Mining Real-World Big Data to Characterize Adverse Drug Reaction Quantitatively: Mixed Methods Study.

BACKGROUND: Adverse drug reactions (ADRs), which are the phenotypic manifestations of clinical drug toxicity in humans, are a major concern in precision clinical medicine. A comprehensive evaluation of ADRs is helpful for unbiased supervision of marketed drugs and for discovering new drugs with high success rates. OBJECTIVE: In current practice, drug safety evaluation is often oversimplified to the occurrence or nonoccurrence of ADRs. Given the limitations of current qualitative methods, there is an urgent need for a quantitative evaluation model to improve pharmacovigilance and the accurate assessment of drug safety. METHODS: In this study, we developed a mathematical model, namely the Adverse Drug Reaction Classification System (ADReCS) severity-grading model, for the quantitative characterization of ADR severity, a crucial feature for evaluating the impact of ADRs on human health. The model was constructed by mining millions of real-world historical adverse drug event reports. A new parameter called Severity_score was introduced to measure the severity of ADRs, and upper and lower score boundaries were determined for 5 severity grades. RESULTS: The ADReCS severity-grading model exhibited excellent consistency (99.22%) with the expert-grading system, the Common Terminology Criteria for Adverse Events. Hence, we graded the severity of 6277 standard ADRs for 129,407 drug-ADR pairs. Moreover, we calculated the occurrence rates of 6272 distinct ADRs for 127,763 drug-ADR pairs in large patient populations by mining real-world medication prescriptions. With the quantitative features, we demonstrated example applications in systematically elucidating ADR mechanisms and thereby discovered a list of drugs with improper dosages. CONCLUSIONS: In summary, this study represents the first comprehensive determination of both ADR severity grades and ADR frequencies. This endeavor establishes a strong foundation for future artificial intelligence applications in discovering new drugs with high efficacy and low toxicity. It also heralds a paradigm shift in clinical toxicity research, moving from qualitative description to quantitative evaluation.

Convolutional Neural Network Visualization for Identification of Risk Genes in Bipolar Disorder.

BACKGROUND: Bipolar disorder (BD) is a type of chronic emotional disorder with a complex genetic structure. However, its genetic molecular mechanism is still unclear, which makes it insufficient to be diagnosed and treated. METHODS AND RESULTS: In this paper, we proposed a model for predicting BD based on single nucleotide polymorphisms (SNPs) screening by genome-wide association study (GWAS), which was constructed by a convolutional neural network (CNN) that predicted the probability of the disease. According to the difference of GWAS threshold, two sets of data were named: group P001 and group P005. And different convolutional neural networks are set for the two sets of data. The training accuracy of the model trained with group P001 data is 96%, and the test accuracy is 91%. The training accuracy of the model trained with group P005 data is 94.5%, and the test accuracy is 92%. At the same time, we used gradient weighted class activation mapping (Grad-CAM) to interpret the prediction model, indirectly to identify high-risk SNPs of BD. In the end, we compared these high-risk SNPs with human gene annotation information. CONCLUSION: The model prediction results of the group P001 yielded 137 risk genes, of which 22 were reported to be associated with the occurrence of BD. The model prediction results of the group P005 yielded 407 risk genes, of which 51 were reported to be associated with the occurrence of BD.