Table 1 Publications relevant to machine learning in immunotherapy response prediction
From: Informing immunotherapy with multi-omics driven machine learning
Task | ML Model | ML-based biomarker | Cancer | Patients | Therapy | Validation method | Performance | Input | Output | Ref |
|---|---|---|---|---|---|---|---|---|---|---|
Predict response | RF *, CNN | Yes | NSCLC | 915 | Anti-PD-(L)1 | 5-fold cross-validation | AUC (0.96–0.97) | 55 SNV locations | Response prediction (DCB, PFS, OS) | |
Predict response | SVM-RFE *, LASSO regularized LR | Yes | Metastatic BLCA | 272 | Anti-PD-L1 | 10-fold cross-validation | AUC (0.93) | TMB related genes | Responder vs. Non-responder and selected genes | |
Predict response | Multi-task linear regression using elastic net regularization | No | SKCM, STAD, BLCA, GBM | 432 | Anti-PD-(L)1 | Hold-out | AUC (0.79–0.84) | RNA-based features | Responder vs. Non-responder | |
Predict PFS | linear SVM * | Yes | Metastatic gastrointestinal cancer | 96 | Anti-PD-(L)1 | 13-fold cross-validation | AUC (0.74) | RNA of 395 genes | DCB vs. non-DCB | |
Predict response | SVM and XGBoost | No | Pan-cancer | Not mentioned | ICI | Hold-out | Accuracy (0.88) | RNA of 2387 genes | Responder vs. Non-responder | |
Predict response | SVM-RFE *, RF | Yes | SKCM | 212 | Anti-PD-1 | 10-fold cross-validation | AUC (0.71–0.87) | RNA + SNV + clinical features | Response prediction | |
Predict response | LR *, NN | Yes | Esophageal adenocarcinoma | 76 | ICI | Hold-out | AUC (0.88–1.00) | RNA | Responder vs. Non-responder and selected genes | |
Predict response | A joint NMF-based model * | Yes | Pan-cancer (12 cancer types) | 764 | Anti-PD-1, anti-PD-L1, anti-PD-L2, anti-CTLA4 | 5-fold cross-validation | AUC (0.74) | RNA | Responder vs. Non-responder | |
Predict response | LASSO regression *,SVM | Yes | NSCLC | 122 | Anti-PD-(L)1 | Hold-out | Significant hazard ratio differences | RNA | Responder vs. Non-responder | |
Predict response | LASSO regression * | Yes | NSCLC, UC, RCC | 366 | Anti-PD-L1 | 5-fold cross-validation | AUC (up to 0.62) | RNA | Responder vs. Non-responder selected gene features | |
Predict response | KNN, Linear SVM, RBF-SVM, GP, RF, DT, NN, AdaBoost, NB, quadratic classifier | No | BCC | 11 | Anti-PD-1 | 5-fold cross-validation | Accuracy (0.61–0.97 from different models) | Top 2,000 highly variable genes of CD8 T cell scRNA-seq data | Responder vs. Non-responder | |
Predict response | NN * | Yes | SKCM, BCC | 43 | Anti-PD-1 | LOOCV | Accuracy (up to 1.00) | scRNA-seq data of CD8 + T cell | Responder vs. Non-responder | |
Predict response | LR | No | GEA | Not mentioned | Anti-PD-1 along with radiation therapy | 10-fold cross-validation | Accuracy (up to 1.00) | Expression of selected genes from PMBC | Responder vs. Non-responder | |
Predict response | RF | No | NSCLC | 213 | Anti-PD-1 | 5-fold cross-validation | AUC (0.76–0.83) | Circulating miRNA + clinical information | Responder vs. Non-responder | |
Predict response and identify response related cfmiR biomarkers | RF * | Yes | Metastatic melanoma | 47 | ICI | Not mentioned | Not mentioned | 162 differentially expressed cfmiRs | Responder vs. Non-responder and selected cfmiRs | |
Predict response | LASSO regression | No | NSCLC | 78 | Anti-PD-(L)1 | 10-fold cross-validation | AUC (0.80) | Differentially methylated CpG sites | Responder vs. Non-responder | |
Predict response | LASSO regularized LR | No | Metastatic melanoma | 65 | ICI | 10-fold cross-validation | AUC (0.96) | 5000 most variable methylated CpG sites | Responder vs. Non-responder | |
Predict response | NN | No | HNSCC | 37 + simulated patients | Anti–PD-1 | 10-fold cross-validation | AUC (0.61–0.90) | Clinical features | Responder vs. Non-responder | |
Predict response | RF *,SVM | Yes | Colorectal cancer | 25 (mice) | Anti-mouse CTLA4, anti-mouse PD-L1 | LOOCV | Not directly showed | Spectra features from Raman spectroscopy | Responder vs. Non-responder and feature contributions | |
Predict response | MIL + DeepTCR | No | Not mentioned | 43 | ICI | Monte Carlo cross-validation | AUC (0.86) | TCR sequencing data + MHC sequencing data | Responder vs. Non-responder | |
Predict response | RF | No | Pan-cancer (16 cancer types) | 1,479 | ICI | 5-fold cross-validation | AUC (up to 0.85) | Genomic features based on DNA variants, RNA, demographic and clinical data | Responder vs. Non-responder | |
Predict response | SVM, NB, RF, KNN, AdaBoost, boosted LR | No | RCC, UC, SKCM, GBM, BCC | 955 | Anti-PD-(L)1, anti-CTLA4, anti-PD-(L)1 plus anti-CTLA-4 combination | 5-fold cross-validation | AUC (0.62–0.81) | Stemness features based on RNA | Responder vs. Non-responder | |
Predict response | XGBoost | No | Metastatic NSCLC | 239 | ICI | 10-fold cross-validation | AUC (0.72–0.74) | 25 variables based on blood immune cell signatures and clinical data | DCB vs. non-DCB | |
TME analysis and response prediction | LR | No | ccRCC | 172 | Anti-PD-(L)1, anti-CTLA4 | Hold-out | AUC (up to 0.93) | RNA of selected genes | Responder vs. Non-responder | |
Predict response | L2 regularized LR * | Yes | Melanoma, gastric cancer, bladder cancer | 729 | ICI | LOOCV, Monte Carlo cross-validation | AUC (0.69–0.79 in different datasets) | Network-based biomarkers + gene-based biomarkers + TME-based biomarkers | Response (Responder vs. Non-responder) and OS prediction | |
Predict response | CNN *, Attention-based multiple-instance learning | Yes | NSCLC | 345 | Anti-PD-(L)1 | 10-fold cross-validation | AUC (up to 0.80) | Radiology, pathology, genomic alternation, TMB | Risk score | |
Predict CAR T cell phenotype for immunotherapy response | NN | No | Not mentioned | NA | CAR T therapy | 10-fold cross-validation | R squared | Array of signaling motifs of a CAR costimulatory domain + initial CAR T cell count | Quantitively phenotype prediction (cytotoxicity and stemness) from a CAR motif combination |
