Machine Learning Maps Cancer Driver Genes Across the Human Protein Interaction Network

Cancer driver genes are notoriously hard to find. They make up less than five percent of all human genes, so any algorithm trained to spot them faces a brutal statistical imbalance: for every true driver, there are roughly twenty genes that are not. A study published in Genomics & Informatics describes the first phase of a tool called the Precision Medicine Gene Network Analyser, which tackles that imbalance head-on by fusing the topology of the human protein interaction network with a deliberately imbalance-aware ensemble of machine learning models.

The researchers, led by Rashmi Siddalingappa and Gugan Kothandan, built their positive class from the COSMIC Cancer Gene Census, extracting 742 candidate driver genes and retaining 699 after mapping them onto a filtered protein-protein interaction network from the STRING database. The negative class consisted of 15,050 network genes absent from the Cancer Gene Census, giving 15,749 labelled genes in total. The underlying network was derived from STRING version 12 by keeping only high-confidence interactions with combined scores of at least 700, then mapping Ensembl protein identifiers to official gene symbols, which produced a graph of 15,749 nodes connected by 456,300 edges.

From that graph the team extracted eight network topology features for every gene: degree centrality, betweenness centrality, PageRank, k-core number, clustering coefficient, closeness centrality, eigenvector centrality and triangle count. These eight numbers formed the complete input vector for every classifier. Degree centrality, which simply counts a gene’s direct interaction partners, turned out to be the single most informative feature, contributing 0.32 of predictive importance, followed by betweenness at 0.24 and PageRank at 0.19; together these three accounted for 75 percent of the model’s predictive power.

The methodological centrepiece is the Imbalance-Aware Network Integrator, or IANI, which corrects for class imbalance at three levels simultaneously. At the data level, synthetic minority oversampling generated new cancer-gene examples while Latin Hypercube Sampling-based undersampling trimmed the majority class in a way that preserved its global distributional structure. At the model level, a deep neural network trained with focal loss down-weighted easy examples and concentrated learning on hard-to-classify minority genes, while logistic regression, random forest and gradient boosting used class-weight adjustments. At the decision level, the ensemble’s classification threshold was tuned against the precision-recall curve to maximise the F1-score, landing at 0.466 rather than the default 0.5. The four classifiers were combined through weighted soft voting, with gradient boosting and the neural network each carrying 30 percent of the vote.

Evaluation was carefully insulated from leakage. The full dataset was split 80-20 before any resampling, leaving a held-out test set of 3,150 genes containing 140 cancer genes and 3,010 non-cancer genes at the natural 4.4 percent positive rate. On that untouched test set, the optimised ensemble lifted ROC-AUC from 0.84 to 0.96, precision from 0.78 to 0.90 and recall from 0.42 to 0.81, with an F1-score of 0.85. The recall improvement, a 93 percent gain over baseline, is the headline number: it shows that the multi-level imbalance correction recovered far more true drivers without sacrificing precision.

The authors are candid that their negative labels are not experimentally validated non-cancer genes. Genes absent from the Cancer Gene Census form an unlabelled background pool, making this formally a positive-unlabelled learning problem, and reported precision and recall represent lower bounds on true performance. Strikingly, of the 89 apparent false positives on the test set, 40 percent had peer-reviewed post-2020 literature evidence of cancer association, and 86 percent of those literature-supported genes now appear in IntOGen 2023 or OncoKB 2024. Examples include ARID1A, SETD2, PBRM1, CREBBP and ASXL1, all validated or recurrently mutated drivers that post-dated the January 2023 COSMIC snapshot used for training.

Independent validation strengthened the case. The trained model was applied, without any retraining, to 568 high-confidence cancer drivers from IntOGen that were absent from the training set, filtered for temporal independence and network membership. The ensemble achieved a ROC-AUC of 0.91 and an AUPRC of 0.76 on this external set, correctly identifying 412 of the 568 genes, a recall of 0.73. The missed genes showed significantly lower network degree than the recovered ones, mirroring the pattern among COSMIC false negatives and suggesting the limitation lies in the topology signal itself rather than the training labels. A degree-matched random baseline analysis confirmed the model was not merely recognising highly connected nodes: predicted cancer genes overlapped COSMIC at 56.4 percent versus 4.4 percent for degree-matched random sets.

Functional and clinical evidence pointed the same way. In DepMap CRISPR knockout screens across 1,086 cancer cell lines, predicted cancer genes showed substantially more negative essentiality scores than predicted non-cancer genes, with 68 percent essential in at least 20 percent of cell lines. In TCGA survival data spanning 33 cancer types and 10,967 patients, mutations in predicted cancer genes were associated with significantly worse overall survival, a median of 42.3 months versus 68.7 months, with a hazard ratio of 1.72. Separately, 78 of the 202 predicted cancer genes, 38.6 percent, were direct targets of FDA-approved or clinically investigated cancer therapies, compared with just 1.1 percent of predicted non-cancer genes.

Beyond classification, the study mapped the network’s architecture. Applying a hub threshold of mean outdegree plus two standard deviations, 118.7, identified 689 hub genes, of which 111, or 16.1 percent, were known cancer genes compared with 4.0 percent among non-hubs, a roughly fourfold enrichment. Cancer hubs also showed significantly higher betweenness centrality than non-cancer hubs, marking them as bridges between network modules. TP53 topped the list with 758 connections, followed by EGFR with 512 and AKT1 with 415, and their strongest interactions were 60 to 67 percent enriched for cancer genes, clustering in p53 signalling and cell cycle regulation pathways.

The authors frame this as Phase 1 of a three-phase programme, with drug-gene association mapping and patient-specific treatment prediction still to come. They also acknowledge real constraints: the STRING network is a static, incomplete average across tissue contexts, the pan-cancer model does not capture tumour-type heterogeneity, and topological centrality indicates importance rather than proving causative roles in oncogenesis, so experimental validation remains essential. Even so, the convergence of supervised classification, hub analysis, external driver sets, cell-line essentiality and patient survival data makes a coherent argument that cancer drivers leave a measurable fingerprint in the shape of the interaction network itself.

Subject of Research: Computational identification of cancer driver genes using protein interaction network topology and ensemble machine learning

Article Title: Precision Medicine Gene Network Analyser: part I—cancer driver gene identification through network topology and ensemble machine learning

Article References: Siddalingappa, R., Hussain, S., S., D., Dheerendra, P., Gornale, S., B. L., M., & Kothandan, G. (2026). Precision Medicine Gene Network Analyser: part I—cancer driver gene identification through network topology and ensemble machine learning. Genomics & Informatics, 24(1), Article 12. https://doi.org/10.1186/s44342-026-00074-7

Image Credits: AI Generated

DOI: 10.1186/s44342-026-00074-7

Keywords: cancer driver genes, machine learning, protein-protein interaction network, STRING, COSMIC Cancer Gene Census, class imbalance, hub genes, IntOGen, DepMap, TCGA, precision oncology, network topology