Trecento : in silico network-driven identification of target combinations for combination therapy
Chua, Huey Eng
Date of Issue2015
School of Computer Engineering
Singapore-MIT Alliance Programme
In recent years, there has been a gradual paradigm shift in the drug industry towards combination therapy after the “one-target one-drug” (mono-therapy) approach fails to increase the number of successful drugs. Although, the use of multiple drugs in a therapy is potentially useful for addressing diseases (e.g., cancer) that implicate multiple genes and pathways, there is also a higher possibility of drug toxicity that is caused by various factors including the com-pounding of off-target effects of individual drugs. Careful selection of target combination is an important step towards reducing attrition rate of new drug combinations during preclinical and clinical phases. Although various techniques have been suggested for target selection, they generally apply to mono-therapies. Many of these approaches (e.g., afﬁnity matrix) cannot be applied directly to target combination identiﬁcation because they lack consideration of inter-action between targets that affects efﬁcacy and toxicity proﬁles of drug combinations hitting these targets. In addition, the few proposed approach for target combination identiﬁcation all suffer from certain key limitations. First, they assume that all targets are equally probable for a combination. However, targets inﬂuence effects of the combinations differently and appropriate choice of targets can potentially improve the target combinations. Second, they demand speciﬁc technical expertise from users (e.g., knowledge of what off-target effects to optimize) and a lack of the required expertise can undermine the effectiveness of the method. In this dissertation, we seek to address these limitations by proposing a framework called TRECENTO (In Silico NeTwork-dRiven IdEntiﬁCation of TargEt CombiNaTions for COmbination Therapy). TRECENTO uses a network-based approach for identifying target combinations in signaling networks. In particular, it addresses three aspects of target combination identiﬁcation, namely, target characterization, target prioritization and synergistic target combination identiﬁcation. For target characterization, we propose TENET, a support vector machine-based approach, to characterize known targets in signaling networks using network topological features. The insights gained from characterizing the targets are used subsequently for target prioritization in our proposed network-based approach, TAPESTRY. Finally, for synergistic target combination identiﬁcation, we propose a heuristics-based simulated annealing algorithm called STEROID. Speciﬁcally, STEROID uses heuristics based on target prioritization and the Loewe additivity theorem to identify synergistic target combinations that are optimized for off-target effects. We evaluated TRECENTO on ﬁve signaling networks that are implicated in diseases (e.g., cancer) or speciﬁc biological processes (e.g., glucose metabolism). Our experimental results reveal several key ﬁndings. First, known targets in different signaling networks are characterized by different sets of topological features. This is different from the observation reported by Hwang and colleagues for several protein-protein interaction networks. Second, the use of network dynamic features can improve target prioritization in certain networks. This ﬁnding highlights the importance of dynamic features and the need to explore them further in network-based analysis of signaling networks. Third, using appropriate heuristics can im-prove the quality of target combinations identiﬁed in terms of off-target effects when compared to state-of-the-art approaches. Note that TRECENTO optimizes off-target effects of the entire system instead of speciﬁc off-target effects. In summary, this thesis is an important ﬁrst step towards improving the quality of target combination selection, a key step in enabling target combination therapies.
DRNTU::Engineering::Computer science and engineering