LORIS (A LOgistic Regression-based Immunotherapy-response Score)

This is a repository of scripts for reproducing the paper titled "LORIS robustly predicts patient outcomes with immune checkpoint blockade therapy using common clinical, pathologic, and genomic features" by Chang et al. (Nature Cancer 2024)

<img src="./images/Fig1.jpg" width = "1000" alt="method" align=center /> Figure: Overview of this work. a. The data. b. Model building, training, and evaluation. c. LORIS facilitates the estimation of ICB efficacy for individual patients.

Briefly, in this work, Chang et al. developed a new clinical score called the LOgistic Regression-based Immunotherapy-response Score (LORIS) using a transparent and concise 6-feature logistic regression model. LORIS outperforms previous signatures in ICB response prediction and can identify responsive patients, even those with low tumor mutational burden or tumor PD-L1 expression. Importantly, LORIS consistently predicts both objective responses and short-term and long-term survival across most cancer types. Moreover, LORIS showcases a near-monotonic relationship with ICB response probability and patient survival, enabling more precise patient stratification across the board. As the method is accurate, interpretable, and only utilizes a few readily measurable features, it could help improve clinical decision-making practices in precision medicine to maximize patient benefit. Specifically, the LORIS can be calculated as follows:

Method 1: Pan-cancer LORIS

LORIS = 1 / (1 + e^(-S))

S = 0.0371 * min(TMB, 50) - 0.8775 * PSTH + 0.5382 * Albumin - 0.033 * min(NLR, 25) + 0.0049 * min(Age, 85) + CTCT - 2.0886

CTCT = - 0.3323 * Bladder - 0.3323 * Breast - 0.102 * Colorectal - 0.0079 * Endometrial + 0.55 * Esophageal + 0.2306 * Gastric
    + 0.0678 * Head&Neck - 0.1189 * Hepatobiliary - 0.0086 * Melanoma + 0.1255 * Mesothelioma + 0.0008 * NSCLC
    - 0.052 * Ovarian - 1.1169 * Pancreatic + 0.5451 * Renal + 0.0542 * Sarcoma - 0.0033 * SCLC

where PSTH is a patient’s systemic therapy history (if a patient received chemotherapy or targeted therapy before ICB therapy, its value is 1, otherwise its value is 0); CTCT is the cancer type calibration term (for a patient, the value of the cancer type he/she has been diagnosed is 1, all other cancer types are 0).

Other features: TMB, measured by panel sequencing (mut/Mb, typical value range 0-50); Albumin (g dL-1); NLR, ratio between absolute counts of blood neutrophils and lymphocytes.

Method 2: NSCLC-specific LORIS

LORIS = 1 / (1 + e^(-S))

S = 0.0353 * min(TMB, 50) + 0.0111*PDL1 - 0.375 * PSTH + 0.2924 * Albumin - 0.0103 * min(NLR, 25) - 1.5593

Features: PDL1, PD-L1 tumor proportion score (%, value range 0-100).

Scripts

The enclosed scripts are designed to faithfully reproduce all the results presented in the original paper.

1. Pan-cancer cohort cancer type composition pie plot

01.cohortCancerComposition_piePlot.Rmd

2. Characterization of patients in different ICB cohorts

02.ICBcohortsPatientCharacterization_table.Rmd

3. Correlation analysis of features that are measured on a continuous scale

03.PanCancer_continuousFeatureCorrelation_heatmap.py

4. Pan-cancer logistic regression: feature selection and feature importance

04.PanCancer_FeaturetImportance.py

5. Pan-cancer model comparison: training and evaluation of different models

05_1.PanCancer_20Models_HyperParams_Search.py

05_2.PanCancer_20Models_evaluation.py

Note: One may use batchSubmit.py and jobscript.sh to submit batch jobs to server for model hyperparameter search and/or evaluation.

05_3.TabNet_paramSearch_evaluation.py

05_4.PanCancer_20Models_evaluation_stat.py

05_5.PanCancer_LLRx_ParamCalculate.py

05_6.PanCancer_LLR6_RF16NBT_Performance_compare.py

05_7.PanCancer_LLR6_RF16NBT_AUC_compare.py

05_8.PanCancer_LLR6_RF6_TMB_thresholds_on_train.Rmd

6. Pan-cancer external validation of LLR6, RF6 and TMB

06_1.PanCancer_LLR6_RF6_TMB_multiMetric_compare.py

06_2.PanCancer_LLR6_RF6_TMB_ViolinPlot_compare.py

06_3.PanCancer_LLR6_vs_LLR6noCancerTerm_ROC_AUC.py

06_4.PanCancer_LLR6_vs_LLR5noPSTH_ROC_AUC.py

06_5.PanCancer_vs_nonNSCLC_LLR6_ROC_AUC.py

7. Pan-cancer prediction of patient survival following ICB or non-ICB treatments by LLR6; monotonic relationship between LORIS and patient ICB response and survival

07_1.PanCancer_LORIS_TMB_vs_resProb_curve.py

07_2.PanCancer_LORIS_TMB_survivalAnalysis_ICB.Rmd

07_3.PanCancer_LORIS_ROC_AUC_nonICB_vs_ICB_OS.py

07_4.PanCancer_LR5noTMB_LORIS_ROC_AUC_nonICB_vs_ICB_OS.py

07_5.PanCancer_LORIS_TMB_survivalAnalysis_nonICB.Rmd

07_6.PanCancer_LORIS_K-Mcurve_individualCancers_ICB.Rmd

8. NSCLC-specific model comparison: training and evaluation of different machine-learning models

08_1.NSCLC_20Models_HyperParams_Search.py

08_2.NSCLC_20Models_evaluation.py

Note: One may use batchSubmit.py and jobscript.sh to submit batch jobs to server for model hyperparameter search and/or evaluation.

08_3.NSCLC_20Models_evaluation_stat.py

08_4.NSCLC_LLRx_10k_ParamCalculate.py

9. NSCLC-specific external validation of LLR6

09_1.NSCLC_LLR6_PDL1_TMB_multiMetric_compare.py

09_2.NSCLC_LLR6_PDL1_TMB_ViolinPlot_compare.py

09_3.NSCLC_vs_PanCancer_LLR6_ROC_AUC.py

09_4.NSCLC_LLR6_vs_LLR2_ROC_AUC.py

09_5.NSCLC_LLR6_vs_LLR5noPSTH_ROC_AUC.py

10. Hazard ratio prediction of patient survival following ICB by NSCLC-specific LLR6

10.NSCLC_LORIS_PDL1_TMB_ForestPlot_individualDatasets.Rmd

11. Predictive power of NSCLC-specific LLR6 model in other cancer types

11.NSCLC_LLR6_on_otherCancers_ROC_AUC.py

12. Pan-cancer & NSCLC-specific formulas for LORIS calculation

12.Formula_LORIS.py

Citation

Chang, TG., Cao, Y., Sfreddo, H.J. et al. LORIS robustly predicts patient outcomes with immune checkpoint blockade therapy using common clinical, pathologic and genomic features. Nat Cancer 5, 1158–1175 (2024). https://doi.org/10.1038/s43018-024-00772-7

Contact

Dr. Eytan Ruppin: eytan.ruppin@nih.gov; Lab page.

Dr. Luc Morris: morrisl@mskcc.org; Lab page.

Dr. Tiangen Chang: tiangen.chang@nih.gov; Google scholar.

LORIS

Install / Use

README