Our publications

We present a multi-omic signature for COVID-19 developed through a comprehensive Quebec initiative that established an extensive dataset of COVID-19 positive and negative patient samples.Moving beyond traditional symptomatic studies that rely on limited descriptors, our research integrates clinical, proteomic, and metabolomic data to classify COVID-19 status using thousands of features. Our multi-view machine learning approach extracts distinctive COVID-19 signatures from multi-omic data with remarkable effectiveness. By applying ensemble methods, we developed accurate and interpretable models for high-dimensional data-containing significantly more features than samples-achieving 89% ± 5% balanced accuracy. Through our novel feature relevance methodology, we identified condensed 12-and 50-feature signatures that enhanced classification accuracy by at least 3% compared to the original feature set. This approach successfully extracted and interpreted a robust multi-omic signature characterizing COVID-19positive individuals from a large, complex dataset, representing a significant advancement in COVID-19 biomarker discovery.

Affiliations

PubMed Disclaimer

Abstract

The market for illicit drugs has been reshaped by the emergence of more than 1100 new psychoactive substances (NPS) over the past decade, posing a major challenge to the forensic and toxicological laboratories tasked with detecting and identifying them. Tandem mass spectrometry (MS/MS) is the primary method used to screen for NPS within seized materials or biological samples. The most contemporary workflows necessitate labor-intensive and expensive MS/MS reference standards, which may not be available for recently emerged NPS on the illicit market. Here, we present NPS-MS, a deep learning method capable of accurately predicting the MS/MS spectra of known and hypothesized NPS from their chemical structures alone. NPS-MS is trained by transfer learning from a generic MS/MS prediction model on a large data set of MS/MS spectra. We show that this approach enables a more accurate identification of NPS from experimentally acquired MS/MS spectra than any existing method. We demonstrate the application of NPS-MS to identify a novel derivative of phencyclidine (PCP) within an unknown powder seized in Denmark without the use of any reference standards. We anticipate that NPS-MS will allow forensic laboratories to identify more rapidly both known and newly emerging NPS. NPS-MS is available as a web server at https://nps-ms.ca/, which provides MS/MS spectra prediction capabilities for given NPS compounds. Additionally, it offers MS/MS spectra identification against a vast database comprising approximately 8.7 million predicted NPS compounds from DarkNPS and 24.5 million predicted ESI-QToF-MS/MS spectra for these compounds.

Affiliations

© The Author(s) 2023. Published by Oxford University Press on behalf of Nucleic Acids Research.

PubMed Disclaimer

Affiliations

PubMed Disclaimer

Affiliations

PubMed Disclaimer

Affiliations

© The Author(s) 2023. Published by Oxford University Press on behalf of Nucleic Acids Research.

PubMed Disclaimer

Abstract

Ensemble methods are a very diverse family ofalgorithms with a wide range of applications. Oneof the most commonly used is boosting, with theprominent Adaboost. Adaboost relies on greed-ily learning base classifiers that rectify the errorfrom previous iterations. Then, it combines themthrough a weighted majority vote, based on theirquality on the entire learning set. In this paper, wepropose a supervised binary classification frame-work that propagates the local knowledge acquiredduring the boosting iterations to the predictionfunction. Based on this general framework, weintroduce SamBA, an interpretable greedy ensem-ble method designed for fat datasets, with a largenumber of dimensions and a small number of sam-ples. SamBA learns local classifiers and combinesthem, using a similarity function, to optimize its ef-ficiency in data extraction. We provide a theoreticalanalysis of SamBA, yielding convergence and gen-eralization guarantees. In addition, we highlightSamBA’s empirical behavior in an extensive exper-imental analysis on both real biological and gen-erated datasets, comparing it to state-of-the-art en-semble methods and similarity-based approaches

Abstract

SuMMIT (Supervised Multi Modal Integration Tool) is a software offering manyfunctionalities for running, tuning, and analyzing experiments of supervised classificationtasks specifically designed for multi-view data sets. SuMMIT is part of a platform 1 thataggregates multiple tools to deal with multiview datasets such as scikit-multimodallearn(Benielli et al., 2021) or MAGE (Bauvin et al., 2021). This paper presents use cases ofSuMMIT, including hyper-parameters optimization, demonstrating the usefulness of sucha platform for dealing with the complexity of multi-view benchmarking on an imbalanceddataset. SuMMIT is powered by Python3 and based on scikit-learn, making it easy touse and extend by plugging one’s own specific algorithms, score functions or adding newfeatures2. By using continuous integration, we encourage collaborative development.Keywords: Multimodal, Supervised, Classification, Benchmarking, Python, ReproducibleResearch, Modularity, Explainability, Interpretability

Identification of proteins is one of the most computationally intensive steps in genomics studies. Itusually relies on aligners that do not accommodate rich information on proteins and require additionalpipelining steps for protein identification. We introduce kAAmer, a protein database engine based onamino‑acid k‑mers that provides efficient identification of proteins while supporting the incorporationof flexible annotations on these proteins. Moreover, the database is built to be used as a microservice,to be hosted and queried remotely.

One fundamental task in genomics is the identification and annotation of DNA coding regions that translate intoproteins via a genetic code. Protein databases increase in size as new variants, orthologous and novel genes, oftenfound in metagenomics studies, are being sequenced. This is particularly true within the microbial world wherebacterial proteomes’ diversity follows their rapid evolution. For instance, UniProtKB (Swiss-Prot/TrEMBL)1 andNCBI RefSeq 2 contain over 100 million bacterial proteins and that number is increasing rapidly.Identification of proteins often relies on accurate, but slow, alignment software such as BLAST or hiddenMarkov model (HMM) profile-based software 3,4 . Although other approaches (such as DIAMOND 5 ) have con-siderably improved the speed of searching proteins in large datasets, from a database standpoint much can bedone to offer a more versatile experience. One such approach would be to expose the database as a permanentservice, which can make use of computational resources for increased performance (e.g. memory mapping)and leveraging the cloud for remote analyses via a HTTP API. Another approach would be to extend the resultset with comprehensive information on protein targets to facilitate subsequent genomics and metagenomicsanalysis pipelines.

Abstract

The Human Metabolome Database or HMDB (https://hmdb.ca) has been providing comprehensive ref-erence information about human metabolites andtheir associated biological, physiological and chemi-cal properties since 2007. Over the past 15 years, theHMDB has grown and evolved significantly to meetthe needs of the metabolomics community and re-spond to continuing changes in internet and comput-ing technology. This year’s update, HMDB 5.0, bringsa number of important improvements and upgradesto the database. These should make the HMDB moreuseful and more appealing to a larger cross-sectionof users. In particular, these improvements include:(i) a significant increase in the number of metabo-lite entries (from 114 100 to 217 920 compounds); (ii)enhancements to the quality and depth of metabolitedescriptions; (iii) the addition of new structure, spec-tral and pathway visualization tools; (iv) the inclusionof many new and much more accurately predictedspectral data sets, including predicted NMR spec-tra, more accurately predicted MS spectra, predictedretention indices and predicted collision cross sec-tion data and (v) enhancements to the HMDB’s searchfunctions to facilitate better compound identification

Abstract

Background: Deep learning methods are a proven commodity in many fields andendeavors. One of these endeavors is predicting the presence of adverse drug–druginteractions (DDIs). The models generated can predict, with reasonable accuracy,the phenotypes arising from the drug interactions using their molecular structures.Nevertheless, this task requires improvement to be truly useful. Given the complexityof the predictive task, an extensive benchmarking on structure-based models for DDIsprediction was performed to evaluate their drawbacks and advantages.

Results: We rigorously tested various structure-based models that predict drug inter-actions using different splitting strategies to simulate different real-world scenarios.In addition to the effects of different training and testing setups on the robustnessand generalizability of the models, we then explore the contribution of traditionalapproaches such as multitask learning and data augmentation

Triple negative breast cancer (TNBC) is one of the most aggressive form of breast cancer (BC) with thehighest mortality due to high rate of relapse, resistance, and lack of an effective treatment. Variousmolecular approaches have been used to target TNBC but with little success. Here, using machinelearning algorithms, we analyzed the available BC data from the Cancer Genome Atlas Network(TCGA) and have identified two potential genes, TBC1D9 (TBC1 domain family member 9) and MFGE8(Milk Fat Globule‑EGF Factor 8 Protein), that could successfully differentiate TNBC from non‑TNBC,irrespective of their heterogeneity. TBC1D9 is under‑expressed in TNBC as compared to non‑TNBCpatients, while MFGE8 is over‑expressed. Overexpression of TBC1D9 has a better prognosis whereasoverexpression of MFGE8 correlates with a poor prognosis. Protein–protein interaction analysis byaffinity purification mass spectrometry (AP‑MS) and proximity biotinylation (BioID) experimentsidentified a role for TBC1D9 in maintaining cellular integrity, whereas MFGE8 would be involved invarious tumor survival processes. These promising genes could serve as biomarkers for TNBC anddeserve further investigation as they have the potential to be developed as therapeutic targets forTNBC.

Mass spectrometry is a valued method to evaluate the metabolomics content of a biological sample.The recent advent of rapid ionization technologies such as Laser Diode Thermal Desorption (LDTD) andDirect Analysis in Real Time (DART) has rendered high-throughput mass spectrometry possible. It isused for large-scale comparative analysis of populations of samples. In practice, many factors resultingfrom the environment, the protocol, and even the instrument itself, can lead to minor discrepanciesbetween spectra, rendering automated comparative analysis difficult. In this work, a sequence/pipelineof algorithms to correct variations between spectra is proposed. The algorithms correct multiple spectraby identifying peaks that are common to all and, from those, computes a spectrum-specific correction.We show that these algorithms increase comparability within large datasets of spectra, facilitatingcomparative analysis, such as machine learning.

Abstract

Untargeted metabolomic measurements using mass spectrometry are a powerful tool foruncovering new small molecules with environmental and biological importance. The smallmolecule identification step, however, still remains an enormous challenge due to fragmentationdifficulties or unspecific fragment ion information. Current methods to address this challenge areoften dependent on databases or require the use of nuclear magnetic resonance (NMR), whichhave their own difficulties. The use of the gas-phase collision cross section (CCS) values obtainedfrom ion mobility spectrometry (IMS) measurements were recently demonstrated to reduce thenumber of false positive metabolite identifications. While promising, the amount of empirical CCSinformation currently available is limited, thus predictive CCS methods need to be developed. Inthis article, we expand upon current experimental IMS capabilities by predicting the CCS valuesusing a deep learning algorithm. We successfully developed and trained a prediction model forCCS values requiring only information about a compound’s SMILES notation and ion type. Theuse of data from five different laboratories using different instruments allowed the algorithm to betrained and tested on more than 2400 molecules. The resulting CCS predictions were found toachieve a coefficient of determination of 0.97 and median relative error of 2.7% for a wide range

Understanding the relationship between the genome of a cell and its phenotype is a central problemin precision medicine. Nonetheless, genotype-to-phenotype prediction comes with great challengesfor machine learning algorithms that limit their use in this setting. the high dimensionality of thedata tends to hinder generalization and challenges the scalability of most learning algorithms.Additionally, most algorithms produce models that are complex and difficult to interpret. We alleviatethese limitations by proposing strong performance guarantees, based on sample compression theory,for rule-based learning algorithms that produce highly interpretable models. We show that theseguarantees can be leveraged to accelerate learning and improve model interpretability. our approachis validated through an application to the genomic prediction of antimicrobial resistance, an importantpublic health concern. Highly accurate models were obtained for 12 species and 56 antibiotics, andtheir interpretation revealed known resistance mechanisms, as well as some potentially new ones. Anopen-source disk-based implementation that is both memory and computationally efficient is providedwith this work. The implementation is turnkey, requires no prior knowledge of machine learning, and iscomplemented by comprehensive tutorials.