Hardcoding feature weights into the model is an absolute no. What you want to do is implement a weighted version of Random Forest, where the decision trees are iteratively grown by selecting features with probabilities corresponding to their Gini impurity index.

Here are a couple of papers that did something along these lines - [1] https://academic.oup.com/bioinformatics/article/24/18/2010/190849?login=true [2] https://www.pnas.org/doi/10.1073/pnas.1711236115

Hope this helps.