Large Heterogeneous Datasets Research Articles

Automated quality control of small animal MR neuroimaging data

Abstract Magnetic resonance imaging (MRI) is a valuable tool for studying brain structure and function in animal and clinical studies. With the growth of public MRI repositories, access to data has finally become easier. However, filtering large datasets for potential poor-quality outliers can be a challenge. We present AIDAqc, a machine-learning-assisted automated Python-based command-line tool for small animal MRI quality assessment. Quality control features include signal-to-noise ratio (SNR), temporal SNR, and motion. All features are automatically calculated and no regions of interest are needed. Automated outlier detection for a given dataset combines the interquartile range and the machine-learning methods one-class support vector machine, isolation forest, local outlier factor, and elliptic envelope. To evaluate the reliability of individual quality control metrics, a simulation of noise (Gaussian, salt and pepper, speckle) and motion was performed. In outlier detection, single scans with induced artifacts were successfully identified by AIDAqc. AIDAqc was challenged in a large heterogeneous dataset collected from 19 international laboratories, including data from mice, rats, rabbits, hamsters, and gerbils, obtained with different hardware and at different field strengths. The results show that the manual inter-rater agreement (mean Fleiss Kappa score 0.17) is low when identifying poor-quality data. A direct comparison of AIDAqc results, therefore, showed only low-to-moderate concordance. In a manual post hoc validation of AIDAqc output, precision was high (>70%). The outlier data can have a significant impact on further postprocessing, as shown in representative functional and structural connectivity analysis. In summary, this pipeline optimized for small animal MRI provides researchers with a valuable tool to efficiently and effectively assess the quality of their MRI data, which is essential for improved reliability and reproducibility.

Development, validation and use of custom software for the analysis of pain trajectories

In chronic musculoskeletal conditions, the prognosis tends to be more informative than the diagnosis for the future course of the disease. Many studies have identified clusters of patients who seemingly share similar pain trajectories. In a dataset of low back pain (LBP) patients, pain trajectories have been identified, and distinct trajectory types have been defined, making it possible to create pattern recognition software that can classify patients into respective pain trajectories reflecting their condition. It has been suggested that the classification of pain trajectories may create clinically meaningful subgroups of patients in an otherwise heterogeneous population of patients with LBP. A software tool was created that combined the ability to recognise the pain trajectory of patients with a system that could create subgroups of patients based on their characteristics. This tool is primarily meant for researchers to analyse trends in large heterogeneous datasets without large losses of data. Prospective analysis of pain trajectories is not directly helpful for clinicians. However, the tool might aid in the identification of patient characteristics which have predictive capabilities of the most likely trajectory a patient might experience in the future. This will help clinicians to tailor their advice and treatment for a specific patient.

Immersive and Interactive 3D Visualization of Large-Scale Geoscientific Data

Abstract Virtual reality (VR) technology has great potential in supporting planetary scientists and the field of geosciences by offering immersive and interactive experiences for data exploration and analysis. This paper shows the opportunities presented by the use of VR technology in the geosciences domain while also identifying the open challenges associated with this integration. Our focus is on highlighting the several benefits that VR brings, including stereo vision, head tracking, and collaborative capabilities. These features can foster knowledge exchange and interdisciplinary research. Nevertheless, the adoption of VR presents certain challenges that need to be addressed. These include maintaining high refresh rates, handling large heterogeneous datasets, and striking a balance between visualization fidelity and performance. Fortunately, significant advancements have been made in high-performance data analysis, progressive data streaming, and real-time visualization, enabling interactive exploration of large-scale datasets within VR environments. To support a broader adoption among domain experts, we propose a visualization approach that scales both with display and with data size. As such, the system can be used to interactively explore large-scale datasets in immersive environments like CAVEs or powerwalls, on HMDs, or on traditional desktop setups. This approach allows for seamless transitions between desktop and VR experiences, leveraging immersive environments for collaboration and outreach activities. In this paper, we present an architectural framework, hardware environments, use cases, lessons learned, and the future potential of VR in this field. Through this scalable system, we anticipate transformative advancements in scientific exploration, collaboration, and knowledge dissemination.

L-PBF High-Throughput Data Pipeline Approach for Multi-modal Integration

Metal-based additive manufacturing requires active monitoring solutions for assessing part quality. Multiple sensors and data streams, however, generate large heterogeneous data sets that are impractical for manual assessment and characterization. In this work, an automated pipeline is developed that enables feature extraction from high-speed camera video and multi-modal data analysis. The framework removes the need for manual assessment through the utilization of deep learning techniques and training models in a weakly supervised paradigm. We demonstrate this pipeline’s capability over 700,000 high-speed camera frames. The pipeline successfully extracts melt pool and spatter geometries and links them to corresponding pyrometry, radiography, and processparameter information. 715 individual prints are examined to reveal melt pool areas that exceeds 0.07 mm2 and pyrometry signal over a threshold (375 pyrometry units) were more likely to have defects. These automated processes enable massive throughput of characterization techniques.

Heterogeneity in strategy use during arbitration between experiential and observational learning

To navigate our complex social world, it is crucial to deploy multiple learning strategies, such as learning from directly experiencing action outcomes or from observing other people’s behavior. Despite the prevalence of experiential and observational learning in humans and other social animals, it remains unclear how people favor one strategy over the other depending on the environment, and how individuals vary in their strategy use. Here, we describe an arbitration mechanism in which the prediction errors associated with each learning strategy influence their weight over behavior. We designed an online behavioral task to test our computational model, and found that while a substantial proportion of participants relied on the proposed arbitration mechanism, there was some meaningful heterogeneity in how people solved this task. Four other groups were identified: those who used a fixed mixture between the two strategies, those who relied on a single strategy and non-learners with irrelevant strategies. Furthermore, groups were found to differ on key behavioral signatures, and on transdiagnostic symptom dimensions, in particular autism traits and anxiety. Together, these results demonstrate how large heterogeneous datasets and computational methods can be leveraged to better characterize individual differences.

Image Quality Assessment Tool for Conventional and Dynamic Magnetic Resonance Imaging Acquisitions.

Image quality assessment of magnetic resonance imaging (MRI) data is an important factor not only for conventional diagnosis and protocol optimization but also for fairness, trustworthiness, and robustness of artificial intelligence (AI) applications, especially on large heterogeneous datasets. Information on image quality in multi-centric studies is important to complement the contribution profile from each data node along with quantity information, especially when large variability is expected, and certain acceptance criteria apply. The main goal of this work was to present a tool enabling users to assess image quality based on both subjective criteria as well as objective image quality metrics used to support the decision on image quality based on evidence. The evaluation can be performed on both conventional and dynamic MRI acquisition protocols, while the latter is also checked longitudinally across dynamic series. The assessment provides an overall image quality score and information on the types of artifacts and degrading factors as well as a number of objective metrics for automated evaluation across series (BRISQUE score, Total Variation, PSNR, SSIM, FSIM, MS-SSIM). Moreover, the user can define specific regions of interest (ROIs) to calculate the regional signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), thus individualizing the quality output to specific use cases, such as tissue-specific contrast or regional noise quantification.

From local to global gene co-expression estimation using single-cell RNA-seq data.

In genomics studies, the investigation of gene relationships often brings important biological insights. Currently, the large heterogeneous datasets impose new challenges for statisticians because gene relationships are often local. They change from one sample point to another, may only exist in a subset of the sample, and can be nonlinear or even nonmonotone. Most previous dependence measures do not specifically target local dependence relationships, and the ones that do are computationally costly. In this paper, we explore a state-of-the-art network estimation technique that characterizes gene relationships at the single cell level, under the name of cell-specific gene networks. We first show that averaging the cell-specific gene relationship over a population gives a novel univariate dependence measure, the averaged Local Density Gap (aLDG), that accumulates local dependence and can detect any nonlinear, nonmonotone relationship. Together with a consistent nonparametric estimator, we establish its robustness on both the population and empirical levels. Then, we show that averaging the cell-specific gene relationship over mini-batches determined by some external structure information (eg, spatial or temporal factor) better highlights meaningful local structure change points. We explore the application of aLDG and its minibatch variant in many scenarios, including pairwise gene relationship estimation, bifurcating point detection in cell trajectory, and spatial transcriptomics structure visualization. Both simulations and real data analysis show that aLDG outperforms existing ones.

Twenty years of network meta-analysis: Continuing controversies and recent developments.

Network meta-analysis (NMA) is an extension of pairwise meta-analysis (PMA) which combines evidence from trials on multiple treatments in connected networks. NMA delivers internally consistent estimates of relative treatment efficacy, needed for rational decision making. Over its first 20 years NMA's use has grown exponentially, with applications in both health technology assessment (HTA), primarily re-imbursement decisions and clinical guideline development, and clinical research publications. This has been a period of transition in meta-analysis, first from its roots in educational and social psychology, where large heterogeneous datasets could be explored to find effect modifiers, to smaller pairwise meta-analyses in clinical medicine on average with less than six studies. This has been followed by narrowly-focused estimation of the effects of specific treatments at specific doses in specific populations in sparse networks, where direct comparisons are unavailable or informed by only one or two studies. NMA is a powerful and well-established technique but, in spite of the exponential increase in applications, doubts about the reliability and validity of NMA persist. Here we outline the continuing controversies, and review some recent developments. We suggest that heterogeneity should be minimized, as it poses a threat to the reliability of NMA which has not been fully appreciated, perhaps because it has not been seen as a problem in PMA. More research is needed on the extent of heterogeneity and inconsistency in datasets used for decision making, on formal methods for making recommendations based on NMA, and on the further development of multi-level network meta-regression.

A Comparative Analysis of Machine Learning Models for the Detection of Undiagnosed Diabetes Patients

Introduction: Early detection of type 2 diabetes is essential for preventing long-term complications. However, screening the entire population for diabetes is not cost-effective, so identifying individuals at high risk for this disease is crucial. The aim of this study was to compare the performance of five diverse machine learning (ML) models in classifying undiagnosed diabetes using large heterogeneous datasets. Methods: We used machine learning data from several years of the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2018 to identify people with undiagnosed diabetes. The dataset included 45,431 participants, and biochemical confirmation of glucose control (HbA1c) were used to identify undiagnosed diabetes. The predictors were based on simple and clinically obtainable variables, which could be feasible for prescreening for diabetes. We included five ML models for comparison: random forest, AdaBoost, RUSBoost, LogitBoost, and a neural network. Results: The prevalence of undiagnosed diabetes was 4%. For the classification of undiagnosed diabetes, the area under the ROC curve (AUC) values were between 0.776 and 0.806. The positive predictive values (PPVs) were between 0.083 and 0.091, the negative predictive values (NPVs) were between 0.984 and 0.99, and the sensitivities were between 0.742 and 0.871. Conclusion: We have demonstrated that several types of classification models can accurately classify undiagnosed diabetes from simple and clinically obtainable variables. These results suggest that the use of machine learning for prescreening for undiagnosed diabetes could be a useful tool in clinical practice.

Neuroimaging-based classification of PTSD using data-driven computational approaches: A multisite big data study from the ENIGMA-PGC PTSD consortium

BackgroundRecent advances in data-driven computational approaches have been helpful in devising tools to objectively diagnose psychiatric disorders. However, current machine learning studies limited to small homogeneous samples, different methodologies, and different imaging collection protocols, limit the ability to directly compare and generalize their results. Here we aimed to classify individuals with PTSD versus controls and assess the generalizability using a large heterogeneous brain datasets from the ENIGMA-PGC PTSD Working group. MethodsWe analyzed brain MRI data from 3,477 structural-MRI; 2,495 resting state-fMRI; and 1,952 diffusion-MRI. First, we identified the brain features that best distinguish individuals with PTSD from controls using traditional machine learning methods. Second, we assessed the utility of the denoising variational autoencoder (DVAE) and evaluated its classification performance. Third, we assessed the generalizability and reproducibility of both models using leave-one-site-out cross-validation procedure for each modality. ResultsWe found lower performance in classifying PTSD vs. controls with data from over 20 sites (60 % test AUC for s-MRI, 59 % for rs-fMRI and 56 % for d-MRI), as compared to other studies run on single-site data. The performance increased when classifying PTSD from HC without trauma history in each modality (75 % AUC). The classification performance remained intact when applying the DVAE framework, which reduced the number of features. Finally, we found that the DVAE framework achieved better generalization to unseen datasets compared with the traditional machine learning frameworks, albeit performance was slightly above chance. ConclusionThese results have the potential to provide a baseline classification performance for PTSD when using large scale neuroimaging datasets. Our findings show that the control group used can heavily affect classification performance. The DVAE framework provided better generalizability for the multi-site data. This may be more significant in clinical practice since the neuroimaging-based diagnostic DVAE classification models are much less site-specific, rendering them more generalizable.

Potential and challenges of applying artificial intelligence (AI) in geosciences to the search for a high-level waste repository in Germany

Abstract. In a science-based site selection procedure (StandAV), the Federal Republic of Germany is searching for a site with the best possible safety for a repository for high-level waste (HLW) over a period of 1 million years. This is a challenging task, given the size of Germany, its geological variability and the verification period required to ensure the best possible long-term safety. In addition, an enormous amount of heterogeneous geodata has to be processed. The application of artificial-intelligence-based methods in geosciences has great potential for dealing with large heterogeneous data sets. In the German Federal Office for the Safety of Nuclear Waste Management (BASE)-funded project “Application of artificial intelligence (AI) in the site selection process for a deep geological repository (FKZ 4721E03210)”, an interdisciplinary assessment tool is being developed to evaluate the applicability of AI methods in the geosciences. This contribution focuses on the potential and challenges of applying AI in geosciences with respect to key geological activities in the StandAV. Limitations that may arise from the use of AI with respect to key activities in the StandAV are highlighted. The necessary conditions for its future applicability are proposed. The results show that AI methods offer clear advantages over conventional methods for data management and handling large geological datasets and for modelling complex long-term and coupled geological processes. However, AI methods are generally only transferable to the geoscientific questions of the StandAV with methodological and subject-specific adaptations. Furthermore, the use of AI requires sufficient data, both in terms of quality and quantity. The study also shows that AI should only be used in a supportive way to address geological issues in key activities and should not have any decision-making power when used in StandAV. For example, AI can carry the risk of data and developer bias, which in turn can have serious consequences for the correct interpretation of results. High demands must be placed on the traceability of the AI methods used. AI methods that do not meet the transparency requirements of the StandAV carry a significant risk of jeopardising public confidence in the participation process. This could increase the general mistrust and scepticism towards AI in the public perception (Krob et al., 2023). It is strongly recommended that all methods should be evaluated and validated iteratively and that the results should be made available to the public when applied to the key activities of the StandAV.

KGETCDA: an efficient representation learning framework based on knowledge graph encoder from transformer for predicting circRNA-disease associations.

Recent studies have demonstrated the significant role that circRNA plays in the progression of human diseases. Identifying circRNA-disease associations (CDA) in an efficient manner can offer crucial insights into disease diagnosis. While traditional biological experiments can be time-consuming and labor-intensive, computational methods have emerged as a viable alternative in recent years. However, these methods are often limited by data sparsity and their inability to explore high-order information. In this paper, we introduce a novel method named Knowledge Graph Encoder from Transformer for predicting CDA (KGETCDA). Specifically, KGETCDA first integrates more than 10 databases to construct a large heterogeneous non-coding RNA dataset, which contains multiple relationships between circRNA, miRNA, lncRNA and disease. Then, a biological knowledge graph is created based on this dataset and Transformer-based knowledge representation learning and attentive propagation layers are applied to obtain high-quality embeddings with accurately captured high-order interaction information. Finally, multilayer perceptron is utilized to predict the matching scores of CDA based on their embeddings. Our empirical results demonstrate that KGETCDA significantly outperforms other state-of-the-art models. To enhance user experience, we have developed an interactive web-based platform named HNRBase that allows users to visualize, download data and make predictions using KGETCDA with ease. The code and datasets are publicly available at https://github.com/jinyangwu/KGETCDA.

SASE: Self-Adaptive noise distribution network for Speech Enhancement with Federated Learning using heterogeneous data

Correlation of speech data is ubiquitous in real-world environments. Heterogeneous data involving different distributions hinder federated learning (FL) significantly, which can lead to drifted global models exhibiting intractable convergence. Based on the differences between speech and noise distributions, this paper proposes self-adaptive noise distribution network for speech enhancement (SASE)—a complex-domain denoising model with FL—to tackle the problem of data heterogeneity. We propose a complex-valued time–frequency gate attention mechanism (TF-GA), which is optimized in both time and frequency domains to enable the extraction of rich speech and noise distribution information. Further, we construct the self-adaptive Gaussian unitary ensemble attention (SA-GUEA) block in the SASE network to make it adaptable to the noise distribution. Data heterogeneity is addressed using the SASE model by developing the CommonVoice (CV) dataset with noise—this large heterogeneous dataset contains several speakers, acoustic environments, and noise obtained from different devices. In a realistic FL experimental environment, we also develop the Loss-based and PESQ-based optimization weighting strategies that intelligently update the server model with a large-scale heterogeneous dataset intelligently, leading to better generalization performance corresponding to non-independent and identical distribution (non-IID) data distributions. Empirical studies based on this theoretical framework demonstrate that the proposed SASE model not only exhibits high applicability to unknown noise on an independent and identical distribution (IID) VoiceBank + DEMAND dataset, but also achieves successful denoising for real environment noise on the non-IID CV Chinese + Noise92 and CV English + TUT datasets.

Machine learning-based colorectal cancer prediction using global dietary data

BackgroundColorectal cancer (CRC) is the third most commonly diagnosed cancer worldwide. Active health screening for CRC yielded detection of an increasingly younger adults. However, current machine learning algorithms that are trained using older adults and smaller datasets, may not perform well in practice for large populations.AimTo evaluate machine learning algorithms using large datasets accounting for both younger and older adults from multiple regions and diverse sociodemographics.MethodsA large dataset including 109,343 participants in a dietary-based colorectal cancer ase study from Canada, India, Italy, South Korea, Mexico, Sweden, and the United States was collected by the Center for Disease Control and Prevention. This global dietary database was augmented with other publicly accessible information from multiple sources. Nine supervised and unsupervised machine learning algorithms were evaluated on the aggregated dataset.ResultsBoth supervised and unsupervised models performed well in predicting CRC and non-CRC phenotypes. A prediction model based on an artificial neural network (ANN) was found to be the optimal algorithm with CRC misclassification of 1% and non-CRC misclassification of 3%.ConclusionsANN models trained on large heterogeneous datasets may be applicable for both younger and older adults. Such models provide a solid foundation for building effective clinical decision support systems assisting healthcare providers in dietary-related, non-invasive screening that can be applied in large studies. Using optimal algorithms coupled with high compliance to cancer screening is expected to significantly improve early diagnoses and boost the success rate of timely and appropriate cancer interventions.

DataCurator.jl: efficient, portable and reproducible validation, curation and transformation of large heterogeneous datasets using human-readable recipes compiled into machine-verifiable templates.

Large-scale processing of heterogeneous datasets in interdisciplinary research often requires time-consuming manual data curation. Ambiguity in the data layout and preprocessing conventions can easily compromise reproducibility and scientific discovery, and even when detected, it requires time and effort to be corrected by domain experts. Poor data curation can also interrupt processing jobs on large computing clusters, causing frustration and delays. We introduce DataCurator, a portable software package that verifies arbitrarily complex datasets of mixed formats, working equally well on clusters as on local systems. Human-readable TOML recipes are converted into executable, machine-verifiable templates, enabling users to easily verify datasets using custom rules without writing code. Recipes can be used to transform and validate data, for pre- or post-processing, selection of data subsets, sampling and aggregation, such as summary statistics. Processing pipelines no longer need to be burdened by laborious data validation, with data curation and validation replaced by human and machine-verifiable recipes specifying rules and actions. Multithreaded execution ensures scalability on clusters, and existing Julia, R and Python libraries can be reused. DataCurator enables efficient remote workflows, offering integration with Slack and the ability to transfer curated data to clusters using OwnCloud and SCP. Code available at: https://github.com/bencardoen/DataCurator.jl.

The Security Challenges of Big Data Analytics: A Systematic Literature Review

The huge amount of data generated from heterogeneous sources such as social networking sites, healthcare applications, sensor networks and many other sources are drastically increasing from time to time swiftly. Big Data is described as extremely large datasets that have grown beyond the capability to manage and analyze them with traditional database processing tools. Big data analytics is the use of advanced analytical techniques against a very large heterogeneous datasets that include structured, semi-structured and unstructured data from different sources. The larger the quantity of data by itself is not advantageous unless analyzed to produce valuable information. This deluge amount of data creates an operational risk in which, the risks arise from storage devices, security of tools or the technologies used to analyze the data. In this paper, we perform a systematic literature review to give comprehensive review of security challenges and risks related to big data analytics. Security mechanisms such as cryptographic and non-cryptographic techniques are used to secure big data during analytics. The security of big data at rest and in transit gets enough investigation while a few researches had done at securing data at processing stage. Even though a number of possible techniques were proposed for big data security, it still suffers performance issues. This article is trying to explor security issues that used for preserving the Confidentiality, Integrity and Availability (CIA triad), non-repudiation as well as Access control in the context of big data analytics. Finally, we identify open future research directions for security of big data analytics. This paper also can serve as a good reference source for the development of modern security-preserving techniques to address various challenges of big data analytics security and privacy-issues.

Quantitative analysis of 6,150 real‐world amyloid Positron Emission Tomography (PET) scans from the Imaging Dementia–Evidence for Amyloid Scanning (IDEAS) study

AbstractBackgroundAmyloid‐PET had been widely used and validated in research settings, using highly selected samples, harmonized acquisition protocols, co‐registration with MRI, and central interpretation by highly experienced experts. In contrast, in clinical settings, more heterogeneous acquisition, reconstruction, and interpretation may compromise the accuracy of the imaging. We quantitatively analyzed real‐world amyloid‐PET scans to assess their validity.MethodIDEAS acquired 18,295 amyloid PET scans at 343 PET facilities in patients with MCI or dementia using 18F‐florbetapir, 18F‐florbetaben, or 18F‐flutemetamol. Scans were visually interpreted at each site as either negative or positive for cortical tracer retention. Scans from consenting patients were archived. As of December 1, 2021, amyloid‐PET scans from 6,263 unique participants were available for analysis. Exclusion for lack of valid images or clinical data and failure of quality checks resulted in 6,150 (98.2%) valid scans. We analyzed the scans using a recently validated PET‐only processing pipeline designed to process heterogeneous amyloid‐PET scans (Iaccarino et al, 2022) and quantified cortical uptake in Centiloid (CL) units. A previously established neuropathology‐based threshold of 24.4 CL was used to define amyloid‐PET positivity independent of visual reads.ResultMean CL was higher in dementia (mean±SD = 53±51) than in MCI (40±48) (mean difference: 13; 95%CI: 10‐15). Mean CL of visually negative scans (3±27) was very close to 0, as expected for patients without amyloid accumulation, and significantly lower than visually positive scans (72±41) (mean difference: 69; 95%CI: 67‐70) (table 1). High concordance was found between local visual reads and CL‐based positivity (86.5%, Cohen’s κ=0.72, figure 1). CL exhibited a bimodal distribution, with most scans clearly positive or negative, and a minority of visual‐quantitative discordant scans surrounding the positivity threshold (figure 2). CL negatively correlated with MMSE (r=‐0.19, p<.001, figure 3). CL further correlated with the level of confidence in the diagnosis of Alzheimer’s Disease (AD), as was indicated by clinicians before the performance of PET (r=0.13, p<.001, figure 4).ConclusionA large heterogeneous dataset of real‐world amyloid‐PET scans analyzed quantitatively, shows high concordance with visual reads, and expected relationships with clinical and neuropsychological measures of AD.

Chickens, Inc.: Was UK Newspapers’ Framing of the Chicken Meat Production Industry Compatible with Holding Corporate Power to Account?

ABSTRACT Journalism is presented as fundamental to democratic accountability in that news media are both able and expected to hold power to account. Such normative expectations, which justify the protections given journalists and news media, are most frequently studied with regards to state power as the natural object of press scrutiny. This article reports on a successful effort to conduct a replicable analysis of a paradigmatic case of corporate power, the UK chicken meat production industry, that asked whether and how newspapers hold corporate power to account. The analytic framework that we developed to support our two-stage framing analysis decomposed accountability into problematization, causal interpretation and attributions of responsibility, thus allowing us to systematically describe how newspapers shape the public debates in a large heterogeneous dataset. We examined a census of relevant articles from seven UK outlets published between 1985 and 2016 (N = 766). While we were pleased to find that our method, if labor intensive, was fully workable, we were concerned to find that media practice was not compatible with holding corporate power to account. These findings raise serious concerns at the levels of this case, for media practice and for media scholarship.

Graph Neural Tree: A novel and interpretable deep learning-based framework for accurate molecular property predictions

Determining various properties of molecules is a critical step in drug discovery. Recently, with the improvement of large heterogeneous datasets and the development of deep learning approaches, more and more scientists have turned their attention to neural network-based virtual preliminary screening to reduce the time and monetary cost of drug discovery. However, the poor interpretability of deep learning masks causality, so models' conclusions are often beyond the comprehension of human users, which reduces the credibility of the model and makes it difficult for chemists to further narrow the huge chemical space based on models’ results. Thus, this study develops a novel framework consisting of Graph Neural Networks for feature extraction, Curriculum-Based Learning Strategies for optimization, and a Learning Binary Neural Tree (LBNT) for prediction, to improve the performance of neural networks and reveal their decision-making process to chemists. The framework encodes molecular graph data with graph neural networks (GNNs), then retrains the encoder with curriculum-based learning strategies to reduce uncertainty and improve accuracy, and finally uses LBNT as the predictor, which joint retrains with the encoder after independently training, for prediction and visualization. The framework is validated on the public datasets and compared to single GNNs with normal training strategies as well as GNN encoders with common machine learning predictors instead of the LBNT predictor. The result reveals that the proposed framework enhances the point prediction accuracy of the completely trained GNN and reduces its uncertainty through curriculum-based learning, and further improves the accuracy by combining LBNT. Besides, compared with common machine learning tools, the LBNT predictor generally has the best performance because of joint retraining with the GNN encoder. The decision-making process of LBNT is also better and easier to explain than that of other models.

Evaluation and comparison of federated learning algorithms for Human Activity Recognition on smartphones

Pervasive computing promotes the integration of smart devices in our living spaces to develop services providing assistance to people. Such smart devices are increasingly relying on cloud-based Machine Learning, which raises questions in terms of security (data privacy), reliance (latency), and communication costs. In this context, Federated Learning (FL) has been introduced as a new machine learning paradigm enhancing the use of local devices. At the server level, FL aggregates models learned locally on distributed clients to obtain a more general model. In this way, no private data is sent over the network, and the communication cost is reduced. Unfortunately, however, the most popular federated learning algorithms have been shown not to be adapted to some highly heterogeneous pervasive computing environments. In this paper, we propose a new FL algorithm, termed FedDist, which can modify models (here, deep neural network) during training by identifying dissimilarities between neurons among the clients. This permits to account for clients’ specificity without impairing generalization. FedDist evaluated with three state-of-the-art federated learning algorithms on three large heterogeneous mobile Human Activity Recognition datasets. Results have shown the ability of FedDist to adapt to heterogeneous data and the capability of FL to deal with asynchronous situations.