Protein interaction network topology predicts in vitro synergy comparable to approaches trained with experimental data

Introduction: Combination therapies are often regarded as the future for treating cancer. However, anticipating the synergistic effects that arise from combined perturbations (i.e., drug combinations) is challenging as it requires an intricate understanding of drug targets’ connections to a disease’s dysregulated molecular pathways. In fact, even sophisticated machine learning methods trained on ample single perturbation and cell line information have reported model-experiment Pearson’s correlation coefficients (PCC) of ~0.24-0.48, with better performing models leveraging drug target and pathway information. Further, network-based methods have successfully been applied to the protein-protein interaction network (interactome) in order to identify disease/signaling pathways, discover new drug targets, and even predict drug synergies. However, to our knowledge, few have considered de novo synergy prediction solely from interaction network topology.

Materials and

Methods: We implemented a network approach to measure the extent to which topological relationships between protein target pairs can be used to predict their combined perturbation effects. First, we used a network diffusion algorithm to model how a single target perturbation can propagate in the interactome and affect the activity of nearby proteins. Expanding this to consider the effects of a dual target perturbation, we measured the relative topological positions (i.e., distance) of a combination’s targets (or target neighborhoods) within the interactome. Given the interactome’s complex topology, we exhaustively tested 36 diffusion-based distance metrics, each encoded by a unique set of network features that quantified various assumptions of diffusion depth, the importance of target neighborhoods, and signal propagation directionality.

To test our framework, we first applied our interactome distances to target pairs screened in a pairwise CRISPR knockout experiment conducted in a chronic myeloid leukemia (CML) cell line. Within this screen, we extracted the most toxic and synergistic target pairs (n = 57) as our validation set. We then turned to five additional breast cancer cell line validation sets derived from a multi-target small molecule screen to determine if our framework extended to combinations with weaker, more additive in vitro signals. Notably, because multiple target pair distances corresponded to a single small molecule combination, we employed a curve fitting procedure to identify the single target pair that best explained a combination’s cell line-specific experimental effects. We implemented this curve fitting procedure with all 36 diffusion-based distances and a comparator shortest-paths measurement.

Results, Conclusions, and Discussions: First, across all diffusion-based distances, we found that our validation set within the CML model was significantly closer than all tested target pairs (t-test, P < 0.05). This suggests that target closeness may sufficiently explain strong experimental effects. Additionally, we found that for these 57 target pairs, there existed a moderately negative correlation between distance and synergy score (PCC = ~-0.40). Turning to the fitted curves within the breast cancer cell line models, we found that nearly all diffusion-based distance metrics were better aligned with synergy compared to shortest-paths measurements, implying that diffusion methods may capture more biologically relevant connections. The optimal diffusion-based distance metric, defined by its high accuracy and robustness across fitted curves, yielded correlations of -0.35 – -0.58, corroborating our findings in the CML model. Separately, as each model was fit per cell line, we identified unique target pair interaction(s) that we speculate may drive a combination’s cell line-specific observed effect (Figure 1).

We next sought to uncover whether the optimal metric’s performance was dependent on any of its encoded network features by quantifying the sensitivity of its synergy alignment to single feature variations. First, distance-synergy correlations were consistently sensitive to increases in the diffusion depth feature with significant trends observed in 3/5 cell line models (asymptotic z-test, P < 0.05). This suggests that in the context of drug combinations, drug efficacy is limited to local target neighborhoods. Similarly, alignments were significantly sensitive to the exclusion of target-neighbor interactions in 2/5 cell line models (asymptotic z-test, P < 0.05), implying that pathway-driven rather than target-centric distance characterizations more adequately represent the complexities of intracellular signaling. While there weren’t strong trends for the signal propagation directionality feature, overall, these preliminary insights can be used to guide other network studies toward topology-informed representations of network signaling.

Here, we demonstrated that in the context of predicting combined perturbation effects, network topology alone is as performant as experiment-informed machine learning models. Therefore, our framework may be a viable approach to make de novo synergy predictions and can be used as a powerful starting point to more efficiently design resource-deprived combination screens.

Acknowledgements (Optional):