Search

Poster presented at the Colloque "Artificial Intelligence and Health", November 25th, 2019, Kremlin-Bic��tre, organized by Universit�� Paris-Saclay and Institut DATAIATitle:Prediction of antibiotic resistance from genomic data using machine learning and deep learning approachesAuthors:Valentin Laroche, Edithe Selwa and Bogdan I. IorgaInstitut de Chimie des Substances Naturelles, CNRS UPR 2301, Universit�� Paris-Saclay, Gif-sur-YvetteABSTRACTIntroduction:The availability of increasing amounts of genomic and biomedical data makes machine learning and deep learning the methods of choice for the prediction of antibiotic resistance, a worldwide health problem of utmost importance. However, several strong limitations have hampered the development of this field.Materials and methods:Input datasets were prepared from 390 genomes of Pseudomonas aeruginosa with known minimum inhibitory concentration (MIC) values for 4 antibiotics, using three approaches: knowledge-based (output of ResMiner, a bacterial genome analysis tool developed in our group), n-grams and k-mers. Dataset enrichment was performed with software publicly available (SMOTE, ROSE) or developed in our group (KNOTE).Results and discussion:Three different encoding modes of genomic data, each one with its advantages and limitations, were used for the preparation of input datasets. The performance of several machine learning and deep learning approaches was evaluated using these datasets. The best performance was obtained with Convolutional Neural Network approach and knowledge-based dataset encoding. However, a clear limitation was the unequal distribution of classes leading to imbalanced datasets. The dataset enrichment performed with publicly available software (SMOTE, ROSE) provided only little improvement in the prediction accuracy, due to errors introduced during enrichment. In these conditions, we have developed a new tool called KNOTE (KNowledge-based Oversampling TEchnique) for the error-free enrichment of imbalanced data from genomic and biomedical studies, to be used in machine learning or deep learning predictions. The use of KNOTE-enriched datasets provided up to 95-98% accuracy in the prediction of antibiotic resistance.Conclusion:We propose a very efficient knowledge-based approach for input data encoding or dataset enrichment to overcome the intrinsic limitations (small amount of experimental samples compared with the number of features, imbalanced datasets) encountered with the use of genomic and biomedical data in machine learning or deep learning predictive models.

International audience; The catalytic domain of the adenyl cyclase (AC) toxin from Bordetella pertussis is activated by interaction with calmodulin (CaM), resulting in cAMP overproduction in the infected cell. In the X-ray crystallographic structure of the complex between AC and the C terminal lobe of CaM, the toxin displays a markedly elongated shape. As for the structure of the isolated protein, experimental results support the hypothesis that more globular conformations are sampled, but information at atomic resolution is still lacking. Here, we use temperature-accelerated molecular dynamics (TAMD) simulations to generate putative all-atom models of globular conformations sampled by CaM-free AC. As collective variables, we use centers of mass coordinates of groups of residues selected from the analysis of standard molecular dynamics (MD) simulations. Results show that TAMD allows extended conformational sampling and generates AC conformations that are more globular than in the complexed state. These structures are then refined via energy minimization and further unrestrained MD simulations to optimize inter-domain packing interactions, thus resulting in the identification of a set of hydrogen bonds present in the globular conformations.

International audience; Adenylyl cyclase (AC) toxin is an essential toxin that allows Bordetella pertussis to invade eukaryotic cells, where it is activated after binding to calmodulin (CaM). Based on the crystal structure of the AC catalytic domain in complex with the C-terminal half of CaM (C-CaM), our previous molecular dynamics simulations (Selwa, E., Laine, E., and Malliavin, T. (2012) Differential role of calmodulin and calcium ions in the stabilization of the catalytic domain of adenyl cyclase CyaA from Bordetella pertussis. Proteins 80, 1028–1040) suggested that three residues (i.e. Arg(338), Asn(347), and Asp(360)) might be important for stabilizing the AC/CaM interaction. These residues belong to a loop-helix-loop motif at the C-terminal end of AC, which is located at the interface between CaM and the AC catalytic loop. In the present study, we conducted the in silico and in vitro characterization of three AC variants, where one (Asn(347); ACm1A), two (Arg(338) and Asp(360); ACm2A), or three residues (Arg(338), Asn(347), and Asp(360); ACm3A) were substituted with Ala. Biochemical studies showed that the affinities of ACm1A and ACm2A for CaM were not affected significantly, whereas that of ACm3A was reduced dramatically. To understand the effects of these modifications, molecular dynamics simulations were performed based on the modified proteins. The molecular dynamics trajectories recorded for the ACm3AC-CaM complex showed that the calcium-binding loops of C-CaM exhibited large fluctuations, which could be related to the weakened interaction between ACm3A and its activator. Overall, our results suggest that the loop-helix-loop motif at the C-terminal end of AC is crucial during CaM binding for stabilizing the AC catalytic loop in an active configuration.

The protein CyaA is an important virulence factor of Bordetella pertussis involved in the whooping cough disease. The AC domain of CyaA is activated in eukaryotic host cells as an adenyl-cyclase by the ubiquitous and calcium sensitive protein calmodulin (CaM). This activation transformes the ATP into cAMP in an uncontrolled way, leading to severe cellular dysfunction. Only the structure of AC complexed with calcium-loaded calmodulin has been resolved by X-ray crystallography. Starting from this structure, molecular dynamics simulations of AC free and in complex with calmodulin enabled to characterize the effect of the calmodulin and calcium ions on the conformational plasticity of the complex. Conformational tendencies of free AC were also studied. The joint analysis of energetic influences and hydrogen bonds revealed a network of interactions between AC and calmodulin, in which three key residues, that may play an allosteric role on the catalytic activity of the adenyl-cyclase, were modified by site-directed mutagenesis. These experimental tests have led to the identification of an allosteric region which assures the communication of the conformational transition information between the binding site of the calmodulin and the catalytic site. A more thorough conformational exploration of the unbound AC state was undertaken by an innovative temperature accelerated molecular dynamics (TAMD) method. It led to a prediction of some sampled conformations of AC in its unbound state. These predictions could be used in the future to stabilize the free state and to facilitate the experimental study of its structure; La protéine CyaA est un facteur de virulence majeur de Bordetella pertussis, impliqué dans la maladie de la coqueluche. Le domaine catalytique AC de CyaA est directement transféré dans la cellule hôte eucaryote, où il est activé comme adénylcyclase par la calmoduline, une protéine ubiquitaire et sensible aux ions calcium. Ainsi, AC transforme l'ATP en AMPc de manière incontrôlée, ce qui conduit à des dérèglements cellulaires. Seule la structure de AC complexé à la calmoduline chargée d'ions calcium avait été résolue par cristallographie aux rayons X. À partir de cette structure, des simulations de dynamique moléculaire de AC libre, et en complexe avec la calmoduline nous ont permis de caractériser l'effet de la calmoduline et des ions calcium sur la plasticité conformationnelle du complexe. Les tendances conformationelles de AC libre ont aussi été étudiées. L'analyse conjointe des influences énergétiques et des liaisons hydrogène a révélé un réseau d'interactions entre AC et la calmoduline, dans lequel trois résidus clés, susceptibles de jouer un rôle allostérique sur l'activité de l'adénylcyclase ont été modifiés par mutagenèse dirigée. Ces tests expérimentaux ont conduits à la mise en évidence d'une région allostérique qui assure la communication de l'information de transition conformationnelle entre le site de fixation de la calmoduline et le site catalytique. Une exploration conformationnelle plus approfondie de AC à l'état non-lié a été entreprise par une méthode innovante de dynamique accélérée par la température (TAMD). Elle nous a conduit à la prédiction de conformations échantillonnées de AC dans son état libre. Ces prédictions pourraient être utilisées à l'avenir pour stabiliser l'état libre et faciliter l'étude expérimentale de sa structure

The protein CyaA is an important virulence factor of Bordetella pertussis involved in the whooping cough disease. The AC domain of CyaA is activated in eukaryotic host cells as an adenyl-cyclase by the ubiquitous and calcium sensitive protein calmodulin (CaM). This activation transformes the ATP into cAMP in an uncontrolled way, leading to severe cellular dysfunction. Only the structure of AC complexed with calcium-loaded calmodulin has been resolved by X-ray crystallography. Starting from this structure, molecular dynamics simulations of AC free and in complex with calmodulin enabled to characterize the effect of the calmodulin and calcium ions on the conformational plasticity of the complex. Conformational tendencies of free AC were also studied. The joint analysis of energetic influences and hydrogen bonds revealed a network of interactions between AC and calmodulin, in which three key residues, that may play an allosteric role on the catalytic activity of the adenyl-cyclase, were modified by site-directed mutagenesis. These experimental tests have led to the identification of an allosteric region which assures the communication of the conformational transition information between the binding site of the calmodulin and the catalytic site. A more thorough conformational exploration of the unbound AC state was undertaken by an innovative temperature accelerated molecular dynamics (TAMD) method. It led to a prediction of some sampled conformations of AC in its unbound state. These predictions could be used in the future to stabilize the free state and to facilitate the experimental study of its structure; La protéine CyaA est un facteur de virulence majeur de Bordetella pertussis, impliqué dans la maladie de la coqueluche. Le domaine catalytique AC de CyaA est directement transféré dans la cellule hôte eucaryote, où il est activé comme adénylcyclase par la calmoduline, une protéine ubiquitaire et sensible aux ions calcium. Ainsi, AC transforme l'ATP en AMPc de manière incontrôlée, ce qui conduit à des dérèglements cellulaires. Seule la structure de AC complexé à la calmoduline chargée d'ions calcium avait été résolue par cristallographie aux rayons X. À partir de cette structure, des simulations de dynamique moléculaire de AC libre, et en complexe avec la calmoduline nous ont permis de caractériser l'effet de la calmoduline et des ions calcium sur la plasticité conformationnelle du complexe. Les tendances conformationelles de AC libre ont aussi été étudiées. L'analyse conjointe des influences énergétiques et des liaisons hydrogène a révélé un réseau d'interactions entre AC et la calmoduline, dans lequel trois résidus clés, susceptibles de jouer un rôle allostérique sur l'activité de l'adénylcyclase ont été modifiés par mutagenèse dirigée. Ces tests expérimentaux ont conduits à la mise en évidence d'une région allostérique qui assure la communication de l'information de transition conformationnelle entre le site de fixation de la calmoduline et le site catalytique. Une exploration conformationnelle plus approfondie de AC à l'état non-lié a été entreprise par une méthode innovante de dynamique accélérée par la température (TAMD). Elle nous a conduit à la prédiction de conformations échantillonnées de AC dans son état libre. Ces prédictions pourraient être utilisées à l'avenir pour stabiliser l'état libre et faciliter l'étude expérimentale de sa structure