Reciprocal perspective as a super learner improves drug-target interaction prediction (MUSDTI)

Abstract

The identification of novel drug-target interactions (DTI) is critical to drug discovery and drug repurposing to address contemporary medical and public health challenges presented by emergent diseases. Historically, computational methods have framed DTI prediction as a binary classification problem (indicating whether or not a drug physically interacts with a given protein target); however, framing the problem instead as a regression-based prediction of the physiochemical binding affinity is more meaningful. With growing databases of experimentally derived drug-target interactions (e.g. Davis, Binding-DB, and Kiba), deep learning-based DTI predictors can be effectively leveraged to achieve state-of-the-art (SOTA) performance. In this work, we formulated a DTI competition as part of the coursework for a senior undergraduate machine learning course and challenged students to generate component DTI models that might surpass SOTA models and effectively combine these component models as part of a meta-model using the Reciprocal Perspective (RP) multi-view learning framework. Following 6 weeks of concerted effort, 28 student-produced component deep-learning DTI models were leveraged in this work to produce a new SOTA RP-DTI model, denoted the Meta Undergraduate Student DTI (MUSDTI) model. Through a series of experiments we demonstrate that (1) RP can considerably improve SOTA DTI prediction, (2) our new double-cold experimental design is more appropriate for emergent DTI challenges, (3) that our novel MUSDTI meta-model outperforms SOTA models, (4) that RP can improve upon individual models as an ensembling method, and finally, (5) RP can be utilized for low computation transfer learning. This work introduces a number of important revelations for the field of DTI prediction and sequence-based, pairwise prediction in general.

Figure 1

Conceptual overview of the proposed MUSDTI predictor.

Figure 2

Experimental design to evaluate the proposed MUSDTI predictor.

Figure 3

Example paired one-to-all score curves. An example pair demonstrating dramatically differing distributions is depicted to emphasize that even though a given drug scores relatively low in the given protein target perspective, that protein is the top-scoring target for that specific drug.

Figure 4

Inference rates of each component model measured over random subsets of 1 million pairs.

Figure 5

Component model performance improvement from the reciprocal perspective cascaded layer over the double-cold dataset.

Figure 6

Component model performance improvement from the reciprocal perspective cascaded layer over the DeepDTA-defined dataset.

Figure 7

Experimental results over the DeepDTA-defined datasets when incrementally incorporating reciprocal perspective component models compared to the SOTA DeepDTA models. The top-performing combined models were circled in the figure (top-20 models) and the first (top-10 models) represent the performance of the proposed MUSDTI model even when the later combined models represent a marginally higher performance. We opted for the component model ensemble that represented the plateaued performance of component models.

Figure 8

Shapeley additive features analysis. The x-axis is sorted left-to-right in increasing magnitude of SHAP value summed over the column while the y-axis is sorted top-down in increasing magnitude of SHAP value summed over the row. Emanating out from the bottom-right are the models and features with increasingly lesser impact on the model decision. Only the top-10 models contributing to the MUSDTI model are depicted along all 14 RP features.