Universal adversarial attacks on deep neural networks for medical image classification.

BACKGROUND: Deep neural networks (DNNs) are widely investigated in medical image classification to achieve automated support for clinical diagnosis. It is necessary to evaluate the robustness of medical DNN tasks against adversarial attacks, as high-stake decision-making will be made based on the diagnosis. Several previous studies have considered simple adversarial attacks. However, the vulnerability of DNNs to more realistic and higher risk attacks, such as universal adversarial perturbation (UAP), which is a single perturbation that can induce DNN failure in most classification tasks has not been evaluated yet. METHODS: We focus on three representative DNN-based medical image classification tasks (i.e., skin cancer, referable diabetic retinopathy, and pneumonia classifications) and investigate their vulnerability to the seven model architectures of UAPs. RESULTS: We demonstrate that DNNs are vulnerable to both nontargeted UAPs, which cause a task failure resulting in an input being assigned an incorrect class, and to targeted UAPs, which cause the DNN to classify an input into a specific class. The almost imperceptible UAPs achieved > 80% success rates for nontargeted and targeted attacks. The vulnerability to UAPs depended very little on the model architecture. Moreover, we discovered that adversarial retraining, which is known to be an effective method for adversarial defenses, increased DNNs' robustness against UAPs in only very few cases. CONCLUSION: Unlike previous assumptions, the results indicate that DNN-based clinical diagnosis is easier to deceive because of adversarial attacks. Adversaries can cause failed diagnoses at lower costs (e.g., without consideration of data distribution); moreover, they can affect the diagnosis. The effects of adversarial defenses may not be limited. Our findings emphasize that more careful consideration is required in developing DNNs for medical imaging and their practical applications.

Difficulty in inferring microbial community structure based on co-occurrence network approaches.

BACKGROUND: Co-occurrence networks-ecological associations between sampled populations of microbial communities inferred from taxonomic composition data obtained from high-throughput sequencing techniques-are widely used in microbial ecology. Several co-occurrence network methods have been proposed. Co-occurrence network methods only infer ecological associations and are often used to discuss species interactions. However, validity of this application of co-occurrence network methods is currently debated. In particular, they simply evaluate using parametric statistical models, even though microbial compositions are determined through population dynamics. RESULTS: We comprehensively evaluated the validity of common methods for inferring microbial ecological networks through realistic simulations. We evaluated how correctly nine widely used methods describe interaction patterns in ecological communities. Contrary to previous studies, the performance of the co-occurrence network methods on compositional data was almost equal to or less than that of classical methods (e.g., Pearson's correlation). The methods described the interaction patterns in dense and/or heterogeneous networks rather inadequately. Co-occurrence network performance also depended upon interaction types; specifically, the interaction patterns in competitive communities were relatively accurately predicted while those in predator-prey (parasitic) communities were relatively inadequately predicted. CONCLUSIONS: Our findings indicated that co-occurrence network approaches may be insufficient in interpreting species interactions in microbiome studies. However, the results do not diminish the importance of these approaches. Rather, they highlight the need for further careful evaluation of the validity of these much-used methods and the development of more suitable methods for inferring microbial ecological networks.

PREvaIL, an integrative approach for inferring catalytic residues using sequence, structural, and network features in a machine-learning framework.

Determining the catalytic residues in an enzyme is critical to our understanding the relationship between protein sequence, structure, function, and enhancing our ability to design novel enzymes and their inhibitors. Although many enzymes have been sequenced, and their primary and tertiary structures determined, experimental methods for enzyme functional characterization lag behind. Because experimental methods used for identifying catalytic residues are resource- and labor-intensive, computational approaches have considerable value and are highly desirable for their ability to complement experimental studies in identifying catalytic residues and helping to bridge the sequence-structure-function gap. In this study, we describe a new computational method called PREvaIL for predicting enzyme catalytic residues. This method was developed by leveraging a comprehensive set of informative features extracted from multiple levels, including sequence, structure, and residue-contact network, in a random forest machine-learning framework. Extensive benchmarking experiments on eight different datasets based on 10-fold cross-validation and independent tests, as well as side-by-side performance comparisons with seven modern sequence- and structure-based methods, showed that PREvaIL achieved competitive predictive performance, with an area under the receiver operating characteristic curve and area under the precision-recall curve ranging from 0.896 to 0.973 and from 0.294 to 0.523, respectively. We demonstrated that this method was able to capture useful signals arising from different levels, leveraging such differential but useful types of features and allowing us to significantly improve the performance of catalytic residue prediction. We believe that this new method can be utilized as a valuable tool for both understanding the complex sequence-structure-function relationships of proteins and facilitating the characterization of novel enzymes lacking functional annotations.

MAPLE 2.3.0: an improved system for evaluating the functionomes of genomes and metagenomes.

MAPLE is an automated system for inferring the potential comprehensive functions harbored by genomes and metagenomes. To reduce runtime in MAPLE analyzing the massive amino acid datasets of over 1 million sequences, we improved it by adapting the KEGG automatic annotation server to use GHOSTX and verified no substantial difference in the MAPLE results between the original and new implementations.

Limitations of a metabolic network-based reverse ecology method for inferring host-pathogen interactions.

BACKGROUND: Host-pathogen interactions are important in a wide range of research fields. Given the importance of metabolic crosstalk between hosts and pathogens, a metabolic network-based reverse ecology method was proposed to infer these interactions. However, the validity of this method remains unclear because of the various explanations presented and the influence of potentially confounding factors that have thus far been neglected. RESULTS: We re-evaluated the importance of the reverse ecology method for evaluating host-pathogen interactions while statistically controlling for confounding effects using oxygen requirement, genome, metabolic network, and phylogeny data. Our data analyses showed that host-pathogen interactions were more strongly influenced by genome size, primary network parameters (e.g., number of edges), oxygen requirement, and phylogeny than the reserve ecology-based measures. CONCLUSION: These results indicate the limitations of the reverse ecology method; however, they do not discount the importance of adopting reverse ecology approaches altogether. Rather, we highlight the need for developing more suitable methods for inferring host-pathogen interactions and conducting more careful examinations of the relationships between metabolic networks and host-pathogen interactions.

Systematic Protein Level Regulation via Degradation Machinery Induced by Genotoxic Drugs.

In this study we monitored protein dynamics in response to cisplatin, 5-fluorouracil, and irinotecan with different concentrations and administration modes using "reverse-phase" protein arrays (RPPAs) in order to gain comprehensive insight into the protein dynamics induced by genotoxic drugs. Among 666 protein time-courses, 38% exhibited an increasing trend, 32% exhibited a steady decrease, and 30% fluctuated within 24 h after drug exposure. We analyzed almost 12,000 time-course pairs of protein levels based on the geometrical similarity by correlation distance (dCor). Twenty-two percent of the pairs showed dCor > 0.8, which indicates that each protein of the pair had similar dynamics. These trends were disrupted by a proteasome inhibitor, MG132, suggesting that the protein degradation system was activated in response to the drugs. Among the pairs with high dCor, the average dCor of pairs with apoptosis-related protein was significantly higher than those without, indicating that regulation of protein levels was induced by the drugs. These results suggest that the levels of numerous functionally distinct proteins may be regulated by common degradation machinery induced by genotoxic drugs.

The moral machine experiment on large language models.

As large language models (LLMs) have become more deeply integrated into various sectors, understanding how they make moral judgements has become crucial, particularly in the realm of autonomous driving. This study used the moral machine framework to investigate the ethical decision-making tendencies of prominent LLMs, including GPT-3.5, GPT-4, PaLM 2 and Llama 2, to compare their responses with human preferences. While LLMs' and humans' preferences such as prioritizing humans over pets and favouring saving more lives are broadly aligned, PaLM 2 and Llama 2, especially, evidence distinct deviations. Additionally, despite the qualitative similarities between the LLM and human preferences, there are significant quantitative disparities, suggesting that LLMs might lean toward more uncompromising decisions, compared with the milder inclinations of humans. These insights elucidate the ethical frameworks of LLMs and their potential implications for autonomous driving.

Identification of chemogenomic features from drug-target interaction networks using interpretable classifiers.

MOTIVATION: Drug effects are mainly caused by the interactions between drug molecules and their target proteins including primary targets and off-targets. Identification of the molecular mechanisms behind overall drug-target interactions is crucial in the drug design process. RESULTS: We develop a classifier-based approach to identify chemogenomic features (the underlying associations between drug chemical substructures and protein domains) that are involved in drug-target interaction networks. We propose a novel algorithm for extracting informative chemogenomic features by using L(1) regularized classifiers over the tensor product space of possible drug-target pairs. It is shown that the proposed method can extract a very limited number of chemogenomic features without loosing the performance of predicting drug-target interactions and the extracted features are biologically meaningful. The extracted substructure-domain association network enables us to suggest ligand chemical fragments specific for each protein domain and ligand core substructures important for a wide range of protein families. AVAILABILITY: Softwares are available at the supplemental website. CONTACT: yamanishi@bioreg.kyushu-u.ac.jp SUPPLEMENTARY INFORMATION: Datasets and all results are available at http://cbio.ensmp.fr/~yyamanishi/l1binary/ .

Revisiting the political biases of ChatGPT.

Although ChatGPT promises wide-ranging applications, there is a concern that it is politically biased; in particular, that it has a left-libertarian orientation. Nevertheless, following recent trends in attempts to reduce such biases, this study re-evaluated the political biases of ChatGPT using political orientation tests and the application programming interface. The effects of the languages used in the system as well as gender and race settings were evaluated. The results indicate that ChatGPT manifests less political bias than previously assumed; however, they did not entirely dismiss the political bias. The languages used in the system, and the gender and race settings may induce political biases. These findings enhance our understanding of the political biases of ChatGPT and may be useful for bias evaluation and designing the operational strategy of ChatGPT.

Adversarial attacks on voter model dynamics in complex networks.

This paper investigates adversarial attacks conducted to distort voter model dynamics in complex networks. Specifically, a simple adversarial attack method is proposed to hold the state of opinions of an individual closer to the target state in the voter model dynamics. This indicates that even when one opinion is the majority the vote outcome can be inverted (i.e., the outcome can lean toward the other opinion) by adding extremely small (hard-to-detect) perturbations strategically generated in social networks. Adversarial attacks are relatively more effective in complex (large and dense) networks. These results indicate that opinion dynamics can be unknowingly distorted.

Nested structure acquired through simple evolutionary process.

Nested structure, which is non-random, controls cooperation dynamics and biodiversity in plant-animal mutualistic networks. This structural pattern has been explained in a static (non-growth) network models. However, evolutionary processes might also influence the formation of such a structural pattern. We thereby propose an evolving network model for plant-animal interactions and show that non-random patterns such as nested structure and heterogeneous connectivity are both qualitatively and quantitatively predicted through simple evolutionary processes. This finding implies that network models can be simplified by considering evolutionary processes, and also that another explanation exists for the emergence of non-random patterns and might provide more comprehensible insights into the formation of plant-animal mutualistic networks from the evolutionary perspective.

Analysis and prediction of nutritional requirements using structural properties of metabolic networks and support vector machines.

Properties of graph representation of genome scale metabolic networks have been extensively studied. However, the relationship between these structural properties and functional properties of the networks are still very unclear. In this paper, we focus on nutritional requirements of organisms as a functional property and study the relationship with structural properties of a graph representation of metabolic networks. In order to examine the relationship, we study to what extent the nutritional requirements can be predicted by using support vector machines from structural properties, which include degree exponent, edge density, clustering coefficient, degree centrality, closeness centrality, betweenness centrality and eigenvector centrality. Furthermore, we study which properties are influential to the nutritional requirements.

Revisiting the hypothesis of an energetic barrier to genome complexity between eukaryotes and prokaryotes.

The absence of genome complexity in prokaryotes, being the evolutionary precursors to eukaryotic cells comprising all complex life (the prokaryote-eukaryote divide), is a long-standing question in evolutionary biology. A previous study hypothesized that the divide exists because prokaryotic genome size is constrained by bioenergetics (prokaryotic power per gene or genome being significantly lower than eukaryotic ones). However, this hypothesis was evaluated using a relatively small dataset due to lack of data availability at the time, and is therefore controversial. Accordingly, we constructed a larger dataset of genomes, metabolic rates, cell sizes and ploidy levels to investigate whether an energetic barrier to genome complexity exists between eukaryotes and prokaryotes while statistically controlling for the confounding effects of cell size and phylogenetic signals. Notably, we showed that the differences in bioenergetics between prokaryotes and eukaryotes were less significant than those previously reported. More importantly, we found a limited contribution of power per genome and power per gene to the prokaryote-eukaryote dichotomy. Our findings indicate that the prokaryote-eukaryote divide is hard to explain from the energetic perspective. However, our findings may not entirely discount the traditional hypothesis; in contrast, they indicate the need for more careful examination.

Vulnerability of deep neural networks for detecting COVID-19 cases from chest X-ray images to universal adversarial attacks.

Owing the epidemic of the novel coronavirus disease 2019 (COVID-19), chest X-ray computed tomography imaging is being used for effectively screening COVID-19 patients. The development of computer-aided systems based on deep neural networks (DNNs) has become an advanced open source to rapidly and accurately detect COVID-19 cases because the need for expert radiologists, who are limited in number, forms a bottleneck for screening. However, thus far, the vulnerability of DNN-based systems has been poorly evaluated, although realistic and high-risk attacks using universal adversarial perturbation (UAP), a single (input image agnostic) perturbation that can induce DNN failure in most classification tasks, are available. Thus, we focus on representative DNN models for detecting COVID-19 cases from chest X-ray images and evaluate their vulnerability to UAPs. We consider non-targeted UAPs, which cause a task failure, resulting in an input being assigned an incorrect label, and targeted UAPs, which cause the DNN to classify an input into a specific class. The results demonstrate that the models are vulnerable to non-targeted and targeted UAPs, even in the case of small UAPs. In particular, the 2% norm of the UAPs to the average norm of an image in the image dataset achieves >85% and >90% success rates for the non-targeted and targeted attacks, respectively. Owing to the non-targeted UAPs, the DNN models judge most chest X-ray images as COVID-19 cases. The targeted UAPs allow the DNN models to classify most chest X-ray images into a specified target class. The results indicate that careful consideration is required in practical applications of DNNs to COVID-19 diagnosis; in particular, they emphasize the need for strategies to address security concerns. As an example, we show that iterative fine-tuning of DNN models using UAPs improves the robustness of DNN models against UAPs.

The brain-derived neurotrophic factor Val66Met polymorphism increases segregation of structural correlation networks in healthy adult brains.

BACKGROUND: Although structural correlation network (SCN) analysis is an approach to evaluate brain networks, the neurobiological interpretation of SCNs is still problematic. Brain-derived neurotrophic factor (BDNF) is well-established as a representative protein related to neuronal differentiation, maturation, and survival. Since a valine-to-methionine substitution at codon 66 of the BDNF gene (BDNF Val66Met single nucleotide polymorphism (SNP)) is well-known to have effects on brain structure and function, we hypothesized that SCNs are affected by the BDNF Val66Met SNP. To gain insight into SCN analysis, we investigated potential differences between BDNF valine (Val) homozygotes and methionine (Met) carriers in the organization of their SCNs derived from inter-regional cortical thickness correlations. METHODS: Forty-nine healthy adult subjects (mean age = 41.1 years old) were divided into two groups according to their genotype (n: Val homozygotes = 16, Met carriers = 33). We obtained regional cortical thickness from their brain T1 weighted images. Based on the inter-regional cortical thickness correlations, we generated SCNs and used graph theoretical measures to assess differences between the two groups in terms of network integration, segregation, and modularity. RESULTS: The average local efficiency, a measure of network segregation, of BDNF Met carriers' network was significantly higher than that of the Val homozygotes' (permutation p-value = 0.002). Average shortest path lengths (a measure of integration), average local clustering coefficient (another measure of network segregation), small-worldness (a balance between integration and segregation), and modularity (a representative measure for modular architecture) were not significantly different between group (permutation p-values ≧ 0.01). DISCUSSION AND CONCLUSION: Our results suggest that the BDNF Val66Met polymorphism may potentially influence the pattern of brain regional morphometric (cortical thickness) correlations. Comparing networks derived from inter-regional cortical thickness correlations, Met carrier SCNs have denser connections with neighbors and are more distant from random networks than Val homozygote networks. Thus, it may be necessary to consider potential effects of BDNF gene mutations in SCN analyses. This is the first study to demonstrate a difference between Val homozygotes and Met carriers in brain SCNs.

HSEpred: predict half-sphere exposure from protein sequences.

MOTIVATION: Half-sphere exposure (HSE) is a newly developed two-dimensional solvent exposure measure. By conceptually separating an amino acid's sphere in a protein structure into two half spheres which represent its distinct spatial neighborhoods in the upward and downward directions, the HSE-up and HSE-down measures show superior performance compared with other measures such as accessible surface area, residue depth and contact number. However, currently there is no existing method for the prediction of HSE measures from sequence data. RESULTS: In this article, we propose a novel approach to predict the HSE measures and infer residue contact numbers using the predicted HSE values, based on a well-prepared non-homologous protein structure dataset. In particular, we employ support vector regression (SVR) to quantify the relationship between HSE measures and protein sequences and evaluate its prediction performance. We extensively explore five sequence-encoding schemes to examine their effects on the prediction performance. Our method could achieve the correlation coefficients of 0.72 and 0.68 between the predicted and observed HSE-up and HSE-down measures, respectively. Moreover, contact number can be accurately predicted by the summation of the predicted HSE-up and HSE-down values, which has further enlarged the application of this method. The successful application of SVR approach in this study suggests that it should be more useful in quantifying the protein sequence-structure relationship and predicting the structural property profiles from protein sequences. AVAILABILITY: The prediction webserver and supplementary materials are accessible at http://sunflower.kuicr.kyoto-u.ac.jp/~sjn/hse/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Network resilience of mutualistic ecosystems and environmental changes: an empirical study.

It is theorized that a mutualistic ecosystem's resilience against perturbations (e.g. species extinction) is determined by a single macroscopic parameter (network resilience), calculable from the network. Given that such perturbations occur owing to environmental changes (e.g. climate change and human impact), it has been predicted that mutualistic ecosystems that exist despite extensive environmental changes exhibit higher network resilience; however, such a prediction has not been confirmed using real-world data. Thus, in this study, the effects of climate change velocity and human activities on mutualistic network resilience were investigated. A global dataset of plant-animal mutualistic networks was used, and spatial analysis was performed to examine the effects. Moreover, the potential confounding effects of network size, current climate and altitude were statistically controlled. It was demonstrated that mutualistic network resilience was globally influenced by warming velocity and human impact, in addition to current climate. Specifically, pollination network resilience increased in response to human impact, and seed-dispersal network resilience increased with warming velocity. The effect of environmental changes on network resilience for plants was remarkable. The results confirmed the prediction obtained based on the theory and imply that real-world mutualistic networks have a structure that increases ecosystem resilience against environmental changes. These findings will enhance the understanding of ecosystem resilience.

Decomposing the effects of ocean environments on predator-prey body-size relationships in food webs.

Body-size relationships between predators and their prey are important in ecological studies because they reflect the structure and function of food webs. Inspired by studies on the impact of global warming on food webs, the effects of temperature on body-size relationships have been widely investigated; however, the impact of environmental factors on body-size relationships has not been fully evaluated because climate warming affects various ocean environments. Thus, here, we comprehensively investigated the effects of ocean environments and predator-prey body-size relationships by integrating a large-scale dataset of predator-prey body-size relationships in marine food webs with global oceanographic data. We showed that various oceanographic parameters influence prey size selection. In particular, oxygen concentration, primary production and salinity, in addition to temperature, significantly alter body-size relationships. Furthermore, we demonstrated that variability (seasonality) of ocean environments significantly affects body-size relationships. The effects of ocean environments on body-size relationships were generally remarkable for small body sizes, but were also significant for large body sizes and were relatively weak for intermediate body sizes, in the cases of temperature seasonality, oxygen concentration and salinity variability. These findings break down the complex effects of ocean environments on body-size relationships, advancing our understanding of how ocean environments influence the structure and functioning of food webs.

Large-scale aggregation analysis of eukaryotic proteins reveals an involvement of intrinsically disordered regions in protein folding.

A subset of the proteome is prone to aggregate formation, which is prevented by chaperones in the cell. To investigate whether the basic principle underlying the aggregation process is common in prokaryotes and eukaryotes, we conducted a large-scale aggregation analysis of ~500 cytosolic budding yeast proteins using a chaperone-free reconstituted translation system, and compared the obtained data with that of ~3,000 Escherichia coli proteins reported previously. Although the physicochemical properties affecting the aggregation propensity were generally similar in yeast and E. coli proteins, the susceptibility of aggregation in yeast proteins were positively correlated with the presence of intrinsically disordered regions (IDRs). Notably, the aggregation propensity was not significantly changed by a removal of IDRs in model IDR-containing proteins, suggesting that the properties of ordered regions in these proteins are the dominant factors for aggregate formation. We also found that the proteins with longer IDRs were disfavored by E. coli chaperonin GroEL/ES, whereas both bacterial and yeast Hsp70/40 chaperones have a strong aggregation-prevention effect even for proteins possessing IDRs. These results imply that a key determinant to discriminate the eukaryotic proteomes from the prokaryotic proteomes in terms of protein folding would be the attachment of IDRs.

Brain structural connectivity and neuroticism in healthy adults.

Understanding the neural correlates of the neurotic brain is important because neuroticism is a risk factor for the development of psychopathology. We examined the correlation between brain structural networks and neuroticism based on NEO Five-Factor Inventory (NEO-FFI) scores. Fifty-one healthy participants (female, n = 18; male, n = 33; mean age, 38.5 ± 11.7 years) underwent the NEO-FFI test and magnetic resonance imaging (MRI), including diffusion tensor imaging and 3D T1WI. Using MRI data, for each participant, we constructed whole-brain interregional connectivity matrices by deterministic tractography and calculated the graph theoretical network measures, including the characteristic path length, global clustering coefficient, small-worldness, and betweenness centrality (BET) in 83 brain regions from the Desikan-Killiany atlas with subcortical segmentation using FreeSurfer. In relation to the BET, neuroticism score had a negative correlation in the left isthmus cingulate cortex, left superior parietal, left superior temporal, right caudal middle frontal, and right entorhinal cortices, and a positive correlation in the bilateral frontal pole, left caudal anterior cingulate cortex, and left fusiform gyrus. No other measurements showed significant correlations. Our results imply that the brain regions related to neuroticism exist in various regions, and that the neuroticism trait is likely formed as a result of interactions among these regions. This work was supported by a Grant-in-Aid for Scientific Research on Innovative Areas (Comprehensive Brain Science Network) from the Ministry of Education, Science, Sports and Culture of Japan.