In a recent Nature publication by Algavi and Elhanan, faculties from Tel Aviv University, Israel, have tackled the challenge of fully understanding the adverse effects of drugs on the gut microbiome. By combining machine learning with the chemical characteristics of drugs and the genetic and metabolic pathways of bacteria, they developed a powerful approach to quantifying the anti-commensal properties and side effects of drugs on the microbial community. With an impressive ROC AUC (area under the receiver-operator curve) of 0.92, their model demonstrates exceptional predictive performance in identifying new drug-microbe interactions in vitro. This groundbreaking research offers promising insights into mitigating the potential harms caused by drug-induced perturbations in the gut microbiological ecosystem.
Comprehension of Drug-Microbe Interactions Can Revolutionize the Healthcare System
Humans and the microflora community simultaneously share an intertwined as well as antagonistic relationship. Microbiota maintains homeostasis, for example, by facilitating individualistic beneficial responses to diet habits; they are also often associated with health-related complications such as hypertension, inflammatory diseases, cancer, metabolic disorders, and so on; due to dysbiosis or aberrations in their populations.
The microbiota directly influences the pharmacokinetics, metabolism, toxicology, and overall efficacy of therapeutics. Even though researching the intricate details of human-microbe association is in its inception phase, considering the aspect of the microflora community has been proven effective wonderfully in multiple therapies.
Concomitantly, drugs also influence the microbial members inhabiting our body, manifesting multiple side effects as a repercussion. Transient variation in gut microbiota following drug administration has been proven in clinical studies. For instance, the paper mentions a medicine, metamorphin, prescribed in Type 2 diabetic patients, that often causes bloating, diarrhea, and nausea as it induces an abundance of virulent E. coli strains.
Therefore, a comprehensive perspective about microbial ecosystem diversion by means of medication is a critical respect that, if taken care of during drug development and prescription, will result in tremendous benefits to patients, and the public health system will progress more towards personalized medication approaches eradicating undesired drug-induced side effects.
The Machine Learning Model
As the authors mention, although several previous studies have researched drug impacts on microbiota, a generalized understanding of this aspect has yet to be unlocked. Moreover, integrating regular clinical practices with drug-microbe interaction analysis is not easily feasible yet. Henceforth, the researchers develop a computational model to complement further and supplement the understanding of microbiota and drug relationships.
Algavi and Elhanan implement machine-learning approaches in which each microbe-drug pair is represented as a vector of features. A microbe strain is described as a set of its genome-encoded biochemical pathways, and drugs are denoted by their physical-chemical and structural features. The model operates 148 microbial features from KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways and 92 drug features as SMILE (Simplified Molecular Input Line Entry System) representations.
The researchers trained Random Forest and Supported Vector Machines (SVMs) Machine learning models with previously in-vitro screened 1197 compounds against 40 microbes, either implementing polynomial or radial basis function (RBF) kernels, and three regularized logistic regression models (ridge, lasso, and elastic net) with the default hyperparameter. The model then predicts a numerical value between 0 and 1, where “0” indicates no effect of the drug in microbe growth inhibition and “1” means growth inhibition.
A permutation importance method is applied to calculate the statistical significance and contribution of each feature in predicting the possible interactions with drugs. Lipophilicity and charge distribution of the compound is found to be significant contributors to the antimicrobial effect certainly because these factors influence the permeability of bacterial membranes. Indole, which is the main byproduct of tryptophan metabolism, causes antibiotic tolerance to bacteria. Lipopolysaccharides and ABC transporters that attribute membrane features have high importance in this aspect.
The Excellency of the Model in Predictive Performance
The Random Forest model predicted new drug–microbe interactions in vitro with an area-under-the-receiver-operator curve (ROC AUC) of 0.972. The model successfully computes the effects of new drugs on different microbial strains and vice versa. Moreover, the model doesn’t solely rely on the chemical similarity of drugs during predictions, as it executed favorable predictive performance when drugs with structural similarities were removed. The model showed good predictive power even when it included drugs with different bioactivity.
In the same way, checking with the ‘leave-one-microbe-out approach,’ the model displayed accuracy in the prediction of the influence of new microbial strain. The model proved its robustness in this aspect when all microbes belonging to the same phylum were excluded from the training set as the ROC AUC was > 0.92. Also, when the model was validated with different datasets, it proved its universal potential for predicting drug-target interactions.
To capture the diversity of intestinal microbiota, the researchers collected metagenomic data from a healthy population of fecal microbiota transplant donors who were not exposed to antibiotics or other drug treatments. Clinically approved small-molecule drugs from the DrugBank database (2585 compounds) were collected, and their physio-chemical properties were determined. The model is trained with the data and predicted impact of 2585 drugs on 409 human microbiota members and generated a catalog of 1,057,265 drug–microbe interactions. Among them, more than 62% of the drugs and 90% of the microbial taxa have not been tested in vitro before. Drug-Microbe interaction patterns were identified from the catalog and also obtained an Anatomical Therapeutic Chemical (ATC) classification of each drug discerning information regarding the physiological system targeted by the drugs.
They examined the model’s efficacy in predicting microbiome alteration in human clinical trials. For example, the model is able to capture the effect of the drug Omeprazole correctly when compared to the microbiome composition pre and post-treatment with the drug. Furthermore, compared with summary statistics data from the ‘Lifelines Dutch microbiome project’ regarding the adverse effects of drugs on microbial growth, the model displayed statistically significant and agreeable results.
The researchers validated the predictions through experimental data, clinical studies, and population cohorts throughout their work. Although the researchers can’t ignore the limitation owing to the limited availability of drug-microbe interaction data. Nonetheless, the model will help researchers develop drugs, understand individualistic responses to drugs attributed to the microflora population in detail, and design personalized pharmaceuticals.
Article Source: Reference Paper
Aditi is a consulting scientific writing intern at CBIRT, specializing in explaining interdisciplinary and intricate topics. As a student pursuing an Integrated PG in Biotechnology, she is driven by a deep passion for experiencing multidisciplinary research fields. Aditi is particularly fond of the dynamism, potential, and integrative facets of her major. Through her articles, she aspires to decipher and articulate current studies and innovations in the Bioinformatics domain, aiming to captivate the minds and hearts of readers with her insightful perspectives.