Table 2 Implementation and technical challenges in drug combination optimization methods

DrugComboRanker

Relies only on cell line data for drug network construction

Lacks comprehensive preclinical and clinical validation;

Does not integrate certain biomedical knowledge like drug side effects;

Limited to gene expression and PPI data, missing other potentially valuable omics data types

Processes large datasets (6100 profiles from 4 cancer cell lines with 1309 drugs);

Requires matrix factorization computations;

Needs to analyze complex biological networks;

Must handle multiple databases simultaneously

Relies on access to CMAP database;

Requires drug interaction data from multiple sources (STITCH, BioGRID);

Needs updated PPI network information;

Limited by available cell line data

Needs experimental validation of predicted drug combinations;

Requires clinical validation beyond cell line studies;

Should validate predictions across different disease contexts;

Needs comparison with other prediction methods

AuDNNsynergy

Need for high quality multi-omics data;

Complex data preprocessing requirements;

Model training requires significant computational resources

Requires substantial computational resources for training due to:

Processing of high-dimensional multi-omics data;

Deep neural network architecture;

Multiple cell line-specific models

Requires comprehensive multi-omics profiles for cell lines;

Data quality and consistency across different omics; platforms

Need for standardized drug combination screening data

Experimental validation of predicted synergistic combinations;

Cross-validation across different cell lines;

Independent test sets for model evaluation;

Need for replication in clinical settings or animal models

MethylMix

Handling missing values;

Batch effect correction;

Computational demands for large datasets

Higher computational demands for sequencing data;

Parallel processing support to improve performance;

Mixture modeling is computationally demanding

Requires complete datasets from both normal and cancer samples;

Needs matched gene expression data;

Dependent on TCGA data availability

Statistical validation through consensus clustering;

Biological validation through pathway analysis;

Clinical validation of identified subtypes

CDA

Extracting biologically meaningful information from genome-wide expression analysis;

Standardizing bioinformatics approaches for signature extraction;

Integrating data from multiple public databases for a unified analysis

May have significant computational requirements

Reliance on public databases for gene expression data and pathway information

In vitro validations are necessary

DIGRE

Requires specialized software/programming;

Complex mathematical modeling;

Need for careful experimental design

Appears moderate given mathematical complexity

Requires extensive experimental data points;

Multiple drug combinations and concentrations needed

Statistical validation needed for some methods;

Reproducibility testing;

Comparison across multiple methods recommended

IUPUI_CCBB

Without considering other factors such as drug metabolism, transport, and target binding may impact predictive accuracy

May have moderate computational requirements.

Requires gene expression profile data following treatment with highly toxic compounds, which may be limited or difficult to obtain

Requires rigorous experimental validation and clinical trials

SynGen

Uses nuclear norm constraints;

Alternates between propensity-based sampling and prediction-driven screening;

Can handle datasets up to several thousand targets

Not specifically stated

Not specifically stated

Not specifically stated

DeepSynergy

Requires double representation of drug pairs to ensure order-invariant predictions;

Needs careful normalization strategy for heterogeneous input data

Not specifically stated

Available at https://www.bioinf.jku.at/software/DeepSynergy/

Nested cross-validation;

Separate validation for known vs novel drug combinations

PRODeepSyn

Integration of heterogeneous data sources;

Processing of high-dimensional sparse omics data

Not specifically stated

O’Neil dataset

Experimental validation of predictions needed;

Cross-validation required for assessing prediction ability on novel drug combinations

DGSSynADR

Integration of heterogeneous biological data;

Construction of complex graph networks;

End-to-end training requirements

Not specifically stated

Available at https://github.com/DHUDBlab/DGSSynADR

Validation through previous findings in case studies;

Ablation studies for component validation

LOBICO

Optimizes the algorithm for better interpretability and practicality

May require moderate computational requirements

Requires comprehensive genomic and transcriptomic data

May need rigorous experimental validation and clinical trials

MKL

Choosing appropriate kernel combination method;

Optimization of kernel weights;

Computational scalability;

Parameter selection

Higher than single kernel methods

Requires multiple feature representations;

Training data with labels needed;

Validation data for parameter tuning

Cross-validation for parameter selection;

Performance validation on test sets;

Kernel weight validation;

Model selection criteria

DrugComboExplorer

Complex network construction and analysis;

Integration of heterogeneous data types;

Parameter optimization for matrix factorization

Not specifically stated

Program available upon request;

Requires access to CMAP and protein interaction databases

Experimental validation of predicted drug combinations;

Clinical validation for practical application

OncosynergyX

Large feature set requires significant processing

Not specifically stated

Not specifically stated

Experimental validation of predicted synergistic combinations;

Clinical trials for promising drug pairs

RWR

Handling large-scale matrices;

Memory requirements

Iterative matrix computations;

Scale depends on network size

Available at https://github.com/alberto-valdeolivas/RWR-MH

Leave-one-out cross-validation;

Known disease-gene associations as ground truth

Mathematical Model Based on the EGFR Signaling Network

Parameter estimation accuracy;

Initial condition sensitivity;

Network topology completeness

Moderate (solvable via 4th order Runge-Kutta method);

Non-stiff system requires standard numerical methods

Rate constants from literature;

Initial conditions based on cell culture studies

Experimental validation of predicted combination effects;

Verification of synergy predictions;

Clinical correlation

Three-node Enzymatic Network Model

Solving nonlinear equations for steady states;

Ensuring system stability;

Computing dose-response relationships within physiological ranges

Required 100,000 parameter sets per network;

Multiple numerical solutions needed per parameter set

Model equations and parameters provided in paper;

No external data dependencies

Experimental validation needed;

Translation to larger networks unclear;

Biological relevance of simplified networks