Studying and mitigating the effects of data drifts on ML model performance at the example of chemical toxicity data