Piyush Chaubey
Assistant Professor, Electrical Engineering, Government Engineering College, Bharatpur, Rajasthan, India
Suganya S
Information Technology, Easwari Engineering College, India
Dr Karthika Padmavathy
Professor, Department of Pathology, Specialization: Pathology, Sri Lalithambigai Medical College and Hospital, Dr M.G.R Educational and Research Institute, Chennai, India
Dr. Radhika Gupta
Professor, Incharge, Sharda University, Knowledge Park, Greater Noida, India
Dr Shalini Kapoor
Professor, Department of Periodontology, SGT Dental College Hospital and Research Centre, SGT University, India
Dr Abhishek Anand
Assistant Professor, Department of Pharmacy Practice, Teerthanker Mahaveer College of Pharmacy, TMU University, Moradabad, India
Tumour-immune interactions during immunotherapy constitute a complex, dynamic biological system characterised by pronounced inter-patient heterogeneity and sparse clinical sampling. Precision oncology is largely focused on accurately predicting the effects of therapy on immune activation, which requires integrative modelling frameworks that balance data-driven inference with mechanistic interpretability. Here, we suggest an adapted machine learning-based biomathematical model to describe the system tumour-immune response over short periods, based on pre-treatment and on-treatment melanoma tumour biopsies taken prior to and following immune checkpoint blockade. Bulk RNA-sequencing data were used to quantify changes in immune effector activity induced by therapy, measured by the cytolytic score. Regularised ElasticNet regression was used to predict changes in immune activation from paired transcriptomic differences, with patient-level cross-validation to ensure generalisability. The ElasticNet model with Yeo-Johnson target transformation achieved high out-of-fold predictive performance (R2 ~ 0.73; RMSE ~ 33.4; MAE ~ 16.5), with strong agreement between observed and predicted immune activation changes. There was a cross-validation fold stability analysis of features that showed strong immune-based transcriptional drivers, such as genes linked to the natural killer cell and cytotoxic T-cell functions. Machine learning-derived transcriptomic immune activation scores were then superimposed onto patient-specific immune activation parameters within a simplified biomathematical tumour-immune ordinary differential equation model to provide a mechanistic understanding of therapy-induced immune stimulation without requiring long-term trajectory simulation. The derived parameters showed high inter-patient heterogeneity in immune responsiveness, as observed in biological variability. In general, the paper shows that the combination of interpretable machine learning and biomathematical modelling provides a powerful, biologically grounded approach to predicting changes in immune state and characterising the response to immunotherapy in cancer systems.
