Limitations of use of machine learning/deep learning in drug development

Many branches of the industry are joining the hype around using deep neural networks or complex machine learning approaches, eg gradient boosting machines, random forests. However, I am sceptical that these “black box” algorithms will make major contributions to¬† drug development, since the problems in clinical trials seems conceptually different from, eg image or speech recognition. Using autonomous driving as an example, the problme of drug development is actually, whether moving the steering wheel (cf drug) influences the probabilty that the car leaves the road (eg blood sugar levels out of range) among multiple predictors which may affect this outcome (eg shape of road, speed of the car, side winds, being towed off). While currently some variety of linear models are used to explicitly analyse the effect of the drug on the outcome, the importance/influence (clinical relevance) of a single predictor among very many (correlated) predictors is difficult to evaluate with complex machine learning/deep learning methods, let alone its statistical significance.

Potential solutions may be to re-formulate the problem by trying to predict the treatment group from outcome (and covariates), which follows the rationale that if one can predict the treatment from the outcome, the treatment must have had an effect on it.

Alternatively, one may try to switch from modelling the signal (drug effect) to modelling the noise (effects of covariates) in data-rich clinical trials (omics data, sensor data, etc), predict the outcomes, and compare the residuals (cf virtual twin approach).

See also similar blog by Pablo Cordero (Hyperparameter):