Table 7.

Overall advantages and limitations of driver prioritisation approaches.

Driver Prioritisation Approach	Advantages	Limitations
Network-Based Driver Prioritisation Using External Reference Networks	•Utilises patient-specific changes in gene expression to evaluate the effect of a mutation in a patient-centric manner	•Requires genomic and transcriptomic information •Dependent on external reference network •Requires a cohort of patients for making comparisons which may need to be batch-corrected •Network-based approaches in general are susceptible to centrality bias
DawnRank	•Uses differential expression information quantitatively	•Requires paired tumour/normal expression data
OncoImpact	•Has measures to combat centrality bias	•Uses differential expression information qualitatively •Requires reference healthy expression data
Hit'nDrive	•Utilises integer linear programming •Requires only tumour expression data	•Requires additional licenced software (CPLEX) •Coded in proprietary, licenced language (MATLAB) •Uses differential expression information qualitatively
SCS		•Uses differential expression information qualitatively •Requires reference healthy expression data
PRODIGY	•Uses differential expression information quantitatively •Incorporates additional pathway information •Has measures to combat centrality bias	•Requires reference healthy expression data
PersonaDrive	•Requires only tumour expression data •Incorporates information from other similar samples •Incorporates additional pathway information	•Uses differential expression information qualitatively
Network-Based Driver Prioritisation Using De-Novo Networks	•Utilises patient-specific changes in gene expression to evaluate the effect of a mutation in a patient-centric manner •Does not rely on external reference networks	•Requires genomic and transcriptomic information •Requires a cohort of patients for making comparisons which may need to be batch-corrected •Network-based approaches in general are susceptible to centrality bias
PNC	•Utilises integer linear programming	•Requires additional licenced software (Gurobi) •Coded in proprietary, licenced language (MATLAB) •Requires paired tumour/normal expression data
pDriver	•Also considers miRNA drivers	•Requires miRNA expression data
PDGPCS	•Incorporates additional pathway information	•Coded in proprietary, licenced language (MATLAB) •Requires paired tumour/normal expression data
Machine Learning-Based Driver Prioritisation	•Only requires genomic information •Can theoretically be expanded to include more features •Once the model is trained, truly requires only a single patient	•Models are reliant on accuracy of “known driver gene” databases •“Black-box” approaches without mechanistic interpretability
iCAGES		•Training model requires true positive and true negative drivers •Only uses ANNOVAR annotations as training features
sysSVM2	•Only true positive drivers are required for training •Expanded list of training features
driverR	•Expanded list of training features •Cancer-type specific	•Training model requires true positive and true negative drivers
IMCDriver		•Does not utilise any external similarity features •Similarity based on co-mutation