Dose-response prediction for in-vitro drug combination datasets: a probabilistic approach

Abstract

In this paper we propose PIICM, a probabilistic framework for dose-response prediction in high-throughput drug combination datasets. PIICM utilizes a Permutation-Invariant version of the Intrinsic Co-regionalization Model for multi-output Gaussian Process regression, to predict dose-response surfaces in untested drug combination experiments. Coupled with an observation model that incorporates experimental uncertainty, PIICM is able to learn from noisily observed cell-viability measurements in settings where the underlying dose-response experiments are of varying quality, utilize different experimental designs, and the resulting training dataset is sparsely observed. We show that the model can accurately predict dose-response in held-out experiments, and the resulting function captures relevant features indicating synergistic interaction between drugs.

Predicting drug combination response surfaces

“…With PIICM we use the same BRAID surfaces as the baseline models as in our comboKR. Like in [23], we consider a subset of concentrations to represent the surfaces since the method is computationally too heavy to run with full set of concentrations in the data (over 60 unique dose values), and the BRAID surfaces are sampled at these concentrations as training data to the model. As the concentrations are assumed to be normalised to [0, 1] interval, we consider the same normalisation scheme for PIIMC, as we use to obtain our normalised kernel evaluations.…”

“…We note that the PIICM method [23] is based on Gaussian processes, modelling the drug interaction surfaces and without considering any drug features when making predictions. Thus, it is not expected to be able to generalise to data outside of the training set in the new drug scenario, further than predicting some generic interactions derived from the other drugs.…”

“…• Important for novel drug combination discovery, our proposed method can be applied to the new drug settings without the need to re-train the model or measure each drug response beforehand. • In comparison to the baseline comboLTR method [21], and another surface-valued prediction approach, PIICM [23], we show that comboKR achieves superior results especially in the more challenging predictive scenario, where testing is performed on drugs not available in the training stage.…”

“…Recently, Rønneberg et al [23] proposed an approach called PIICM, that considers predicting full response surfaces instead of individual dose-combination responses. They proposed a probabilistic prediction model based on Gaussian process regression where the covariance matrices for pairwise drug interactions are parameterised and learned.…”

“…In this work, we propose a novel approach for predicting directly the full continuous drug combination dose-response surfaces with a kernel-based structured output prediction model, called comboKR. In contrast to the PIICM model [23], comboKR is based on an inductive learning approach, which predicts the drug combination response surface from input drug features that are easily available from drug databases. Moreover, to overcome the practical issues arising from the heterogeneous experimental design often used in drug combination response measurements, we propose a novel normalisation scheme for comparing drug interaction surfaces.…”