Addala, V. | Search bioRxiv

bioRxiv2025

Explainable machine learning identifies features and thresholds predictive of immunotherapy response

BackgroundImmunotherapy has improved patient survival for multiple cancer types, including melanoma. While a variety of molecular features have been linked to response to immune checkpoint inhibitors (ICI) treatment, clinically established biomarkers, such as tumour mutation burden (TMB) and PD-L1 expression, have shown limitations in accurately categorising responders versus non-responders. Due to the complex nature of ICI response, which includes cancer intrinsic and extrinsic features within the tumour microenvironment (TME), using a single biomarker to predict response is insufficient, necessitating the need to identify accurate clinical and multi-omic molecular predictors. MethodsWe integrate clinical, DNA and RNA sequencing data from four datasets, comprising 138 melanoma patients treated with ICI, to develop machine learning models for predicting ICI response. The performance of each model was evaluated using an independent dataset of patients with cutaneous melanoma (n=53). Interactions between trained models and features were rationalised using the explainability method SHAP. ResultsThe most optimal model was the multi-omic random forest model, with AUC-ROC of 0.78 when predicting response in the independent test dataset. Using SHAP, we predicted thresholds for mutational signatures, neoantigen load, immune cell-type abundance and immune receptor LAG3 expression. The relationship between these influential features and their SHAP scores revealed numerical thresholds constituting good and poor patient response. ConclusionsThis approach highlights patient response to ICI is influenced by both cancer intrinsic and extrinsic features, as well as identifies candidate biomarkers that could inform the use of ICI and potentially assist in the selection of combination therapy in melanoma treatment.

bioinformatics↗

bioRxiv2022

Generalising uncertainty improves accuracy and safety of deep learning analytics applied to oncology

Trust and transparency are critical for deploying deep learning (DL) models into the clinic. DL application poses generalisation obstacles since training/development datasets often have different data distributions to clinical/production datasets that can lead to incorrect predictions with underestimated uncertainty. To investigate this pitfall, we benchmarked one pointwise and three approximate Bayesian DL models used to predict cancer of unknown primary with three independent RNA-seq datasets covering 10,968 samples across 57 primary cancer types. Our results highlight simple and scalable Bayesian DL significantly improves the generalisation of uncertainty estimation (e.g., p-value = 0.0013 for calibration). Moreover, we demonstrate Bayesian DL substantially improves accuracy under data distributional shifts when utilising uncertainty thresholding by designing a prototypical metric that evaluates the expected (accuracy) loss when deploying models from development to production, which we call the Area between Development and Production curve (ADP). In summary, Bayesian DL is a hopeful avenue of research for generalising uncertainty, which improves performance, transparency, and therefore safety of DL models for deployment in real-world.

bioinformatics↗