A plagiarism story

Original

Cancer is a complex spatialized multi-scale process, starting from genetic mutations and potentially leading to metastasis. Moreover, it has multi-scale (from from genetic to environmental) causes. Therefore, it can be studied from various perspectives from the intracellular (molecular) to the population levels.
ML-ABM is a promising paradigm to model cancer development and discover new therapies (Wang and Deisboeck, 2008, 2013). [...]
Zhang et al. developed an ML-ABM of a brain tumor named Glioblastoma Multiforme (GBM) (Zhang et al., 2007, 2009a, 2011, 2009b). This model explicitly defines the relations between scales and uses different modeling approaches: ordinary differential equations (ODE) at the intracellular level, discrete rules typically found in ABM at the cellular level and partial differential equations (PDE) at the tissue level (see fig. 6).
Moreover, this model also relies on a multi-resolution approach: heterogenous clusters, i.e., composed of migrating and proliferating cells are simulated at a high resolution while homogenous clusters of dead cells are simulated at a lower resolution. [...] This model has been implemented on graphics processing units (GPU), leading to an efficient parallel simulator (Zhang et al., 2011). Then, this model illustrates two interesting features of the ML-ABM approach: the coupling of heterogeneous models and the dynamic adaptation of the level of detail of simulations to save computational resources. [...]
Lepagnot and Hutzler (2009) model the growth of avascular tumors to study the impact of PAI-1 molecules on metastasis. To deal with the problem complexity (a tumor may be composed of millions of cells) two levels are introduced: the cell and the tumor’s core levels (fig. 6). Indeed, such cancers are generally structured as a kernel of necrosed or quiescent cells surrounded by living tumor cells. As necrosed and quiescent cells are mostly inactive, tumor’s core is reified as a single upper-level agent, interacting with cells and PAI-1 molecules at its boundary. A more comprehensive analysis of this model can be found in Gil-Quijano et al. (2012).

Figure 6:

Original

There exists a wide range of disease gene prioritization methods that are based on the analysis of the topological properties of PPI networks. These methods commonly rely on the observation that the products of genes that are associated with similar diseases have a higher likelihood of physically interacting [11]. It is important to note here the distinction between genes and their products. Genome-wide association studies focus on identifying genes that are associated with a disease of interest. Network-based prioritization aims to aid this effort by inferring functional associations between genes based on the interactions among their products, i.e., proteins. For this reason, any reference to interactions between genes in this paper refers to the interactions between their products.
Existing methods for network-based disease gene prioritization can be classified into two main categories; (i) localized methods, i.e., methods based on direct interactions and shortest paths between known disease genes and candidate genes [3,9,16], (ii) global methods, i.e., methods that model the information flow in the cell to assess the proximity and connectivity between known disease genes and candidate genes. Several studies show that global approaches, such as random walk and network propagation, clearly outperform local approaches [13,14,17]. For this reason, we focus on global methods in this paper.
For a given disease of interest D, the input to the disease gene prioritization problem consists of two sets of genes, the seed set S and the candidate set C. The seed set S specifies prior knowledge on the disease, i.e., it is the set of genes known to be asso- ciated with D and diseases similar to D. Each gene v ∈ S is also associated with a simi- larity score s(v, D), indicating the known degree of association between v and D. The similarity score for gene v is computed as the maximum phenotypic similarity between D and any other disease associated with v, based on clinical description of diseases (a detailed discussion on computation of phenotypic similarity scores can be found in the Methods section below). The candidate set C specifies the genes, one or more of which are potentially associated with disease D (e.g., these genes might lie within a linkage interval that is identified by association studies). The overall objective of network based disease prioritization is to use a human PPI network G = (V , ε, w), to compute a score a(v, D) for each gene v ∈ C, such that a(v, D) represents the likelihood of v to be associated with D.
The PPI network G = (V,ε,w) consists of a set of gene products V and a set of undirected interactions E between these gene products, where uv ∈ E represents an interaction between u ∈ V and v ∈ V. Since PPI data might be obtained from various resources, interactions are often assigned confidence scores indicating their reliability. In other words, the network is also associated with a function w : E → (0, 1], such that w(uv) indicates the reliability of interaction uv ∈ E. Finally, the set of interacting part- ners of a gene product v∈V is defined as N(v) ={u∈V:uv∈ε }and the total reliability of known interactions of v is defined as W(v) = u∈N(v) w(uv) which we refer as weighted degree throughout this paper. Global prioritization methods use this network information to compute a by propagating s over G. Candidate proteins are then ranked according to a and novel genes that are potentially associated with the disease of interest are identified based on this ranking.
etc.