3.1.1 Context

In this axis, the CAPSID team develops methods related to knowledge discovery from databases (KDD, 36). The diversity of biological databases and resources is such today that it is more and more difficult to consider each database independently from the others 53. A limited subset of these resources is devoted to the 3D structure of biological objects (proteins, nucleic acids, glycanes...). Structural information is also contained in databases classifying protein domains as building blocks of proteins that can be reused in different proteins sharing the same function (Pfam, CATH and InterPro are well-known examples of such databases) 48, 58, 30. There are millions of proteins across all living species but only tens of thousands of domains that are combined in proteins. Thus, complex tasks such as predicting protein function or interactions can be simplified when envisaged at the domain level.

Due to the great diversity of databases, Knowledge Graphs (KGs) are more and more used to represent and integrate biological information. There is no single definition of KGs as these graphs cover a large variety of domains and data representation contexts (for instance the GAFAM companies advertize various KG uses). The main feature that differentiates KGs from classical graphs is the fact that both nodes (or entities) and edges (or relations) in the graph are heterogeneous and belong to various types described in the KG schema (metagraph). The field of biological knowledge discovery in KG is expanding rapidly 46. Most biological KGs today are developed for drug repurposing tasks (e.g. HetioNet 42 or DRKG). Clinicians are also very interested in network science carried out on rich knowledge graphs as a mean to interpret biomarker studies. However, there is still a need for curated, reliable biological KGs and for efficient knowledge discovery methods in KGs.

3.1.2 Knowledge discovery from protein structural databases

Concerning protein structural databases, we aim to explore novel classification paradigms exploiting existing resources about protein folds and domains 26, 27, 48, 58. In particular it will be interesting to use Kpax, our structural alignment tool 54, to define domain-domain similarity matrices. A non-trivial issue with clustering is how to handle similarity rather than mathematical distance matrices, and how to find the optimal number of clusters, especially when using a non-Euclidean similarity measure. We will adapt the algorithms and the calculation of quality indices to the Kpax similarity measure. More fundamentally, it will be necessary to integrate this classification step in the more general KDD process leading from data to knowledge.

For example, protein domain classification is relevant for studying domain-domain interactions (DDI). Our previous work on Knowledge-Based Docking (KBDOCK, 38, 40) will be updated and extended using newly published DDIs. Methods for inferring new DDIs from existing protein-protein interactions (PPIs) will be developed. Efforts should be made for validating such inferred DDIs so that they can be used to enrich DDI classification and predict new PPIs.

3.1.3 Function Annotation in Large Protein Graphs

Knowledge of the functional properties of proteins can shed considerable light on how they might interact. However, huge numbers of protein sequences in public databases such as UniProt/TrEMBL lack any functional annotation, and the functional annotation of such sequences is a highly challenging problem. We are developing graph-based and machine learning techniques to annotate automatically the available unannotated sequences with functional properties such as Enzyme Commission (EC) numbers and Gene Ontology (GO) terms (note that these terms are organised hierarchically allowing generalization/specialization reasoning). The idea is to transfer annotations from expert-reviewed sequences present in the UniProt/SwissProt database (about 560 thousands entries) to unreviewed sequences present in the UniProt/TrEMBL database (about 80% of 180 millions entries). For this, we have to learn from the UniProt/SwissProt database how to compute the similarity of proteins sharing identical or similar functional annotations. Various similarity measures can be tested using cross-validation approches in the UniProt/SwissProt database. For instance, we can use primary sequence or domain signature similarities. More complex similarities can be computed with graph-embedding techniques.

3.1.4 Knowledge discovery algorithms in large biological knowledge graphs

KGs are particularly useful and appropriate in biology, to represent and integrate the complex contents of biological databases 52. We intend to design algorithms for leveraging information embedded in biological KGs (also known as complex networks). In biology, KGs mostly represent PPIs, integrated with various properties attached to proteins, such as pathways, drug binding or relation with diseases. Setting up similarity measures for proteins in a knowledge graph is a difficult challenge. Our objective is to extract useful knowledge from such graphs in order to better understand and highlight the role of multi-component assemblies in various types of cell or organisms. Ultimately, knowledge graphs can be used to model and simulate the functioning of such molecular machinery in the context of the living cell, under physiological or pathological conditions.

3.2.1 Context

This axis deals with 3D protein structure and interactions. In fact, the long-lasting problem of predicting a 3D structure from a protein sequence has been solved in 2021 by the AlphaFold2 (DeepMind) 44 or RosettaFold methods 29. This success, revealed in the CASP14 (Critical Assessment of Stucture Prediction) challenge, was possible not only thanks to AI methods but also because the amount of experimental 3D structures has reached a sufficient size in the Protein Data Bank (PDB). For the same type of reasons, the rigid docking problem (in which the bodies to dock are rigid) seems to be on the way to being solved as well 31, 63. However, research is still required to address the problem of docking disordered proteins or flexible nucleic acids that will fold as they bind to proteins. This is the direction taken by the team since the arrival of Isaure Chauvot de Beauchêne, the inventor of a fragment-based approach for RNA docking onto proteins.

Modeling protein - and even more RNA - flexibility accurately during docking is very computationally expensive. This is due to the very large number of internal degrees of freedom in each molecule, associated with twisting motions around covalent bonds. Therefore, it is highly impractical to use detailed force-field or geometric representations in a brute-force docking search. Instead, most docking algorithms use fast heuristic methods to perform an initial rigid-body search in order to locate a relatively small number of candidate binding orientations, and these are then refined using a more expensive interaction potential or force-field model, which might also include flexible refinement using molecular dynamics (MD), for example.

3.2.2 Coarse-Grained Models

Many approaches have been proposed in the literature to take into account protein (and more recently RNA/DNA) flexibility during docking. The most thorough methods rely on expensive atomistic simulations using MD. However, much of a MD trajectory is unlikely to be relevant to a docking encounter unless it is constrained to explore a putative protein-protein/NA interface. Consequently, MD is normally only used to refine a small number of candidate rigid body docking poses. A much faster - but more approximate - method is to use "coarse-grained" (CG) normal mode analysis (NMA) techniques to reduce the number of flexible degrees of freedom to just one or a handful of the most significant vibrational modes 35, 47, 49, 51. In our experience, docking ensembles of NMA conformations do not give much improvement over basic FFT-based soft docking 61, and it is very computationally expensive to use side-chain repacking to refine candidate soft docking poses 39.

In the last few years, CG force-field models have become increasingly popular in the MD community because they allow very large biomolecular systems to be simulated using conventional MD programs 28. Typically, a CG force-field representation replaces the 5-15 atoms in each amino acid with 2-4 “pseudo-atoms” (each pseudo-atom represents few atoms of an amino-acid as a single bead). It then assigns each pseudo-atom a small number of parameters to represent its chemo-physical properties. By directly attacking the quadratic nature of pair-wise energy functions, coarse-graining can speed up MD simulations by up to three orders of magnitude. Nonetheless, such CG models can still produce useful models of very large multi-component assemblies 57. Furthermore, this type of CG model effectively integrates many internal DOFs to build a smoother but still physically realistic energy surface 43. We are currently developing a CG scoring function for RNA-protein docking by fragments assembly.

3.2.3 Assembling Multi-Component Complexes and Integrative Structure Modelling

We also want to develop related approaches for integrative structure modeling using cryo-electron microscopy (cryo-EM). Thanks to recent developments in cryo-EM instruments and technologies, it is now feasible to capture low resolution images of very large macromolecular machines. However, while such developments offer the intriguing prospect of being able to trap biological systems in unprecedented levels of detail, there will also come with an increasing need to analyse, annotate, and interpret the enormous volumes of data that will soon flow from the latest instruments. In particular, a new challenge that is emerging is how to fit previously solved high resolution protein structures into low resolution cryo-EM density maps. However, the problem here is that large molecular machines will have multiple sub-components, some of which will be unknown, and many of which will fit each part of the map almost equally well. Thus, the general problem of building high resolution 3D models from cryo-EM data is like building a complex 3D jigsaw puzzle in which several pieces may be unknown or missing, and none of which will fit perfectly. We wish to proceed firstly by putting more emphasis on the single-body terms in the scoring function 34, and secondly by using fast CG representations and knowledge-based distance restraints to prune large regions of the search space, as initiated with the EROS-DOCK software 6, 56.

3.2.4 Protein-Nucleic Acid Interactions

As well as playing an essential role in the translation of DNA into proteins, RNA molecules carry out many other essential biological functions in cells, often through their interactions with proteins. A critical challenge in modeling such interactions computationally is that the RNA is often highly flexible, especially in single-stranded (ssRNA) regions of its structure. These flexible regions are often very important because it is through their flexibility that the RNA can adjust its 3D conformation in order to bind to a protein surface. However, conventional protein-protein docking algorithms generally assume that the 3D structures to be docked are rigid, and so are not suitable for modeling protein-RNA interactions. There is therefore much interest in developing dedicated protein-RNA docking algorithms which can take RNA flexibility into account. This research topic has been initiated with the recruitement of Isaure Chauvot de Beauchêne in 2016 and is becoming a major activity in the team. A novel flexible docking algorithm is currently under development in the team. It first docks small fragments of ssRNA (typically three nucleotides at a time) onto a protein surface, and then combinatorially reassembles those fragments in order to recover a contiguous ssRNA structure on the protein surface 32, 33.

As the correctness of the initial docking of the fragments settles an upper limit to the correctness of the full model, we are now focusing on improving that step. A key component of our docking tool is the energy function of the protein-fragment interactions that is used both to drive the sampling (positioning of the fragments) by minimization, and to discriminate the correct final positions from decoys (i.e., false positives). We are developing a new approach to create knowledge-based parameters for coarse-grain energy functions from public structural data, in collaboration with Sjoerd de Vries (INSERM). Such approach will be applied first to ssRNA-protein complexes, then to other types of complexes such as protein-peptides.

Another key requirement for this approach is an exhaustive but non-redundant library of possible internal conformations of RNA fragments. Our library is built by clustering hundreds of thousands of experimentally known RNA structures, based on an approximate geometric similarity criteria. We want to develop new algorithms for the clustering of 3D conformations based on internal coordinates and on epsilon-net theory, in order to optimise the representativity and computational cost of the library.

In the future, we will improve the combinatorial algorithm used for reassembling the docked fragments using both experimental constraints and knowledge-based constraints pertaining from the research carried out in Axis 1.

7.1.1 InteR3Mdb

• Name:
  Database for interactions between RNA and RRM (RNA Recognition Motif)
• Keywords:
  Databases, Proteins, Nucleic Acids, 3D interaction, Biological sequences
• Scientific Description:
  InteR3Mdb is a comprehensive database gathering and integrating public data on the 3D structures and sequences and on in vitro experiments, related to all instances known so far of a very conserved protein domain, the RNA Recognition Motif (RRM). Special effort has been made to enter in the database all available and curated information about the interactions of these RRM instances with RNA.
• Functional Description:
  InteR3Mdb is a comprehensive database built on a relational data model. As a tool, it comprises a web-interface offering filtering and multicriteria search functionalities about RRMs, as well as an API interface.
• Release Contributions:
  InteR3Mdb V1.0 is released with an extended public user interface offering filtering and multi-criteria querying functionalities. The user manual, data dictionary and help documentation are now fully available as well as the main database content statistics (https://inter3mdb.loria.fr)
• News of the Year:
  During year 2022, the InteR3Mdb data model has evolved to integrate results about binding affinities between RRM-containing proteins and RNA. The list of represented RRM domains has been updated. A public interface (https://inter3mdb.loria.fr) has been created with filtering and multicriteria search functionalities. Documentation (Data Dictionary, Statistics) is available on the web site.
• URL:
  https://­inter3mdb.­loria.­fr
• Contact:
  Hrishikesh Dhondge

7.1.2 ProtNAff

• Name:
  Protein - Nucleic Acids Filters and Fragments
• Keywords:
  Structural alphabet, Structural Biology, Nucleic Acids
• Scientific Description:
  The modeling of nucleic acids (NA) - protein interactions can greatly help the design of therapeutic NA. Atomistic models of NA fragments can be used to model the 3D structures of NA-protein complexes, to subdivide the handling of RNAs great flexibility. One way to obtain relevant RNA fragments is to extract them from existing 3D structures of interactions corresponding to the context one wants to model (such as surrounding NA 2D structures, specific protein families, specific sequences) and to the objectives to the study.
• Functional Description:
  ProtNAff is a python-based software for (i) the automated parsing, correction and annotation of all protein-nucleic acid structures in the public Protein Data Bank, (ii) the creation of libraries of non-redundant RNA/DNA structural fragments, (iii) the selection of sets of structures by customised queries, and (iv) the computation of statistics on sets of RNA/DNA - protein structures.
• URL:
  https://­github.­com/­isaureCdB/­NAfragDB
• Publication:
  hal-02393039
• Contact:
  Isaure Chauvot de Beauchêne
• Participants:
  Isaure Chauvot de Beauchêne, Antoine Moniot, Sjoerd De Vries

7.1.3 DISCAN-2022

• Name:
  Distributed and Incremental Structural Clustering Algorithm for Networks
• Keywords:
  Big data, Clustering, Unsupervised graph clustering SCDG, Distributed computing
• Functional Description:
  This is a distributed implementation of a Distributed and Incremental Structural Clustering Algorithm for Networks called DISCAN. For more details see the paper "A distributed and incremental algorithm for large-scale graph clustering" by Wissem Inoubli, Sabeur Aridhi, Haithem Mezni, Mondher Maddouri, Engelbert Mephu Nguifo. ⟨10.1016/j.future.2022.04.013⟩.
• URL:
  https://­github.­com/­inoubliwissem/­remote-master
• Publication:
  hal-03659549
• Contact:
  Wissem Inoubli
• Partner:
  Université Clermont Auvergne

8.1.1 Biomedical Knowledge Discovery

In the context of our collaboration with clinicians at the CHRU Nancy in the RHU FIGHT-HF project (2015-2021 ; see preceding report), we have proposed an approach based on an inductive database to support iterative data mining 8. In this approach, we extend the KDD process model with the help of an inductive database and we design the first generic model of Inductive Clinical DataBase (ICDB) aimed at hosting both patient data and learned models. We report experiments conducted on patient data within the FIGHT-HF program. The ICDB approach allows us to identify biomarker combinations, which are specific and predictive of heart fibrosis phenotype and provide hypotheses about the underlying mechanisms. Two main scenarios were considered, a local-to-global KDD scenario and a trans-cohort alignment scenario. This promising proof of concept enables us to draw the contours of a next-generation Knowledge Discovery Environment (KDE).

In parallel, we continue to exploit the network science platform developed during the FIGHT-HF project. Queries on the Heart Failure Graph Knowledge Box (HF-GKBox) have provided useful information to clinicians for interpreting their results concerning the association of two biomarkers (NT-proBNP and Stem Cell Factor) present in blood plasma with cardiovascular outcomes in end-stage hemodialysis patients with renal disease 15.

In the context of the PraktikPharma ANR project coordinated by Adrien Coulet (HeKA team), we investigated adverse drug reaction mechanisms with knowledge graph mining 23.

8.1.2 Graph-based Approaches for Machine Learning and Protein Annotation

The developments conducted by Bishnu Sarker to improve his method for protein function annotation (GrAPFI for Graph-based Automatic Protein Function annotation) have been published. GrAPFI is based on label propagation in graphs. It relies on a graph whose nodes represent proteins and whose edges are weighted by the domain similarity between proteins. When applied to function annotations taken from Gene Ontology, the method was improved by post-processing of the predicted annotations, leveraging the hierarchical structure of the ontology. It was shown that the post-processing step also improves other methods of protein function annotation 16.

Similarity functions are essential for label propagation algorithms in graphs. As an extension of our Tempo-Graphs project (2019-2021) with our Brazilian partners at Federal University of Ceara (UFC, Fortaleza), we developed a similarity function for sparse binary data with application on protein function annotation 19. The proposed similarity function is based on the analysis of the best existing similarity functions for the protein function annotation task. We performed experiments in a simple pairwise similarity scenario and also using our proposal as part of a more complex protein function annotation method.

Moreover, through the past co-supervision by Sabeur Aridhi of Wissem Inoubli’s PhD at the University of Tunis (2018-2021), the team has acquired recognized expertise in big data methods for large graphs and contributed to the development of a distributed, incremental algorithm for large-scale graph clustering 10. The corresponding software DISCAN-2022 is described above (see section New Software).

8.1.3 Knowledge graph mining with embedding-based methods

In the context of Md Kamrul Islam's PhD project, we addressed the problem of link prediction in large knowledge graphs (KGs) using KG embedding methods. These methods aim to learn low dimensional vector representations of entities and relations in a KG. Such representations (in a latent space) facilitate the link prediction task, which serves inference and completion in a KG. In this context, it is important to achieve both an efficient KG embedding and explainable predictions. During learning of efficient embeddings, sampling negative triples is an important step as KGs only come with observed positive triples. We proposed an efficient simple negative sampling (SNS) method based on the assumption that the entities which are closer to the corrupted entity in the embedding space are able to provide high-quality negative triples 11.

As for explainability, it actually constitutes a thriving research question especially when it comes to analyse KGs with their rich semantics rooted in description logics. Hence, we proposed a new rule mining method which exploits the learned embeddings 11.

We evaluated our SNS sampling method plugged to several KG embedding models by calculating the performance of the link prediction task on well-known datasets. Experimental results show that the SNS improves the prediction performance of KG embedding models, and outperforms the existing sampling methods. To assess the performance of our rule mining method with and without SNS, we mine and evaluate rules on three popular datasets. The extracted rules are evaluated as knowledge nuggets extracted from the KG and also as a support for explainable link prediction. The achieved results are good and open the way to many improvements and new perspectives. Md Kamrul Islam has defended his PhD on December 16, 2022.

8.1.4 Biological network modeling

Boolean Networks (BNs) refer to a simple formalism used to study complex biological systems when the prediction of exact reaction times is not of interest. BNs play a key role in understanding the dynamics of the studied systems and in predicting their disruption in case of complex human diseases. The BioModels database is a well-known repository of peer-reviewed models represented in the Systems Biology Markup Language (SBML). Most of these models are quantitative, but in some use cases, qualitative models—such as BNs—are better suited. In the context of Athénais Vaginay's PhD project, we proposed SBML2BN, a pipeline dedicated to the automatic transformation of quantitative SBML models to Boolean networks 18. Our approach takes advantage of several SBML elements (reactions, rules, events) as well as a numerical simulation of the concentration of the species over time to constrain both the structure and the dynamics of the Boolean networks to synthesise. Finding all the BNs complying with given structure and dynamics was formalised as an optimisation problem formulated in the answer-set programming framework.

We ran SBML2BN on more than 200 quantitative SBML models, and we could construct Boolean networks which are compatible with the structure and the dynamics of the SBML models 18. A more recent work relies on abstract simulation of a chemical reaction network (CRN) to avoid the tricky binarization task 21. Athénais Vaginay will defend her PhD in May 2023.

8.2.1 Inferring epsilon-nets of RNA 3D-fragments

Our fragment-based approach to dock ssRNA on proteins requires libraries of 3D conformations of RNA fragments. Last year, we reported ProtNAff, a python-based software for (i) the automated parsing and annotation of protein-RNA experimental structures, (ii) the selection of (parts of) such structures by customized queries, (iii) the creation of fragment libraries. Since then, we have applied it to characterize the conformational specificity of RNA in different states: single-stranded or double-stranded, unbound or bound to proteins, of different sequences, etc. The ProtNAff software and these analyses were published in a peer-reviewed journal 14.

The RNA fragment libraries must approximate all possible local RNA conformations within a given precision, but be of small enough cardinality to be usable for combinatorial assemblies of fragments. Such a set of well chosen prototypes can be obtained by clustering all fragments of experimentally solved structure, using a suitable clustering algorithm and a measure of dissimilarity between fragments. When the representativeness criterion is a distance, the problem reduces to that of inferring an epsilon-net 1 ($\epsilon$-net) of minimal cardinality for a finite set of points in a metric space. After the construction of the $\epsilon$-net, each point $P_i$ from the initial set should be at a distance $d_i<\epsilon$ from at least one point in the $\epsilon$-net. The minimal cardinality that can be achieved is called the $\epsilon$-covering number.

In collaboration with Yann Guermeur from team ABC (LORIA), we investigated methods to construct minimal $\epsilon$-nets from a given set of points and a given distance measure. Although the literature provides us with upper bounds on the covering numbers 2, the proofs are non constructive. To the best of our knowledge, there was no (efficient) algorithm to compute $\epsilon$-nets of small cardinality in the framework of interest. We have therefore developed an algorithm to derive $\epsilon$-nets, that operates in a reproducing kernel Hilbert space 3, by combining two well-known tools of empirical inference: the hierarchical agglomerative clustering and the computation of minimum enclosing balls. Tested on a well-known set of images used for benchmarking classification tools, it produces $\epsilon$-nets whose cardinality is smaller than those obtained with state-of-the-art methods. This work has been presented at the WSOM international conference 20.

We then applied this approach on RNA fragments, with some arrangements. A commonly used measure of dissimilarity in structural biology is the root mean squared deviation (RMSD) of atomic positions, whose exact computation requires a pairwise structural superposition of the 3D structures. But this superposition is highly time-consuming and not applicable for a very large initial set of fragments (typically ${10}^4$ to ${10}^5$ in our cases). In our approach, we first approximate the real RMSD by the RMSD after superposition on one random fragment of the set, which provides an upper bound of the real RMSD. When a ball calculation is to be performed, all the fragments involved are superimposed again on a common reference structure. To know which balls are really useful to compute, bounds are used (to limit superpositions). This allowed us to reduce by a factor  4 the size of our fragment libraries, compared to the results obtained by the star-shape clustering implemented in the first version of ProtNAff. This work was carried out with ABC in the context of Antoine Moniot's PhD, co-supervised by Yann Guermeur and Isaure Chauvot de Beauchêne, and successfully defended on December 12, 2022.

8.2.2 Modeling and design of RNA-RRM complexes

RNA docking with stacking constraints. Our H2020 ITN project RNAct aims at designing new RNA-binding proteins based on the evolutionary conserved protein domain4, called RNA Recognition Motif (RRM). In this context, we have created in 2020 the InteR3Mdb database that contains all known 3D structures of RRMs and RRM-RNA complexes. It allowed us to infer, for each amino acid at every evolutionary conserved position in the sequence5, the propensity to bind a given nucleotide type (A/C/U/G) with a given interaction type (electrostatic, Hydrogen-bond, etc). We are now using this resource to derive constraints for the modeling of an RNA bound to a given RRM. As a first step, we concentrated on the $\pi -\pi$ stacking interactions occurring between a base of the RNA and the aromatic ring of an amino-acid at evolutionary conserved positions in an RRM. We inferred, from superposition and clustering of all existing RNA-RRM structures, the main possible positions of a stacking RNA base toward an RRM structure. Then, on a test-case RRM structure, we modeled the RNA binding positions by docking (i.e. sampling low-energy positions for RNA fragments) with an energy penalty added to the RNA-protein force-field for deviations from the reference position. The approach was tested on 12 experimental RRM-RNA structures: it retrieved correct positions (RMSD < 1 Å from the target position) for all test cases, versus only 1 test case by docking without constraints.

Cross-mapping of protein domain structural instances. To address the issue of inconsistencies of protein classification in different databases, we developed a tool named Cross-Mapper for domain Structural instances (CroMaSt) that is a workflow to identify and assign a confidence level to each experimental 3D structure of a given domain. This workflow was primarily developed for RRMs, but can easily be adapted to any structural domain. The workflow has been formalized using the Common Workflow Language (CWL) 6. Its rationale is based on the cross-mapping of PDB (Protein Data Bank) entries retrieved as RRM, in the Pfam and CATH domain classifications. Entries that are mapped in both classifications are considered as the “core” RRMs. Those that are specific to only one classification are further analyzed using structural alignment, in order to determine whether they are RRM-like or false RRMs 24. The workflow can be used to create datasets for specific domains with a good confidence level, which are useful for characterizing domain structural diversity or for further analyses such as machine learning, evolutionary studies or synthetic biology. An article on CroMaSt and the rich diversity of RRM domains has been submitted to Bioinformatics in November 2022 (under review).

Empirical inference of an RNA-protein energy function. In collaboration with Sjoerd de Vries 7, we have implemented a new method to create energy parameters for RNA-protein interactions in coarse-grained representations. Each amino-acid of the protein and each nucleotide of the RNA is represented by 2 to 7 pseudo-atoms (“beads”). For each model of an RNA-protein interaction, the energy is computed as the sum of the bead-bead energies, and the model with the lowest energy is considered as the most probable. Each bead-bead energy depends solely on the 2 bead types (among 17 RNA types and 32 protein types) and the inter-bead distance. For each pair of bead types, the energy function has the shape of a Lennard-Jones potential, with 2 parameters that determine the distance of minimal energy and the value of the minimal energy. The current parameters were extracted in 2010 by statistics on existing RNA-protein crystal structures and optimized by a random Monte Carlo-like strategy. They were initially tailored for double-stranded RNA, and their performance is poor on single-stranded RNA (ssRNA). One main goal of Anna Kravchenko’s PhD is to optimize those parameters for the discrimination of correct models for a given ssRNA-protein complex. To achieve that, we have set up a novel “histogram-based” approach. For each pair of bead types, we build a log-odds histogram of the occurrences of distances (discretized into bins) in correct/incorrect models from docking runs on a training set. We then use these histograms, which correspond to the residual error of the energy function, to score each individual model in docking runs on a test set. We tested this corrected approach on a benchmark of 131 complexes of known structures of ssRNA trinucleotides bound to 56 proteins and Anna Kravchenko presented the results as an oral communication at the 35th Rhine-knee Regiomeeting for Structural Biology (5-7 October 2022, Gerardmer, France). In perspective, if we can find a way to transform a set of histograms into a set of docking parameters then we can use this approach to also improve the sampling step, by increasing the raw number of correct models among all generated models.

8.2.3 3D Modeling of protein complexes - Virtual Screening

Modeling DNA-protein complexes to fight antibiotic resistance. In the context of the CITRAM project (collaboration with the DynAMic and LPCT labs), we study a new type of relaxase (RelSt3) as a key protein for horizontal DNA transfer, one of the processes responsible for the spread of antibiotic resistance in bacteria. At the very beginning of this process, DNA interacts with the relaxase both as a classical double-helix and as a transient single-stranded state. We therefore developped a ssDNA docking tool to model the interaction of the relaxase RelSt3 with ssDNA. To benchmark this tool, we created the first dataset of ssDNA-protein structures. From the public general database of experimental structures (PDB), we extracted, processed and clustered 284 complexes containing protein-bound ssDNA. A main difficulty in docking is the conformational differences between the unbound and the DNA-bound protein structures, the latter being not predictable from the former. Using many unbound structures for docking can increase the chance that at least one is close enough to the bound one to provide accurate models. Therefore, we also retrieved all unbound structures from the Protein Data Bank for each protein of our 284 ssDNA-protein structures. We then used this dataset (i) to compare intrinsic and extrinsic conformational variability, i.e. structural changes induced or not by ssDNA binding, and (ii) to prove the advantage of using multiple protein structures in ssDNA docking 13.

In parallel, we modeled the interaction of the HtH domain8 of RelSt3 with double-helix DNA, using known structures of homologous complexes and molecular dynamics simulations. Our models identified key residues for interaction with DNA, which were confirmed by the loss of relaxase activity in vitro after mutating those residues. Part of this work has been published in the Nucleic Acids Research journal, with our microbiologist partners 12.

COVID-19 spike binding to ACE2. In collaboration with Laurent Vuillon and Aria Gheeraert at the LAMA (Laboratoire de Mathématique, CNRS-Université Savoie Mont Blanc) and with Laurent Chaloin and Olivier Moncorgé at the IRIM (Institut de Recherche en Infectiologie de Montpellier, CNRS-Université de Montpellier), we investigated disparities between the SARS-CoV-2 wild-type and five variants that emerged at the end of 2020, focusing on the Spike/ACE2 complex and using experimental structures and molecular dynamics simulations. Dihedral angle PCA and dynamical perturbation networks showed specificities in the Spike dynamics and in the Spike/ACE2 interface respectively, in 4 and 3 variants compared to wild-type respectively. Atomic contacts PCA shows how the L452R and T478K mutations act synergistically on neighboring residues to provoke drastic changes compared to the wild-type behavior 9.

Structural basis of donor-specific antibody response in graft rejection. In the context of the Inria-Inserm PhD project of Diego Amaya Ramirez, in collaboration with the Immunology and Histocompatibility Laboratory at the APHP Saint-Louis Hospital in Paris, we study the structural differences between donor and recipient HLA proteins to understand and possibly predict the immune response triggered in the recipient, which can result in rejection of the transplanted organ. A dataset of 207 HLA 3D structures has been created, both from PDB and AlphaFold predictions. Molecular dynamics runs (10 ns) have been run to simulate protein surface flexibility. These runs are analyzed at the level of single residues or surface patches centered on such residues. Single residues are selected as "confirmed eplets" (polymorphic amino-acids responsible of generating donor-specific antibodies9) or "non eplets" (conserved amino-acids in all HLA proteins, assumed not to induce any donor-specific antibody response). Based on a recent publication suggesting that antibodies tend to bind to less flexible regions on the surface of the proteins 45, we studied the flexibility of HLA surface along molecular dynamics runs, at eplet versus non eplet positions. Our results reveal a significant difference which could indicate that HLA-specific antibodies would also preferentially bind to less flexible regions 22. The datasets built during this preliminary study constitute the basis for further machine learning approaches helping to discriminate between compatible and non compatible recipient-donor HLA pairs.

Virtual screening Virtual screening of small molecules is an essential part of the team's expertise that is often called upon by external partners in biology. In collaboration with Brazilian partners from EMBRAPA, we developped a computational strategy to identify new efficient and environmentally-friendly plant protection solutions. Anti-mycotoxin compounds were isolated by virtual screening using as a target a key enzyme specific of the pathogen fungus 7. In parallel, we participated in an innovative study with colleagues at the University of Maringa in which deep learning methods were combined with classical ligand- and target-based chemoinformatic methods to identify new drugs against the tuberculosis pathogen 17.

10.1.1 Inria associate team not involved in an IIL or an international program

10.2.1 H2020 projects

ANR EPIHLA

Participants: Marie-Dominique Devignes [contact person], Diego Amaya Ramirez, Bernard Maigret.

• Title:
  HLA compatibility in organ transplantation : from antigens to epitopes (EPIHLA)
• Duration:
  October 2022-October 2025
• Coordinator:
  Pr. Jean-Luc Taupin (Inserm U976, Saint-Louis Hospital, Paris)
• Inria contact:
  Marie-Dominique Devignes
• Partner Institutions:
  • Inserm U976 IRSL Saint-Louis Hospital (Paris)
  • LORIA CNRS (Nancy)
  • INSERM U1016 Cochin Institute (Paris)
  • CNRS U144 Institut Curie (Paris)
• Summary:
  The EPIHLA project has two major aims. (1) It aims at correctly representing HLA molecule 3D structure and superimposing predicted conformations in order to identify 3D differences that could constitute epitopes and eplets, targets of donor-specific antibodies. (2) It aims at developing the capacity to isolate and clone anti-HLA antibody genes from patients’ B lymphocytes. The results will provide decisive new information on the understanding of humoral alloreactivity and will make it possible to better anticipate transplant rejection. This project is partially based on the Inria-Inserm PhD project of Diego Amaya Ramirez (2019-2022). This PhD ("HLA genetic system and organ transplantation: understanding the basics of immunogenicity to improve donor - receptor compatibility when assigning grafts to recipients") is co-supervised by Marie-Dominique Devignes and Pr. Jean-Luc Taupin.

11.1.1 Scientific events: organisation

11.1.2 Scientific events: selection

11.1.3 Journals

11.1.4 Leadership within the scientific community

• Marie-Dominique Devignes is coordinating the Interoperability working group at the Institut Français de Bioinformatique with a bioinformatician colleague from Nantes (Alban Gaignard, Institut du Thorax, Plateforme BiRD). This working group is a member of the ELIXIR Interoperability platform. She represents IFB for this part of its activity in meetings at the national and international levels.

11.1.5 Scientific expertise

• Marie-Dominique Devignes reviewed grant applications for the European Science Foundation (call FWO-Odysseus 2022) and for BPI-France (call i-demo 2022).
• Malika Smaïl-Tabbone is enrolled as an expert in data science and AI in two European projects coordinated by a leading clinician (Pr. Magnus Bäck): an HORIZON Europe project entitled CARE-IN-HEALTH (CARdiovascular Resolution of Inflammation to promote HEALTH) and an ERA PerMed project entitled OmegaPerMed (Optimizing omega-3 supplementation to resolve inflammation in a personalized medicine cardiovascular disease prevention). She is responsible for the design and execution of the data science- and AI-related tasks in both projects.
• Malika Smaïl-Tabbone and Marie-Dominique Devignes contributed to the elaboration of the interdisciplinary LUE program TRAVEL (Co-morbidiTies, tRAjectories of liVEs and Longevity), which gathers the three digital science laboratories in Nancy together with Nancy Hospital and research laboratories in public health, biology and social sciences.

11.1.6 Research administration

• Marie-Dominique Devignes is responsible for the Transverse Axis "Digital Health" at the LORIA.
• Marie-Dominique Devignes was member of the hiring committee for an assistant professor (Maître de Conférences) at Université Paris-Saclay in May 2022.

11.2.1 Teaching

• Malika Smaïl-Tabbone is an associate professor at the Université de Lorraine with a full service. She is co-responsible with Pascal Moyal of the IMSD track ("Ingénierie Mathématique pour la Science des Données") in the Applied Mathematics Master's degree at the Université de Lorraine. She is also a member of the pedagogic team of the CMI BSE ("Cursus Master Ingénieur Biologie-Santé-Environnement") and also in charge of corporate relations.
• Sabeur Aridhi is an assistant professor at the Université de Lorraine with a full service. He is responsible for the major in IAMD ("Ingénierie et Applications des Masses de Données") at TELECOM Nancy.
• Marie-Dominique Devignes teaches every year 10 to 16h in the CMI BSE.
• Athenaïs Vaginay had an ATER position from September 2021 to August 2022.
• Diego Amaya Ramirez started an ATER position in October 2022.

11.2.2 Supervision

• PhD: Antoine Moniot. "Modélisation des complexes ARN/protéines par assemblage de fragments structuraux". Co-supervised by Isaure Chauvot de Beauchêne and Yann Guermeur (ABC team, Loria). PhD defended on December 12, 2022. Available soon at https:theses.hal.science.
• PhD: Md Kamrul Islam. "Explainable link prediction in large complex graphs - application to drug repurposing". Co-supervised by Malika Smaïl-Tabbone and Sabeur Aridhi. Thesis defended on December 16, 2022. Available soon at https:theses.hal.science.
• PhD in progress: Athenaïs Vaginay, October 2018, co-supervised by Malika Smaïl and Taha Boukhobza (CRAN).
• PhD in progress: Diego Amaya Ramirez, October 2019, co-supervised by Marie-Dominique Devignes and Jean-Luc Taupin (Inserm, Hôpital Saint-Louis, Paris).
• PhD in progress: Anna Kravchenko, October 2019,co-supervised by Isaure Chauvot de Beauchêne and Malika Smaïl-Tabbone.
• PhD in progress: Hrishikesh Dhondge, October 2019, co-supervised by Isaure Chauvot de Beauchêne and Marie-Dominique Devignes.

11.2.3 PhD thesis juries

• ALFERKH Lina, 10 February 2022, Université Paris-Saclay (Isaure Chauvot de Beauchêne, examiner)
• PETRELIS Alexandros, 29 April 2022, Université de Lorraine (Marie-Dominique Devignes, examiner)
• VERAS Marcelo, 24 June 2022, Federal University of Ceará, Brazil (Sabeur Aridhi, examiner)
• AKID-ZEKRI Hajer, 12 December 2022, co-tutelle Université de Strasbourg and Université de Sfax (Malika Smaïl-Tabbone, reviewer)

11.2.4 Other juries

• Malika Smaïl-Tabbone is member of the Capes NSI (Numérique et Sciences Informatiques) jury since 2022.

11.3.1 Articles and contents

Two PhD students from the RNAct project have produced short presentation videos now available on YouTube (Anna Kravchenko and Hrishikesh Dhondge).

International journals

• 7 articleV.Vessela Atanasova, E.Emmanuel Bresso, B.Bernard Maigret, N. F.Natalia Florencio Martins and F.Florence Richard-Forget. Computational Strategy for Minimizing Mycotoxins in Cereal Crops: Assessment of the Biological Activity of Compounds Resulting from Virtual Screening.Molecules278April 2022, 2582
  HALDOIback to text
• 8 articleE.Emmanuel Bresso, J.-P.Joao-Pedro Ferreira, N.Nicolas Girerd, M.Masatake Kobayashi, G.Grégoire Preud’homme, P.Patrick Rossignol, F.Fayez Zannad, M.-D.Marie-Dominique Devignes and M.Malika Smaïl-Tabbone. Inductive database to support iterative data mining: Application to biomarker analysis on patient data in the Fight-HF project.Journal of Biomedical Informatics135November 2022, 104212
  HALDOIback to text
• 9 articleA.Aria Gheeraert, L.Laurent Vuillon, L.Laurent Chaloin, O.Olivier Moncorgé, T.Thibaut Very, S.Serge Perez, V.Vincent Leroux, I.Isaure Chauvot de Beauchêne, D.Dominique Mias-Lucquin, M.-D.Marie-Dominique Devignes, I.Ivan Rivalta and B.Bernard Maigret. Singular Interface Dynamics of the SARS-CoV-2 Delta Variant Explained with Contact Perturbation Analysis.Journal of Chemical Information and Modeling6212June 2022, 3107-3122
  HALDOIback to text
• 10 articleW.Wissem Inoubli, S.Sabeur Aridhi, H.Haithem Mezni, M.Mondher Maddouri and E.Engelbert Mephu Nguifo. A distributed and incremental algorithm for large-scale graph clustering.Future Generation Computer Systems134September 2022, 334-347
  HALDOIback to text
• 11 articleK.Kamrul Islam, S.Sabeur Aridhi and M.Malika Smaïl-Tabbone. Negative Sampling and Rule Mining for Explainable Link Prediction in Knowledge Graphs.Knowledge-Based Systems250August 2022, 109083
  HALDOIback to textback to text
• 12 articleH.Haifa Laroussi, Y.Yanis Aoudache, E.Emilie Robert, V.Virginie Libante, L.Louise Thiriet, D.Dominique Mias-Lucquin, B.Badreddine Douzi, Y.Yvonne Roussel, I.Isaure Chauvot de Beauch Êne, N.Nicolas Soler and N.Nathalie Leblond-Bourget. Exploration of DNA processing features unravels novel properties of ICE conjugation in Gram-positive bacteria.Nucleic Acids Research5014August 2022, 8127–8142
  HALDOIback to text
• 13 articleD.Dominique Mias-Lucquin and I.Isaure Chauvot de Beauchêne. Conformational variability in proteins bound to single-stranded DNA: a new benchmark for new docking perspectives.Proteins - Structure, Function and Bioinformatics903March 2022, 625-631
  HALDOIback to text
• 14 articleA.Antoine Moniot, Y.Yann Guermeur, S. J.Sjoerd Jacob de Vries and I.Isaure Chauvot de Beauchêne. ProtNAff: Protein-bound Nucleic Acid filters and fragment libraries.Bioinformatics38162022-07-012022, 3911–3917
  HALDOIback to text
• 15 articleP.P Rossignol, K.K Duarte, E.E Bresso, Å.Åsberg A, M. D.M D Devignes, N.N Eriksson, N.N Girerd, R.R Glerup, A.A Jardine, H.H Holdaas, Z.Z Lamiral, C.C Leroy, Z.Z Massy, W.W März, B.B Krämer, P.P Wu, R.R Schmieder, I.I Soveri, J.J Christensen, M.M Svensson, F.F Zannad and B.B Fellström. NT-proBNP and stem cell factor plasma concentrations are independently associated with cardiovascular outcomes in end-stage renal disease hemodialysis patients.European Heart Journal Open26November 2022
  HALDOIback to text
• 16 articleB.Bishnu Sarker, N.Navya Khare, M.-D.Marie-Dominique Devignes and S.Sabeur Aridhi. Improving automatic GO annotation with semantic similarity.BMC Bioinformatics23S2December 2022, 433
  HALDOIback to text
• 17 articleJ. V.João Vítor Perez De Souza, E. S.Erika Seki Kioshima, L. S.Letícia Sayuri Murase, D. d.Diego de Souza Lima, F. A.Flavio Augusto Vicente Seixas, B.Bernard Maigret and R. F.Rosilene Fressatti Cardoso. Identification of new putative inhibitors of Mycobacterium tuberculosis 3-dehydroshikimate dehydratase from a combination of ligand- and structure-based and deep learning in silico approaches.Journal of Biomolecular Structure and DynamicsFebruary 2022, 1-10
  HALDOIback to text
• 18 articleA.Athénaïs Vaginay, T.Taha Boukhobza and M.Malika Smaïl-Tabbone. From quantitative SBML models to Boolean networks.Applied Network Science71December 2022, 73
  HALDOIback to textback to text
• 19 articleM.Marcelo Veras, B.Bishnu Sarker, S.Sabeur Aridhi, J.João Gomes, J.José Macêdo, E. M.Engelbert Mephu Nguifo, M.-D.Marie-Dominique Devignes and M.Malika Smaïl-Tabbone. On the design of a similarity function for sparse binary data with application on protein function annotation.Knowledge-Based Systems238February 2022, 107863
  HALDOIback to text

International peer-reviewed conferences

• 20 inproceedingsA.Antoine Moniot, I.Isaure Chauvot de Beauchêne and Y.Yann Guermeur. Inferring Epsilon-nets of Finite Sets in a RKHS.Proceedings of the 14th International Workshop, Lecture Notes in Networks and Systems - LNNSAdvances in Self-Organizing Maps, Learning Vector Quantization, Clustering and Data Visualization. WSOM - 2022533Prague, Czech RepublicSpringerAugust 2022, 53-62
  HALDOIback to text
• 21 inproceedingsJ.Joachim Niehren, A.Athénaïs Vaginay and C.Cristian Versari. Abstract Simulation of Reaction Networks via Boolean Networks.CMSB2022: 20th International Conference on Computational Methods in Systems BiologyBucarest, RomaniaSpringerSeptember 2022
  HALback to text

Other scientific publications

• 22 inproceedingsD.Diego Amaya-Ramirez, R.Romain Lhotte, M.Magali Devriese, C.Constantin Hays, J.-L.Jean-Luc Taupin and M.-D.Marie-Dominique Devignes. Reduced structural flexibility of eplet amino acids in HLA proteins.ECCB 2022 21st European Conference on Computational Biology Planetary Health and BiodiversityBarcelona, SpainSeptember 2022
  HALback to text
• 23 miscE.Emmanuel Bresso, P.Pierre Monnin, C.Cédric Bousquet, F.-É.François-Élie Calvier, N. C.Ndeye Coumba Ndiaye, N.Nadine Petitpain, M.Malika Smaïl-Tabbone and A.Adrien Coulet. Investigating ADR mechanisms with Explainable AI: a feasibility study with knowledge graph mining.Rennes, FranceJuly 2022
  HALback to text
• 24 inproceedingsH.Hrishikesh Dhondge, I.Isaure Chauvot de Beauchêne and M.-D.Marie-Dominique Devignes. CroMaSt: A workflow for domain family curation through cross-mapping of structural instances between protein domain databases.ECCB2022- 21st European Conference on Computational BiologySitges, SpainSeptember 2022
  HALDOIback to text