[abhikshah]

ML is fairly common in genomics, but for identifying predictive variables for cancer status, it's difficult. The training set is a matrix where rows are people (where some have cancer and some don't) and columns are genomic features (mutation, methylation, etc). You can easily have hundreds of thousands of features but getting even a thousand cancer patients enrolled in a study and sequenced is expensive and slow.

So, even though there are many "AI in biotech" companies out there, for predicting cancer status, most eventually end up hand crafting a small number of features based on extensive knowledge of cancer biology. The ML model tends to be simple and far less important than the features.