What are the best alternatives to ROC-AUC for validating transcriptomic/gene expression signatures?

Hello everyone,

I am currently developing a gene expression panel to predict patient response/stratification. For validation and performance analysis, I have been using the standard Area Under the Receiver Operating Characteristic curve (ROC-AUC) to measure the signature's ability to separate my target classes.

However, I am looking to expand my validation beyond just ROC-AUC. I want to include other robust, threshold-independent metrics that behave similarly (evaluating overall discriminative power) but might capture different nuances—especially regarding class imbalances or specific precision-recall dynamics.

Could the community suggest good alternatives that are well-accepted in bioinformatics for signature/classifier validation? Specifically:

AUC-PR (Precision-Recall AUC) / Average Precision (AP): I understand this is highly recommended for genomics when classes are imbalanced. Are there specific R packages or workflows preferred for this?

Matthews Correlation Coefficient (MCC): How well-regarded is this for publishing transcriptomic panels compared to AUC?

Other methods: Are there other threshold-free metrics (like C-index for survival-linked signatures, or specific permutation-based separation scores) that would look rigorous to reviewers?

Any advice on how to structure a multi-metric validation workflow for gene signatures would be greatly appreciated!

Thanks in advance.

roc auc

• 107 views • link

updated 1 week ago by SeqBench •

• written 1 month ago by Pumla •

1 answer

One correction first: MCC isn't threshold-free. It's computed from a confusion matrix, so it requires you to have already picked a cutoff. AUC-ROC and AUC-PR are the threshold-free pair; MCC, F1 and balanced accuracy all sit downstream of a threshold choice. Worth being precise about that in a manuscript, reviewers do notice.

I'd also gently push back on the framing. More metrics doesn't make a validation more rigorous, and six discrimination metrics that all agree tends to read as padding rather than thoroughness. AUC and AP measure much the same thing with different weighting of the imbalance -- reporting both is reasonable, reporting five is decoration.

What's actually missing from your list is calibration. AUC, AP and MCC all only care about ranking or classification; none of them tell you whether a predicted probability of 0.7 corresponds to 70% of those patients responding. For a signature framed as predicting patient response that matters quite a lot, and TRIPOD-style reporting expects it. A calibration curve plus Brier score gives you a genuinely different axis of performance instead of another discrimination number.

The other thing reviewers press on for clinical panels is incremental value: does the signature add anything over the clinical variables already available? That's a likelihood ratio test between nested models, or decision curve analysis if you want to make the clinical-utility argument. DCA tends to land better with reviewers than another AUC variant.

And more important than any of the metrics: where the numbers come from. An independent external cohort beats any amount of cross-validation. If you only have one cohort, nested CV with feature selection inside the inner loop -- selecting genes on the full dataset and then cross-validating is the single most common reason these panels fail to replicate, and it inflates every metric on your list simultaneously. Bootstrap CIs rather than bare point estimates.

C-index is the right choice for survival endpoints, but it's AUC generalised to censored data, so treat it as the survival-flavoured version rather than an additional independent metric.

• 0 views • link

written 1 week ago by SeqBench •