Suggestion for workflow to predict pathogenicity of SNV and INDELs

Question

[luffy]

I. Prediction using ML medthods

Methods to use Support vector machine (Random Forest?)

• Download variant data from Clinvar, other DBs and from literature
• Remove unnecessary columns, missing data and keep c.change, p.change, clinical significance, etc.
• Use sklearn/pandas(onehotencoder/getdummies) to convert data into binary
• for Position Specific Method, take c.change/p.change remove c/p and make 3 columns(wild,loc,new) and convert all other charaters into strings
• split data into train and test and classifiy using SVC
• draw confusion matrix and check
• determine cross vaildation using gridsearchcv
• draw confusion matrix and check
• classifiy with the newly determined c and gamma
• classifiy with patients variants

II. Prediction of effects of variants with known prediction tools

• compare with results of previous step (tools eg. SIFT, polyphen etc)

III. Compare and corelate both results with disease severity - compare and corelate results first two steps with the patient's disease severity and infarance

Questions

• Does this workflow make sense?
• Any suggestion/advice/opinion to imporve the workflow
• For small datasets SVM and random forest(since it uses decision tree) is better when compared to others?
• Can i use only pathogenic and likely pathogenic or use variants of uncertain significance too for training?

Objective is to predict/corelate pathogenicity of variant with patient's phenotype/severity

why i am predicting rather then using known prediction tools is because i would be using SVM with Position Specific Method on a rare germline disorder with specific genes which are known to cause the disorder.

Thanks in advance for your time