatory complex, a key regulator of actin cytoskeleton, suggesting that Cyfip2 could be implicated in the dendritic spine phenotype of FXS. Here, we generated and characterized Cyfip2-mutant (Cyfip2(+/-)) mice. We found that Cyfip2(+/-) mice exhibited behavioral phenotypes similar to Fmr1-null (Fmr1(-/y)) mice, an animal model of FXS. Synaptic plasticity and dendritic spines were normal in Cyfip2(+/-) hippocampus. However, dendritic spines were altered in Cyfip2(+/-) cortex, and the dendritic spine phenotype of Fmr1(-/y) cortex was aggravated in Fmr1(-/y); Cyfip2(+/-) double-mutant mice. In addition to the spine changes at basal state, metabotropic glutamate receptor (mGluR)-induced dendritic spine regulation was impaired in both Fmr1(-/y) and Cyfip2(+/-) cortical neurons. Mechanistically, mGluR activation induced mRNA translation-dependent increase of Cyfip2 in wild-type cortical neurons, but not in Fmr1(-/y) or Cyfip2(+/-) neurons. These results suggest that misregulation of Cyfip2 function and its mGluR-induced expression contribute to the neurobehavioral phenotypes of FXS.

Many neurodegenerative disorders, such as Alzheimer’s, Parkinson’s and polyglutamine diseases, share a common pathogenic mechanism: the abnormal accumulation of disease-causing proteins, due to either the mutant protein’s resistance to degradation or overexpression of the wild-type protein. We have developed a strategy to identify therapeutic entry points for such neurodegenerative disorders by screening for genetic networks that influence the levels of disease-driving proteins. We applied this approach, which integrates parallel cell-based and Drosophila genetic screens, to spinocerebellar ataxia type 1 (SCA1), a disease caused by expansion of a polyglutamine tract in ataxin 1 (ATXN1). Our approach revealed that downregulation of several components of the RAS–MAPK–MSK1 pathway decreases ATXN1 levels and suppresses neurodegeneration in Drosophila and mice. Importantly, pharmacological inhibitors of components of this pathway also decrease ATXN1 levels, suggesting that these components represent new therapeutic targets in mitigating SCA1. Collectively, these data reveal new therapeutic entry points for SCA1 and provide a proof-of-principle for tackling other classes of intractable neurodegenerative diseases. Cross-species genetic screens reveal that decreased mitogen-activated protein kinase signalling reduces polyglutamine-expanded ataxin 1 levels and toxicity in models of spinocerebellar ataxia type 1. The toxic accumulation of mutant — or sometimes wild-type — protein in the brain is a common cause of pathology in inherited neurodegenerative diseases. Decreasing this accumulation can reverse disease phenotype. Huda Zoghbi and colleagues have developed a cross-species genetic-screening strategy for the identification of druggable targets that could decrease the levels of ataxin 1, the disease-causing protein of spinocerebellar ataxia type 1. This strategy uncovered multiple therapeutic targets (RAS–MAPK–MSK1) that, when inhibited, decrease ataxin 1 levels and suppress its toxicity in human, mouse and Drosophila systems.

Previously, we reported that ATXN1 oligomers are the primary drivers of toxicity in Spinocerebellar ataxia type 1 (SCA1; Lasagna-Reeves et al., 2015). Here we report that polyQ ATXN1 oligomers can propagate locally in vivo in mice predisposed to SCA1 following intracerebral oligomeric tissue inoculation. Our data also show that targeting these oligomers with passive immunotherapy leads to some improvement in motor coordination in SCA1 mice and to a modest increase in their life span. These findings provide evidence that oligomer propagation is regionally limited in SCA1 and that immunotherapy targeting extracellular oligomers can mildly modify disease phenotypes. DOI: http://dx.doi.org/10.7554/eLife.10891.001

SUMMARY RNA splicing entails the coordinated interaction of more than 150 proteins in the spliceosome, one of the most complex of the cell’s molecular machines. We previously discovered that the RNA-binding motif protein 17 (RBM17), a component of the spliceosome, is essential for survival and cell maintenance. Here, we find that it interacts with the spliceosomal factors U2SURP and CHERP and that they reciprocally regulate each other’s stability, both in mouse and in human cells. Individual knockdown of each of the three proteins induces overlapping changes in splicing and gene expression of transcripts enriched for RNA-processing factors. Our results elucidate the function of RBM17, U2SURP, and CHERP and link the activity of the spliceosome to the regulation of downstream RNA-binding proteins. These data support the hypothesis that, beyond driving constitutive splicing, spliceosomal factors can regulate alternative splicing of specific targets., Graphical Abstract, In Brief De Maio et al. find that the splicing factor RBM17 establishes a physical and functional relation with U2SURP and CHERP. Knockdown of these U2 snRNP-associated spliceosomal components reveals their synergistic activity toward regulation of a given set of transcripts rather than a more predictable transcriptome-wide inhibition of splicing.

There have been no objective assessments to determine whether boys with MECP2 duplication have autism or whether female carriers manifest phenotypes. This study characterizes the clinical and neuropsychiatric phenotypes of affected boys and carrier females.Eight families (9 males and 9 females) with MECP2 duplication participated. A detailed history, physical examination, electroencephalogram, developmental evaluation, Autism Diagnostic Observation Schedule, and Autism Diagnostic Interview-Revised were performed for each boy. Carrier females completed the Symptom Checklist-90-R, Wechsler Abbreviated Scale of Intelligence, Broad Autism Phenotype Questionnaire, and detailed medical and mental health histories. Size and gene content of each duplication were determined by array comparative genome hybridization. X-chromosome inactivation patterns were analyzed using leukocyte DNA. MECP2 and IRAK1 RNA levels were quantified from lymphoblast cell lines, and western blots were performed to assess MeCP2 protein levels.All of the boys demonstrated mental retardation and autism. Poor expressive language, gaze avoidance, repetitive behaviors, anxiety, and atypical socialization were prevalent. Female carriers had psychiatric symptoms, including generalized anxiety, depression, and compulsions that preceded the birth of their children. The majority exhibited features of the broad autism phenotype and had higher nonverbal compared to verbal reasoning skills.Autism is a defining feature of the MECP2 duplication syndrome in boys. Females manifest phenotypes despite 100% skewing of X-inactivation and normal MECP2 RNA levels in peripheral blood. Analysis of the duplication size, MECP2 and IRAK1 RNA levels, and MeCP2 protein levels revealed that most of the traits in affected boys are likely due to the genomic region spanning of MECP2 and IRAK1. The phenotypes observed in carrier females may be secondary to tissue-specific dosage alterations and require further study. Ann Neurol 2009;66:771-782.

Recent studies indicate that soluble oligomers drive pathogenesis in several neurodegenerative proteinopathies, including Alzheimer and Parkinson disease. Curiously, the same conformational antibody recognizes different disease-related oligomers, despite the variations in clinical presentation and brain regions affected, suggesting that the oligomer structure might be responsible for toxicity. We investigated whether polyglutamine-expanded ATAXIN-1, the protein that underlies spinocerebellar ataxia type 1, forms toxic oligomers and, if so, what underlies their toxicity. We found that mutant ATXN1 does form oligomers and that oligomer levels correlate with disease progression in the Atxn1154Q/+ mice. Moreover, oligomeric toxicity, stabilization and seeding require interaction with Capicua, which is expressed at greater ratios with respect to ATXN1 in the cerebellum than in less vulnerable brain regions. Thus, specific interactors, not merely oligomeric structure, drive pathogenesis and contribute to regional vulnerability. Identifying interactors that stabilize toxic oligomeric complexes could answer longstanding questions about the pathogenesis of other proteinopathies. DOI: http://dx.doi.org/10.7554/eLife.07558.001, eLife digest Spinocerebellar ataxia type 1 (SCA1) is a progressive neurodegenerative disease in which damage to the brain regions that control movement results in the gradual loss of coordination and motor skills. The disease is a caused by a mutation in the gene that codes for a protein called ATAXIN-1. In healthy individuals this protein contains up to 39 copies of an amino acid called glutamine. However, the mutant gene can encode for 40 or more copies of glutamine, which results in a longer-than-usual ATAXIN-1 protein with toxic properties. Within the brain, some of the toxic ATAXIN-1 proteins form insoluble deposits, while the rest remain soluble. At first it was assumed that the insoluble deposits were responsible for the neurodegeneration seen in SCA1. However, closer examination revealed that these deposits form mainly in brain regions that do not degenerate, which suggests that they might instead have a protective role. This is consistent with evidence from research into other brain disorders, including Alzheimer's disease, which suggests that the soluble form of the toxic proteins might be causing these diseases. Lasagna-Reeves et al. now provide the first direct evidence that the soluble form of the toxic ATAXIN-1 proteins are indeed harmful in a mouse model of SCA1. Experiments reveal that these soluble proteins accumulate in the brain regions that undergo degeneration in SCA1, such as the cerebellum, but not in those regions that remain intact. Moreover, the motor skills and coordination of the mice get worse as the level of soluble toxic ATAXIN-1 increases. Lasagna-Reeves et al. go on to show that a protein called capicua stabilizes the toxic ATAXIN-1 proteins, which keeps them from forming insoluble deposits. Since capicua is particularly abundant in the cerebellum, this could explain the high levels of toxic ATAXIN-1 in this region and why it is vulnerable to degeneration. Future experiments are needed to investigate whether proteins equivalent to capicua might play a similar role in stabilizing toxic proteins in Alzheimer's, Parkinson's and Huntington's diseases, and whether preventing this stabilization could have therapeutic potential. DOI: http://dx.doi.org/10.7554/eLife.07558.002