Unreliable Performance Research Articles

Crossfeat: a transformer-based cross-feature learning model for predicting drug side effect frequency.

Safe drug treatment requires an understanding of the potential side effects. Identifying the frequency of drug side effects can reduce the risks associated with drug use. However, existing computational methods for predicting drug side effect frequencies heavily depend on known drug side effect frequency information. Consequently, these methods face challenges when predicting the side effect frequencies of new drugs. Although a few methods can predict the side effect frequencies of new drugs, they exhibit unreliable performance owing to the exclusion of drug-side effect relationships. This study proposed CrossFeat, a model based on convolutional neural network-transformer architecture with cross-feature learning that can predict the occurrence and frequency of drug side effects for new drugs, even in the absence of information regarding drug-side effect relationships. CrossFeat facilitates the concurrent learning of drugs and side effect information within its transformer architecture. This simultaneous exchange of information enables drugs to learn about their associated side effects, while side effects concurrently acquire information about the respective drugs. Such bidirectional learning allows for the comprehensive integration of drug and side effect knowledge. Our five-fold cross-validation experiments demonstrated that CrossFeat outperforms existing studies in predicting side effect frequencies for new drugs without prior knowledge. Our model offers a promising approach for predicting the drug side effect frequencies, particularly for new drugs where prior information is limited. CrossFeat's superior performance in cross-validation experiments, along with evidence from case studies and ablation experiments, highlights its effectiveness.

Pre-Trained Language Model Ensemble for Arabic Fake News Detection

Fake news detection (FND) remains a challenge due to its vast and varied sources, especially on social media platforms. While numerous attempts have been made by academia and the industry to develop fake news detection systems, research on Arabic content remains limited. This study investigates transformer-based language models for Arabic FND. While transformer-based models have shown promising performance in various natural language processing tasks, they often struggle with tasks involving complex linguistic patterns and cultural contexts, resulting in unreliable performance and misclassification problems. To overcome these challenges, we investigated an ensemble of transformer-based models. We experimented with five Arabic transformer models: AraBERT, MARBERT, AraELECTRA, AraGPT2, and ARBERT. Various ensemble approaches, including a weighted-average ensemble, hard voting, and soft voting, were evaluated to determine the most effective techniques for boosting learning models and improving prediction accuracies. The results of this study demonstrate the effectiveness of ensemble models in significantly boosting the baseline model performance. An important finding is that ensemble models achieved excellent performance on the Arabic Multisource Fake News Detection (AMFND) dataset, reaching an F1 score of 94% using weighted averages. Moreover, changing the number of models in the ensemble has a slight effect on the performance. These key findings contribute to the advancement of fake news detection in Arabic, offering valuable insights for both academia and the industry

Precision Self‐Assembly of Supramolecules with Heterogeneous Derivatives

AbstractSupramolecular self‐assembly with well‐defined building blocks like lipids, deoxyribonucleic acid, or ligands relies on accessible molecular structures and predictable interactions. However, assembling heterogeneous, undefined blocks, such as disordered proteins, amorphous solids, and catecholic derivatives, remains challenging due to their unpredictable assembly, leading to irreversible aggregation, severe precipitation, and unreliable performance. Here, the first programmable, sustainable, and durable self‐assembly strategy of supramolecules with heterogenous is presented, derived blocks via harmonizing multiple molecular interactions. This approach achieves reversible assembly/disassembly, ≈73.7% reduced precipitation, and salt‐ and alkaline‐durability under freeze‐thaw cycles in model catecholic derivatives, functioning effectively as robust adhesive primers and hydrogel interfacial strengtheners. Moreover, through molecular force measurements and computational simulations, the first general criterion and benchmark for high precision supramolecular self‐assembly is proposed, applicable to complex derivatives and interactions: with blocks bearing multiple binding sites existing, the co‐assembling blocks should bear at least two binding sites with minimum binding strength (≈17 to ≈37 kJ mol−1) to prevent disassembly. This study paves the way and provides benchmarks for precision self‐assembly of diverse supramolecules using heterogeneous derivatives for adhesion technology, nanomaterial synthesis and bio‐inspired applications.

Self-Assembly of Silole-Based Aggregation-Induced Emission Compounds with Green Fluorescent Protein under Physiological Conditions for Traceable and Versatile Drug Delivery.

Biomacromolecules are viewed as promising drugs due to their specific functions in biological processes, biocompatibility, and pharmacological efficacy. Injective administration, chosen to avoid intestinal barriers, may in turn lead to immediate decay in the circulation system, unreliable targeting performance, or the induction of immune responses. For some biomacromolecules, chemically modified proteins have been developed for practical use. Various cargo or carrier systems are under development but have been delayed by technical difficulties. We present self-assembled nanocapsules with diameters ranging from 100 to 500 nm that can be deployed in physiological buffers to enclose various substances present in the buffers at the same time. Our amphiphilic nanocapsule, consisting of silole-core dendrimer products as the hydrophobic part and green fluorescent protein (GFP) derivatives as the hydrophilic part, connects and assembles spontaneously when mixed in solutions while engulfing dissolved or dispersed compounds together in a dose-dependent manner and shows unique optical characteristics because the dendrimer products exhibit aggregation-induced emission. Furthermore, the emission of the dendrimer causes considerable fluorescence resonance energy transfer (FRET) to GFP derivatives upon association. We could easily monitor assemblies by FRET states and particle sizes and have confirmed a stable presence in the buffer for at least a month. Further tracking of nanocapsules by fluorescence confirmed efficient uptake into some cancer cells. Nanocapsules based on GFP variants with or without a cell-surface-specific tag demonstrated that the tag improved the potential for specific targeted delivery. There were also indications that the nanocapsules became unstable after cellular uptake in the intracellular environment. We report here the simple preparation of traceable, stable, and biocompatible self-assembled nanocapsules as the basis for a versatile drug delivery system.

Modulating conductive filaments via thermally stable bilayer organic memristor

The basic unit of information in conductive bridge random access memory based on the redox mechanism of metal ions is physically stored in a conductive filament (CF). Therefore, the overall performance of the device is indissolubly related to the properties of such CFs, as unreliable performance often originates from unstable CFs behavior. However, accurately controlling the dissolution of CFs during device operation can be challenging due to their non-uniformity. This paper proposes a type of memristor based on a solid polymer electrolyte with a polyvinylpyrrolidone/polyvinyl alcohol composite layer structure. The properties of the composite layer are employed to regulate the number of CFs and the growth/fracture location, while the damage to the device by Joule heat is prevented. The device exhibits low operating voltage (0.5 V), stable switching conditions (2000 cycles), and a long holdup time (>3 × 104 s).

Limited value of EEG source imaging for decoding hand movement and imagery in youth with brain lesions

ABSTRACT Brain–computer interfaces based on electroencephalography (EEG) often exhibit unreliable performance in the classification of motor tasks. Recent research has shown that EEG source imaging (ESI) has the potential to outperform sensor domain approaches in various movement decoding tasks. However, ESI research to date has predominantly focused on the adult population, so its performance in youth with disabilities is unknown. In this study, we compared the offline classification performance of two ESI approaches (with and without modeling white matter conductivity anisotropy) to that of a sensor domain approach in the classification of left- versus right-hand movement execution and imagery tasks. Magnetic resonance images (MRI) were acquired from nine pediatric participants with brain lesions. Subsequently, cortical activity was recorded from 64 channels. MRI data were used to estimate participant-specific EEG sources. Various feature extraction and classification approaches were investigated in both sensor and source domains. Generally, ESI classification performance did not exceed chance levels and was statistically equivalent to sensor approaches except for isolated participants. However, ESI offered +9.61% improvement over the sensor domain (p = 0.031) in decoding motor execution in a participant with unilateral ventriculomegaly. Future research ought to delineate the specific task and participant characteristics which warrant the source domain approach.

Enspectrumix: Novel adaptive methodology for fault component extraction from vibration mixtures

Machine condition monitoring is important in industry to monitor machine health conditions and avoid unexpected machine breakdowns. Vibration mixtures contain sufficient machine health information, but they have lots of vibration interferences. Therefore, it is crucial to adaptively extract fault components from vibration mixtures. Blind-filter-based methods have been extensively studied to extract fault components and they use impulsiveness and cyclo-stationarity properties as prior knowledge to optimize signal processing parameters. The difficulties of state-of-the-art methods include that: (1) statistical indices (e.g., kurtosis, Gini index, correlated kurtosis, etc.) cannot simultaneously characterize impulsiveness and cyclo-stationarity properties so that blind-filter-based methods are always affected by impulsive noise and low-frequency components, resulting in unreliable performance for fault components extraction; (2) optimization of signal processing parameters is time-consuming; (3) band-in noise cannot be removed; (4) only a frequency band rather than several frequency bands can be adaptively selected, resulting in the incomplete fault components extraction. Instead of using the aforementioned prior knowledge, this paper reports a new perspective that use healthy vibration mixtures as prior knowledge to improve the existing blind-filter-based methods and provide more reliable fault components extraction for practical implementation in industry. Firstly, an ensemble difference spectrum-based new index is proposed to provide a completely new perspective for traditionally existing statistical indices. Subsequently, it is theoretically proved that the new index satisfies two new properties for quickly and better quantifying fault components. Next, a new methodology named Enspectrumix is designed to decompose a raw vibration mixture into sub-signals and use the newly designed index to adaptively select informative sub-signals existing in different frequency bands. The effectiveness and superiorities of the proposed Enspectrumix are validated by two real-world dataset cases. Moreover, the Enspectrumix can be extended to extract concerned difference components in many other domains.

Genomic tools for early selection among Thoroughbreds and Polo Argentino horses for practicing polo

The Polo Argentino (PA) horse is a recognized breed, developed originally by mixing crossbred and Thoroughbred (TB) horses to play polo. Early PA selection is difficult due to unreliable performance estimations. This study investigated the usefulness of genomic markers previously linked to morphological and functional traits as a tool for the early selection of PA. To this, we genotyped 520 PA and 30 TB horses using the Equine GGPArray (Illumina, n = 71,778 SNPs). Analyses included a genetic characterization of six genetic markers associated with behavioral (DRD4), muscular development (MSTN), and body size (LCORL, HMGA6, ZFAT, and LASP1) genes. Genetic differences in the DRD4, MSTN, and LCORL SNP were found between the two breeds, in the last two FST index between breeds was 0.13 and 0.6, respectively (p < 0.01). In DRD4, G allele was the more prevalent in PA (0.56 vs 0.45 in TB, p < 0.05), but no differences were observed between the genotypes associated with phenotypes. In MSTN, heterozygous genotypes were the most common in PA (48 %), with a significant decrease in AA (Hardy-Weinberg p < 0.05), suggesting a negative selection against it in polo horses. In body size, HMGA2 was monomorphic in all horses, while ZFAT and LASP1 SNP showed higher variability. Interestingly, 99 % of PA showed a TT genotype in LCORL (only 66 % in TB), demonstrating selection for smaller horses. Our results suggest that empirical selection in PA has generated an incipient genomic differentiation in discrete traits which could be used as a marker-assisted selection tool for early selection of polo horses.

Low-cost automated flat-sheet membrane casting: An open-source, advanced manufacturing approach

Novel membrane materials developed in research labs struggle to gain widespread industrial adoption due in part to insufficient reproducibility and unreliable performance data. Open-source hardware approaches, especially 3D printing, enable the democratization of automation within a laboratory setting, and in the context of membranes, can minimize the inherent variability associated with manual methods of membrane casting. In this study, the native hardware and firmware of an inexpensive, conventional 3D printer was extensively modified for the purpose of flat-sheet membrane casting. Replicate poly (ether-ether ketone) (PEEK) membranes were cast with a thickness coefficient of variation of ∼10 % using the modified device. Cast membranes were used to assess the importance of controlling shear rate by characterizing both intra- and inter-film variability. Statistical differences in pure water permeability were observed across tested shear rates, with distinct morphological changes occurring to the membrane substructure. Overall, the technology developed in this study is shown to be an extremely useful approach for improving the process of developing membranes at the bench-scale.

A two-step approach for deploying heterogeneous vessels and designing reliable schedule in liner shipping services

Schedule reliability is a crucial metric for measuring liner container shipping service. Due to the inherent uncertainties at sea and ports, schedule delays and unreliable on-time performance have long plagued shipping companies. In practice, the heterogeneity of container vessels also creates challenges for deployment and schedule optimization. To address the problem, we develop a two-step approach with two programming sub-models. In the first step, we tackle the challenges of ship deployment and scheduling by considering heterogeneous vessels and capturing the interarrival time between adjacent ports of call. In the second step, we use the conditional value-at-risk (CVaR) to address the uncertainty risk avoidance in schedule design. To solve the stochastic programming models, we employ a solution framework based on the outer linear approximation method and the risk measure method, where CVaR is specified as the risk measure. We illustrate the application of our model on a case study concerning a real route operated by COSCO Shipping Lines. The results show that the total cost of liner shipping service with heterogeneous fleet on the route can be reduce by 4.77% compared with homogeneous fleet. And the cost of risk with the reliable schedule can be decreased by 8.04% compared with the ideal schedule. The research conclusions can provide some useful decision-making references for shipping companies to deploy heterogeneous vessels and design reliable schedules.

Optimised design of downhole turbodrills with bending-torsional tilting blade

Deep dry hot rock mining drilling operations confronts several challenges including the high-temperature conditions in the downhole environment, the unreliable performance of surface drives, and the difficult drillability of geothermal reservoir rocks. As a result, the use of downhole motors in drilling operations becomes an imperative selection. The all-metal turbodrill presently stands as the sole downhole motor capable of effectively drilling in deep wells with temperatures exceeding 180°. Although it can withstand high temperature, the output torque is small. In this study, we focus on the optimised design of the turbodrill blades with the objective of enhancing the output torque, addressing the prevailing issue of limited output torque. First, the structural parameters and parametric design are determined based on turbine blade design theory. Utilizing the quintic polynomial method, a two-dimensional blade profile is generated. Subsequently, an output torque model applicable to unsteady flow conditions is derived using the moment of momentum theorem in its integral form; Utilizing the designed blade profile, the Unigraphics NX (UG) software is used to illustrate the combination of spatial twisting, bending and tilting of the blade for the combination of the modelling method, and a three-dimensional bending-torsional tilting blade was achieved. Subsequently, numerical analysis, employing Computational Fluid Dynamics (CFD) simulation software, was conducted to examine the impact of intricate spatial modelling on the blade's flow field parameters. The optimal parameters for blade spatial modelling were ascertained to be a twist lofting radius of 63 mm, a bending angle of 30°, and an inclination angle of 3°. The blade in the hydraulic efficiency is reduced by only 2.7% of the case, the output torque increased by 18.87%, improving the performance of the turbodrill, so that the turbodrill in the geothermal development drilling has a good application prospect.

Overcoming the Dilemma of In Vivo Stable Adhesion and Sustained Degradation by the Molecular Design of Polyurethane Adhesives for Bone Fracture Repair.

Bone adhesive is a promising candidate to revolutionize the clinical treatment of bone repairs. However, several drawbacks have limited its further clinical application, such as unreliable wet adhesive performance leading to fixation failure and poor biodegradability inhibiting bone tissue growth. By incorporating catechol groups and disulfide bonds into polyurethane (PU) molecules, an injectable and porous PU adhesive is developed with both superior wet adhesion and biodegradability to facilitate the reduction and fixation of comminuted fractures and the subsequent regeneration of bone tissue. The bone adhesive can be cured within a reasonable time acceptable to a surgeon, and then the wet bone adhesive strength is near 1.30MPa in 1h. Finally, the wet adhesive strength to the cortical bone will achieve about 1.70MPa, which is also five times more than nonresorbable poly(methyl methacrylate) bone cement. Besides, the cell culture experiments also indicate that the adhesives show excellent biocompatibility and osteogenic ability in vitro. Especially, it can degrade in vivo gradually and promote fracture healing in the rabbit iliac fracture model. These results demonstrate that this ingenious bone adhesive exhibits great potential in the treatment of comminuted fractures, providing fresh insights into the development of clinically applicable bone adhesives.

Novel nonparametric control charts for monitoring dispersion of count data

AbstractCount data are common in practice. Statistical process control for count data thus has attracted much attention in recent years. Most existing methods on this topic focus on the detection of mean shifts of count data based on parametric modeling. However, their assumed parametric models (e.g., the Poisson probability model) are often invalid in practice due mainly to the potential impact of some latent confounding risk factors, which would lead to unreliable performance of the related control charts. In addition, it is highly desirable and important to monitor the dispersion of count data when the Poisson probability model is invalid. To this end, new nonparametric cumulative sum control charts and their corresponding self‐starting versions are suggested in this paper for monitoring the dispersion of count data based on data categorization and categorical data analysis. Numerical results show that the proposed method can provide more effective and robust monitoring of count data in comparison with some representative existing methods. A real‐data example is used to demonstrate its implementation and application.

Drift–Diffusion Simulation of Intermediate Band Solar Cell: Effect of Intermediate Band Continuity Constraint

Self-consistent drift–diffusion model has been widely employed to simulate the device performance of intermediate band solar cell (IBSC) under practical device configuration. However, one of the remained issues in the drift–diffusion modeled-based works is the difficulty to reach the IB carrier continuity through the self-consistent manner. In most of the previous reports the constraints were relaxed or just partially satisfied; which render the unreliable performance results and misguide the device design strategy. In this work, in order to solve this issue and to validate our results, we performed extensive simulations to fully disclose the significant effect of the IB continuity constraints by taking InAs/GaAs quantum dot-based IBSC as a model device using the semiconductor modules in COMSOL Multiphysics combined with the Fortran codes. We found that under rigorous satisfaction of IB continuity constraint, the band potential profiles for the IBSC with either doped or nondoped IB under various light illumination conditions are nearly identical to those under the dark conditions. Moreover, from the simulated current–voltage curve dependence on the light concentration ratio, we found the device performance based on drift–diffusion under rigorous IB continuity constraint showed similar tendency to the features simulated based on detailed balance principle except the much-lowered power conversion efficiency. Our work demonstrated here, serves as an accurate and reliable IBSC device design approach toward better IB material screening, efficiency improvement, optical management, and extended application in the emerging field such as the perovskite material-based IBSC.

94 Characterizing Yearling Beef Steers Grazing on Smooth Bromegrass Pasture Using Global Positioning Technology

Abstract Extensive grazing systems often vary in pasture size and terrain, making it challenging for ranchers to accurately locate livestock through visual appraisal. Using satellite-based global positioning systems (GPS), accurate position data can be provided to track animal movements that could inform grazing management decisions. This study aimed to: 1) assess two modern GPS devices for functionality by calculating fix success rate, and 2) evaluate the relationship between average daily distance traveled and average daily weight gain (ADG) performance on yearling steers grazing smooth bromegrass pasture using the GPS devices. Yearling steers (n = 25) grazed a 10.8-ha smooth bromegrass pasture for 156 days from May to September 2021. The temperature ranged from 4.1 to 38.8ºC with an average 22.7ºC and received 490 mm of rainfall. Cattle ADG was 0.65 ± 0.13 kg/d. Each steer was assigned two GPS collars for 28 days to compare two types of sensors. The sensors evaluated were i-gotU GT-600 travel loggers (GU) and Yabby LoRaWAN GPS sensors (YB) utilizing the IoT (Internet of Things) technology. The YB sensor possessed real-time monitoring and remote data access, while data from the GU sensor needed to be removed from the animals and manually downloaded. The YB sensors had a manufacturer declared finest positioning accuracy of 4.6 m, while that of the GU sensor being 9.1 m. With a 10-min sampling rate, battery life of the YB sensors lasted 4 weeks, while the GU sensors demonstrated unreliable battery performance that ranged between 1 to 4 weeks. Fixed success rate (FSR) was calculated for each sensor. Using PROC MIXED of SAS (v 9.4), on average the GU sensors had a 67.1% FSR and the YB sensors observed an 89.9% FSR (P &amp;lt; 0.0001). The YB sensor outweighed the GU sensor on battery life and FSR, thus was utilized throughout the duration of the grazing season. Individual animal traveled distance was calculated after converting the GPS coordinates in decimal degrees to Universal Transverse Mercator coordinates (WGS 1984 UTM zone 14N) in meters. The Euclidean distance between consecutive coordinates was calculated and summed daily to represent the daily traveled distance of an animal. Using all coordinates recorded from YB, steers traveled 2,718 ± 809 m/d. Statistical analysis was conducted using PROC CORR of SAS to determine the correlation between the daily averages of distance traveled and ADG. All values were considered significant α ≤ 0.05. There was no correlation observed between distance traveled and ADG (P ≥ 0.76). Results demonstrate that modern GPS sensors utilizing IoT technology have benefits over traditional GPS sensors and can be used as a tool for extensive grazing livestock tracking.

Towards Ensuring Software Interoperability Between Deep Learning Frameworks

Abstract With the widespread of systems incorporating multiple deep learning models, ensuring interoperability between target models has become essential. However, due to the unreliable performance of existing model conversion solutions, it is still challenging to ensure interoperability between the models developed on different deep learning frameworks. In this paper, we propose a systematic method for verifying interoperability between pre- and post-conversion deep learning models based on the validation and verification approach. Our proposed method ensures interoperability by conducting a series of systematic verifications from multiple perspectives. The case study confirmed that our method successfully discovered the interoperability issues that have been reported in deep learning model conversions.

Model-informed precision dosing of vancomycin for rapid achievement of target area under the concentration-time curve: A simulation study.

In this study, we aimed to evaluate limited sampling strategies for achieving the therapeutic ranges of the area under the concentration-time curve (AUC) of vancomycin on the first and second day (AUC0-24 , AUC24-48 , respectively) of therapy. A virtual population of 1,000 individuals was created using a population pharmacokinetic (popPK) model, which was validated and incorporated into our model-informed precision dosing (MIPD) tool. The results were evaluated using six additional popPK models selected based on a study design of prospective or retrospective data collection with sufficient concentrations. Bayesian forecasting was performed to evaluate the probability of achieving the therapeutic range of AUC, defined as a ratio of estimated/reference AUC within 0.8-1.2. The Bayesian posterior probability of achieving the AUC24-48 range increased from 51.3% (a priori probability) to 77.5% after using two-point sampling at the trough and peak on the first day. Sampling on the first day also yielded a higher Bayesian posterior probability (86.1%) of achieving the AUC0-24 range compared to the a priori probability of 60.1%. The Bayesian posterior probability of achieving the AUC at steady state (AUCSS ) range by sampling on the first or second day decreased with decreased kidney function. We demonstrated that second-day trough and peak sampling provided accurate AUC24-48 , and first-day sampling may assist in rapidly achieving therapeutic AUC24-48 , although the AUCSS should be re-estimated in patients with reduced kidney function owing to its unreliable predictive performance.

Graphene Quantum Dots as an Oxygen Reservoir for Topotactic Phase Transition‐Based Memristive Devices

AbstractA novel class of transition metal oxides, capable of reversible topotactic phase transition between the oxygen‐deficient brownmillerite and oxygen‐rich perovskite, has emerged as a promising material for memristive and magnetoelectric devices. However, the absence of a local oxygen source in the device structure necessitates an oxygen exchange process between the surrounding atmosphere and the switching layer during operation, which can lead to unreliable device performance. In this study, graphene quantum dots (GQDs) are introduced into a SrFe0.5Co0.5Ox memristive device as an oxygen reservoir for the nanoscale topotactic redox process. The SrFe0.5Co0.5Ox memristive devices with GQDs exhibit reliable resistive switching performance compared to SrFe0.5Co0.5Ox devices without GQDs. To understand the effect of GQDs on the device structure, a pulse endurance test is carried out in a high vacuum. The devices with GQDs show rather good endurance behavior, while devices without GQDs exhibit endurance failure. These results provide a deeper understanding of the potential use of GQDs in enhancing the performance of SrFe0.5Co0.5Ox memristive devices, with implications for tuning nanoscale topotactic phase transition for multi‐functional properties.

HoaKV: High-Performance KV Store Based on the Hot-Awareness in Mixed Workloads

Key–value (KV) stores based on the LSM-tree have become the mainstream of contemporary store engines, but there are problems with high write and read amplification. Moreover, the real-world workload has a high data skew, and the existing KV store lacks hot-awareness, leading to its unreliable and poor performance on the highly skewed real-world workload. In this paper, we propose HoaKV, which unifies the key design ideas of hot issues, KV separation, and hybrid indexing technology in a system. Specifically, HoaKV uses the heat differentiation in KV pairs to manage the hot data and the cold data and conducts real-time dynamic adjustment data classification management. It also uses partial KV separation technology to manage differential KV pairs for large and small KV pairs in the cold data. In addition, HoaKV uses hybrid indexing technology to index the hot data and the cold data, respectively, to improve the performance of reading, writing, and scanning at the same time. In the mixed read and write workloads experments show that HoaKV performs significantly better than several state-of-the-art KV store technologies such as LevelDB, RocksDB, PebblesDB, and WiscKey.

Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking

Graph neural networks (GNNs) have been widely used in molecular property prediction, but explaining their black-box predictions is still a challenge. Most existing explanation methods for GNNs in chemistry focus on attributing model predictions to individual nodes, edges or fragments that are not necessarily derived from a chemically meaningful segmentation of molecules. To address this challenge, we propose a method named substructure mask explanation (SME). SME is based on well-established molecular segmentation methods and provides an interpretation that aligns with the understanding of chemists. We apply SME to elucidate how GNNs learn to predict aqueous solubility, genotoxicity, cardiotoxicity and blood–brain barrier permeation for small molecules. SME provides interpretation that is consistent with the understanding of chemists, alerts them to unreliable performance, and guides them in structural optimization for target properties. Hence, we believe that SME empowers chemists to confidently mine structure-activity relationship (SAR) from reliable GNNs through a transparent inspection on how GNNs pick up useful signals when learning from data.