The undesirable effects of drugs (ADR) are an important cause of admission to the hospital and stopping treatment worldwide. Conventional approaches often do not manage to detect the rare or delayed effects of drugs. In order to improve early detection, a research team from the Sofia Medical University has developed an in -depth learning model to predict the probability of MADR based solely on the chemical structure of a drug.
The model was built using a neural network formed using reference pharmacovigilance data. The input characteristics have been derived from smile codes – a standard format representing the molecular structure. Predictions have been generated for six major ADRs: hepatotoxicity, nephrotoxicity, cardiotoxicity, neurotoxicity, hypertension and photosensitivity.
“We could conclude that he has succeeded in identifying many expected reactions while producing relatively few false positives,” write researchers in their article published in the journal Pharmacyconcluding that he “demonstrates an acceptable clarification in the forecast of the ADR”.
The model tests with well -characterized drugs have led to coherent predictions with known side effects. For example, he estimated a probability of hepatotoxicity of 94.06% for erythromycin, 88.44% for nephrotoxicity and 75.8% for hypertension in cisplatin. In addition, 22% photosensitivity was planned for cisplatin, while 64.8% of photosensitivity was estimated for the experimental compound Ezeprogind. For enadolina, a new molecule, the model has returned low probability scores in all ADRs, suggesting a minimum risk.
In particular, these results demonstrate the potential of the model as a decision -making tool in the discovery of drugs in early phase and monitoring regulatory security. The authors recognize that the performance of the infrastructure could be further improved by incorporating factors such as dose levels and parameters specific to the patient.