Noise-precision tradeoff in predicting combinations of mutations and drugs

Many biological problems involve the response to multiple perturbations. Examples include response to combinations of many drugs, and the effects of combinations of many mutations. Such problems have an exponentially large space of combinations, which makes it infeasible to cover the entire space experimentally. To overcome this problem, several formulae that predict the effect of drug combinations or fitness landscape values have been proposed. These formulae use the effects of single perturbations and pairs of perturbations to predict triplets and higher order combinations. Interestingly, different formulae perform best on different datasets. Here we use Pareto optimality theory to quantitatively explain why no formula is optimal for all datasets, due to an inherent bias-variance (noise-precision) tradeoff. We calculate the Pareto front of log-linear formulae and find that the optimal formula depends on properties of the dataset: the typical interaction strength and the experimental noise. This study provides an approach to choose a suitable prediction formula for a given dataset, in order to best overcome the combinatorial explosion problem.Author summary: Sometimes a combination of drugs works much better than each drug alone. Finding such drug cocktails is a pressing challenge in order to combat drug resistance and to improve drug effects. However, it is impossible to test all combinations of multiple drug experimentally. Therefore, researchers are looking for computational rather than experimental approaches to overcome this problem. One approach is to measure the effect of few drugs and plug it into a formula that predicts the effect of many drugs together. Existing prediction formulae typically perform best on the dataset that they were developed on, but less well on other datasets. Here we explain this observation and give a guide for the choice of an optimal prediction formula for a given dataset. The optimal formula depends on two main properties of the dataset: 1) The interaction strength between the drugs and 2) The experimental noise in the data. This study may help researchers discover effective combinations of multiple drugs and multiple perturbations in general.