The Clinical Gap PD-1 inhibitors are FDA-approved for lung adenocarcinoma (LUAD), yet the objective response rate is only 15-20%, with most patients developing acquired resistance and experiencing tumor progression. Identifying true responders before treatment would prevent unnecessary toxicity and cost.
Limitations of Current Biomarkers PD-L1 expression and tumor mutational burden (TMB) are the most widely used predictive biomarkers but have significant limitations: PD-L1 has non-standardized cutoffs and technical challenges, while TMB counts mutations without considering which specific mutations drive immunotherapy response.
The sBiosNet Approach This study developed a semi-supervised biological sparse neural network (sBiosNet) that integrates genomic mutation profiles and copy number variation (CNV) data with biological pathway information from the Reactome database to predict both response and survival under anti-PD-1 therapy in LUAD.
Best Performance sBiosNet achieved AUROC of 0.888 and AUPR of 0.919 for predicting responders vs. non-responders on the validation cohort, and AUROC 0.853 on an independent external cohort - outperforming random forest and SVM baselines.
Biological Sparsification Standard neural networks learn from any correlation in the data, including spurious ones. sBiosNet is structurally constrained using the Reactome biological pathway database - neurons correspond to known biological pathways, forcing the model to learn only biologically plausible relationships between mutations and treatment outcomes.
Semi-Supervised Learning LUAD patients with available response data are scarce. sBiosNet uses semi-supervised learning to leverage both labeled responder/non-responder data from LUAD patients and unlabeled data from non-LUAD patients treated with PD-1 inhibitors, generating pseudo-labels for unlabeled cases through the MixMatch algorithm.
Transfer Learning The model is pre-trained on non-LUAD cancer patients who received PD-1 inhibitors, then fine-tuned on LUAD data. This transfer approach allows learning of general immune checkpoint response biology before specializing to LUAD, improving performance when LUAD training data is limited.
Multi-Omics Integration Genomic mutations are encoded as binary mutation matrices and CNV profiles are encoded as amplification and deletion matrices. Only genes meeting minimum frequency thresholds were included (mutations greater than 1%, CNV variation greater than 5%), reducing noise while preserving biologically relevant signals.
Rizvi Cohort (Training) 185 LUAD patients from Memorial Sloan Kettering Cancer Center with genomic mutation profiles, CNV profiles, and anti-PD-1 therapy outcomes formed the primary training set, supplemented by 54 non-LUAD patients as unlabeled pre-training data.
Miao Cohort (Validation) 47 LUAD patients from Dana-Farber Cancer Institute, with whole-exome sequencing and best response to anti-PD-1 therapy assessed by RECIST v1.1 criteria, served as the internal validation cohort to assess generalization across institutions.
Frigola Cohort (External) 25 LUAD patients from an independent third institution with molecular profiles and anti-PD-1/PD-L1 treatment outcomes provided external validation to test performance in a truly independent population.
TCGA Cohort (Negative Control) 501 LUAD patients from TCGA with genomic profiles and survival data but no anti-PD-1 therapy were used as negative controls. Confirming that sBiosNet survival predictions do not apply to untreated patients validates that the model is specifically capturing immunotherapy-related biology.
Identified Key Genes DeepLIFT algorithm-based interpretability analysis identified specific genes whose mutations and CNV status most strongly influence sBiosNet's predictions, including TP53, FGF3, FGFR4, and EGFR, which affect LUAD response to PD-1 inhibitors through their roles in regulating biological pathways.
Pathway Mechanisms The relevant pathways identified include immune regulation, cell cycle control, and growth factor signaling. Mutations in these pathways can either sensitize tumors to immune attack or promote immune evasion - explaining why individual gene mutations matter beyond the overall mutation count captured by TMB.
Survival Stratification Low-risk LUAD patients identified by sBiosNet achieved significantly longer overall survival and progression-free survival under anti-PD-1 therapy compared to high-risk patients - confirming that the model predicts not just response but the survival consequence of that response.
Negative Control Validation When sBiosNet was applied to the TCGA untreated LUAD cohort, survival stratification was not significant, confirming that the model specifically captures biology relevant to anti-PD-1 response rather than general LUAD prognosis.
Ablation Evidence Ablation experiments systematically removed each component (biological sparsification, multi-omics integration, transfer learning, semi-supervised learning) to measure individual contributions. All four components significantly improved performance, validating the design rationale.
Beyond TMB TMB counts the total number of mutations but ignores which mutations matter for immune recognition. sBiosNet focuses on specific mutations and their pathway consequences, capturing the qualitative immunological relevance of the mutation profile that TMB misses.
Data Efficiency By using semi-supervised learning and transfer learning, sBiosNet achieves strong performance even with the small LUAD training datasets currently available (less than 200 patients), addressing a critical constraint in LUAD immunotherapy response prediction.
Interpretability The DeepLIFT-based attribution approach allows clinicians to understand which specific mutations drive a particular patient's predicted response or resistance - making sBiosNet a transparent, interpretable model rather than a black box.
Small Cohort Sizes Even with semi-supervised learning, the training cohorts are small (less than 250 LUAD patients total), which limits confidence in the model's generalizability across the full spectrum of LUAD clinical presentations and treatment settings.
Need for Prospective Validation All current validation is retrospective. Prospective clinical validation using sBiosNet to guide treatment decisions - and measuring whether this guidance improves outcomes compared to standard biomarker-based selection - is the essential next step.
Tumor Heterogeneity Challenge The model uses bulk sequencing data that represents an average across all tumor cells. Spatial heterogeneity within the tumor - different mutation profiles in different regions - could limit prediction accuracy in heterogeneous tumors.
Expanding the Multi-Omics Feature Set Adding transcriptomic data (RNA expression), epigenomic features (methylation), and immune cell infiltration data from tumor microenvironment analysis could improve prediction accuracy and provide additional mechanistic insights into why specific patients respond.