Computational models of liver fibrosis progression for hepatitis C virus chronic infection

Abstract

Background: Chronic infection with hepatitis C virus (HCV) is a risk factor for liver diseases such as fibrosis, cirrhosis and hepatocellular carcinoma. HCV genetic heterogeneity was hypothesized to be associated with severity of liver disease. However, no reliable viral markers predicting disease severity have been identified. Here, we report the utility of sequences from 3 HCV 1b genomic regions, Core, NS3 and NS5b, to identify viral genetic markers associated with fast and slow rate of fibrosis progression (RFP) among patients with and without liver transplantation (n = 42).

Methods: A correlation-based feature selection (CFS) method was used to detect and identify RFP-relevant viral markers. Machine-learning techniques, linear projection (LP) and Bayesian Networks (BN), were used to assess and identify associations between the HCV sequences and RFP.

Results: Both clustering of HCV sequences in LP graphs using physicochemical properties of nucleotides and BN analysis using polymorphic sites showed similarities among HCV variants sampled from patients with a similar RFP, while distinct HCV genetic properties were found associated with fast or slow RFP. Several RFP-relevant HCV sites were identified. Computational models parameterized using the identified sites accurately associated HCV strains with RFP in 70/30 split cross-validation (90-95% accuracy) and in validation tests (85-90% accuracy). Validation tests of the models constructed for patients with or without liver transplantation suggest that the RFP-relevant genetic markers identified in the HCV Core, NS3 and NS5b genomic regions may be useful for the prediction of RFP regardless of transplant status of patients.

Conclusions: The apparent strong genetic association to RFP suggests that HCV genetic heterogeneity has a quantifiable effect on severity of liver disease, thus presenting opportunity for developing genetic assays for measuring virulence of HCV strains in clinical and public health settings.

Figure 1

2D linear projection (LP) graphs. Base vectors of projection represent RFP-relevant features. Sites are identified by their positions in the HCV genome and physicochemical properties are shown in parenthesis as X1-X5. See Table 1 for detail. LP graphs of HCV 1b isolates (n = 42) sampled from TOH and IC patients based on A) 12-feature and B) 9-feature projections are shown. To the right of each LP graph is the same graph except for the condition that data points were jittered [41] to highlight membership and size of clusters.

Figure 2

RFP-specificity of LP models. A) 2D graph of the LP model shown in Fig 1. HCV strains distributed into 4 clusters, from left to right: cluster 1 (fast RFP-IC, n = 1), cluster 2 (fast RFP-IC, n = 2; fast RFP-TOH, n = 4), cluster 3 (slow RFP-IC, n = 13; slow RFP-TOH, n = 12, and fast RFP-IC, n = 1; fast RFP-TOH, n = 3) and cluster 4 (fast RFP-IC, n = 3; fast RFP-TOH, n = 3). The graph below, shows mapping of computed probability potentials in LP model defining three RFP-class spaces of HCV strains (fast-RFP in blue, slow-RFP in red), where color density of areas are directly proportional to the probability of association to the respective RFP-class. Plots, where x-axis represents predicted probabilities and y-axis denotes observed RFP-class of HCV strains, show classification performances in validation tests of the B) LP-IC and C) LP-TOH model.

Figure 3

BNC models of RFP-relevant HCV sites. Nodes in the graph represent 25 nt sites (Table 1) and arcs between them represent relationships. Numbering of nodes in BNC denotes genomic position in Con1 as reference and colour represent genomic region. Node representing RFP is coloured in red. Models learned from HCV sequence profiles sampled from A) non-transplanted patients and from B) transplanted patients are shown.