Machine-learning and mechanistic modeling of metastatic breast cancer after neoadjuvant treatment.

Clinical trials involving systemic neoadjuvant treatments in breast cancer aim to shrink tumors before surgery while simultaneously allowing for controlled evaluation of biomarkers, toxicity, and suppression of distant (occult) metastatic disease. Yet neoadjuvant clinical trials are rarely preceded by preclinical testing involving neoadjuvant treatment, surgery, and post-surgery monitoring of the disease. Here we used a mouse model of spontaneous metastasis occurring after surgical removal of orthotopically implanted primary tumors to develop a predictive mathematical model of neoadjuvant treatment response to sunitinib, a receptor tyrosine kinase inhibitor (RTKI). Treatment outcomes were used to validate a novel mathematical kinetics-pharmacodynamics model predictive of perioperative disease progression. Longitudinal measurements of presurgical primary tumor size and postsurgical metastatic burden were compiled using 128 mice receiving variable neoadjuvant treatment doses and schedules (released publicly at https://zenodo.org/records/10607753). A non-linear mixed-effects modeling approach quantified inter-animal variabilities in metastatic dynamics and survival, and machine-learning algorithms were applied to investigate the significance of several biomarkers at resection as predictors of individual kinetics. Biomarkers included circulating tumor- and immune-based cells (circulating tumor cells and myeloid-derived suppressor cells) as well as immunohistochemical tumor proteins (CD31 and Ki67). Our computational simulations show that neoadjuvant RTKI treatment inhibits primary tumor growth but has little efficacy in preventing (micro)-metastatic disease progression after surgery and treatment cessation. Machine learning algorithms that included support vector machines, random forests, and artificial neural networks, confirmed a lack of definitive biomarkers, which shows the value of preclinical modeling studies to identify potential failures that should be avoided clinically.

Practical identifiability analysis of a mechanistic model for the time to distant metastatic relapse and its application to renal cell carcinoma.

Distant metastasis-free survival (DMFS) curves are widely used in oncology. They are classically analyzed using the Kaplan-Meier estimator or agnostic statistical models from survival analysis. Here we report on a method to extract more information from DMFS curves using a mathematical model of primary tumor growth and metastatic dissemination. The model depends on two parameters, α and µ, respectively quantifying tumor growth and dissemination. We assumed these to be lognormally distributed in a patient population. We propose a method for identification of the parameters of these distributions based on least-squares minimization between the data and the simulated survival curve. We studied the practical identifiability of these parameters and found that including the percentage of patients with metastasis at diagnosis was critical to ensure robust estimation. We also studied the impact and identifiability of covariates and their coefficients in α and µ, either categorical or continuous, including various functional forms for the latter (threshold, linear or a combination of both). We found that both the functional form and the coefficients could be determined from DMFS curves. We then applied our model to a clinical dataset of metastatic relapse from kidney cancer with individual data of 105 patients. We show that the model was able to describe the data and illustrate our method to disentangle the impact of three covariates on DMFS: a categorical one (Führman grade) and two continuous ones (gene expressions of the macrophage mannose receptor 1 (MMR) and the G Protein-Coupled Receptor Class C Group 5 Member A (GPRC5a) gene). We found that all had an influence in metastasis dissemination (µ), but not on growth (α).

Macro-scale models for fluid flow in tumour tissues: impact of microstructure properties.

Understanding the dynamics underlying fluid transport in tumour tissues is of fundamental importance to assess processes of drug delivery. Here, we analyse the impact of the tumour microscopic properties on the macroscopic dynamics of vascular and interstitial fluid flow. More precisely, we investigate the impact of the capillary wall permeability and the hydraulic conductivity of the interstitium on the macroscopic model arising from formal asymptotic 2-scale techniques. The homogenization technique allows us to derive two macroscale tissue models of fluid flow that take into account the microscopic structure of the vessels and the interstitial tissue. Different regimes were derived according to the magnitude of the vessel wall permeability and the interstitial hydraulic conductivity. Importantly, we provide an analysis of the properties of the models and show the link between them. Numerical simulations were eventually performed to test the models and to investigate the impact of the microstructure on the fluid transport. Future applications of our models include their calibration with real imaging data to investigate the impact of the tumour microenvironment on drug delivery.

Experimental and computational modeling for signature and biomarker discovery of renal cell carcinoma progression.

BACKGROUND: Renal Cell Carcinoma (RCC) is difficult to treat with 5-year survival rate of 10% in metastatic patients. Main reasons of therapy failure are lack of validated biomarkers and scarce knowledge of the biological processes occurring during RCC progression. Thus, the investigation of mechanisms regulating RCC progression is fundamental to improve RCC therapy. METHODS: In order to identify molecular markers and gene processes involved in the steps of RCC progression, we generated several cell lines of higher aggressiveness by serially passaging mouse renal cancer RENCA cells in mice and, concomitantly, performed functional genomics analysis of the cells. Multiple cell lines depicting the major steps of tumor progression (including primary tumor growth, survival in the blood circulation and metastatic spread) were generated and analyzed by large-scale transcriptome, genome and methylome analyses. Furthermore, we performed clinical correlations of our datasets. Finally we conducted a computational analysis for predicting the time to relapse based on our molecular data. RESULTS: Through in vivo passaging, RENCA cells showed increased aggressiveness by reducing mice survival, enhancing primary tumor growth and lung metastases formation. In addition, transcriptome and methylome analyses showed distinct clustering of the cell lines without genomic variation. Distinct signatures of tumor aggressiveness were revealed and validated in different patient cohorts. In particular, we identified SAA2 and CFB as soluble prognostic and predictive biomarkers of the therapeutic response. Machine learning and mathematical modeling confirmed the importance of CFB and SAA2 together, which had the highest impact on distant metastasis-free survival. From these data sets, a computational model predicting tumor progression and relapse was developed and validated. These results are of great translational significance. CONCLUSION: A combination of experimental and mathematical modeling was able to generate meaningful data for the prediction of the clinical evolution of RCC.

Population modeling of tumor growth curves and the reduced Gompertz model improve prediction of the age of experimental tumors.

Tumor growth curves are classically modeled by means of ordinary differential equations. In analyzing the Gompertz model several studies have reported a striking correlation between the two parameters of the model, which could be used to reduce the dimensionality and improve predictive power. We analyzed tumor growth kinetics within the statistical framework of nonlinear mixed-effects (population approach). This allowed the simultaneous modeling of tumor dynamics and inter-animal variability. Experimental data comprised three animal models of breast and lung cancers, with 833 measurements in 94 animals. Candidate models of tumor growth included the exponential, logistic and Gompertz models. The exponential and-more notably-logistic models failed to describe the experimental data whereas the Gompertz model generated very good fits. The previously reported population-level correlation between the Gompertz parameters was further confirmed in our analysis (R2 > 0.92 in all groups). Combining this structural correlation with rigorous population parameter estimation, we propose a reduced Gompertz function consisting of a single individual parameter (and one population parameter). Leveraging the population approach using Bayesian inference, we estimated times of tumor initiation using three late measurement timepoints. The reduced Gompertz model was found to exhibit the best results, with drastic improvements when using Bayesian inference as compared to likelihood maximization alone, for both accuracy and precision. Specifically, mean accuracy (prediction error) was 12.2% versus 78% and mean precision (width of the 95% prediction interval) was 15.6 days versus 210 days, for the breast cancer cell line. These results demonstrate the superior predictive power of the reduced Gompertz model, especially when combined with Bayesian estimation. They offer possible clinical perspectives for personalized prediction of the age of a tumor from limited data at diagnosis. The code and data used in our analysis are publicly available at https://github.com/cristinavaghi/plumky.

Deciphering the response and resistance to immune-checkpoint inhibitors in lung cancer with artificial intelligence-based analysis: when PIONeeR meets QUANTIC.

This project aims to generate dense longitudinal data in lung cancer patients undergoing anti-PD1/PDL1 therapy. Mathematical modelling with mechanistic learning algorithms will help decipher the mechanisms underlying the response or resistance to immunotherapy. A better understanding of these mechanisms should help identifying actionable items to increase the efficacy of immune-checkpoint inhibitors.

Metronomic reloaded: Theoretical models bringing chemotherapy into the era of precision medicine.

Oncology has benefited from an increasingly growing number of groundbreaking innovations over the last decade. Targeted therapies, biotherapies, and the most recent immunotherapies all contribute to increase the number of therapeutic options for cancer patients. Consequently, substantial improvements in clinical outcomes for some disease with dismal prognosis such as lung carcinoma or melanoma have been achieved. Of note, the latest innovations in targeted therapies or biotherapies do not preclude the use of standard cytotoxic agents, mostly used in combination. Importantly, and despite the rise of bioguided (a.k.a. precision) medicine, the administration of chemotherapeutic agents still relies on the maximum tolerated drug (MTD) paradigm, a concept inherited from theories conceptualized nearly half a century ago. Alternative dosing schedules such as metronomic regimens, based upon the repeated and regular administration of low doses of chemotherapeutic drugs, and adaptive therapy (i.e. modulating the dose and frequency of cytotoxics administration to control disease progression rather than eradicate it at all cost) have emerged as possible strategies to improve response rates while reducing toxicities. The recent changes in paradigm in the way we theorize cancer biology and evolution, metastatic spreading and tumor ecology, alongside the recent advances in the field of immunotherapy, have considerably strengthened the interest for these alternative approaches. This paper aims at reviewing the recent evolutions in the field of theoretical biology of cancer and computational oncology, with a focus on the consequences these changes have on the way we administer chemotherapy. Here, we advocate for the development of model-guided strategies to refine doses and schedules of chemotherapy administration in order to achieve precision medicine in oncology.

Computational Modelling of Metastasis Development in Renal Cell Carcinoma.

The biology of the metastatic colonization process remains a poorly understood phenomenon. To improve our knowledge of its dynamics, we conducted a modelling study based on multi-modal data from an orthotopic murine experimental system of metastatic renal cell carcinoma. The standard theory of metastatic colonization usually assumes that secondary tumours, once established at a distant site, grow independently from each other and from the primary tumour. Using a mathematical model that translates this assumption into equations, we challenged this theory against our data that included: 1) dynamics of primary tumour cells in the kidney and metastatic cells in the lungs, retrieved by green fluorescent protein tracking, and 2) magnetic resonance images (MRI) informing on the number and size of macroscopic lesions. Critically, when calibrated on the growth of the primary tumour and total metastatic burden, the predicted theoretical size distributions were not in agreement with the MRI observations. Moreover, tumour expansion only based on proliferation was not able to explain the volume increase of the metastatic lesions. These findings strongly suggested rejection of the standard theory, demonstrating that the time development of the size distribution of metastases could not be explained by independent growth of metastatic foci. This led us to investigate the effect of spatial interactions between merging metastatic tumours on the dynamics of the global metastatic burden. We derived a mathematical model of spatial tumour growth, confronted it with experimental data of single metastatic tumour growth, and used it to provide insights on the dynamics of multiple tumours growing in close vicinity. Together, our results have implications for theories of the metastatic process and suggest that global dynamics of metastasis development is dependent on spatial interactions between metastatic lesions.

Classical mathematical models for description and prediction of experimental tumor growth.

Despite internal complexity, tumor growth kinetics follow relatively simple laws that can be expressed as mathematical models. To explore this further, quantitative analysis of the most classical of these were performed. The models were assessed against data from two in vivo experimental systems: an ectopic syngeneic tumor (Lewis lung carcinoma) and an orthotopically xenografted human breast carcinoma. The goals were threefold: 1) to determine a statistical model for description of the measurement error, 2) to establish the descriptive power of each model, using several goodness-of-fit metrics and a study of parametric identifiability, and 3) to assess the models' ability to forecast future tumor growth. The models included in the study comprised the exponential, exponential-linear, power law, Gompertz, logistic, generalized logistic, von Bertalanffy and a model with dynamic carrying capacity. For the breast data, the dynamics were best captured by the Gompertz and exponential-linear models. The latter also exhibited the highest predictive power, with excellent prediction scores (≥80%) extending out as far as 12 days in the future. For the lung data, the Gompertz and power law models provided the most parsimonious and parametrically identifiable description. However, not one of the models was able to achieve a substantial prediction rate (≥70%) beyond the next day data point. In this context, adjunction of a priori information on the parameter distribution led to considerable improvement. For instance, forecast success rates went from 14.9% to 62.7% when using the power law model to predict the full future tumor growth curves, using just three data points. These results not only have important implications for biological theories of tumor growth and the use of mathematical modeling in preclinical anti-cancer drug investigations, but also may assist in defining how mathematical models could serve as potential prognostic tools in the clinic.

Computational markers for personalized prediction of outcomes in non-small cell lung cancer patients with brain metastases.

Intracranial progression after curative treatment of early-stage non-small cell lung cancer (NSCLC) occurs from 10 to 50% and is difficult to manage, given the heterogeneity of clinical presentations and the variability of treatments available. The objective of this study was to develop a mechanistic model of intracranial progression to predict survival following a first brain metastasis (BM) event occurring at a time [Formula: see text]. Data included early-stage NSCLC patients treated with a curative intent who had a BM as the first and single relapse site (N = 31). We propose a mechanistic mathematical model able to derive computational markers from primary tumor and BM data at [Formula: see text] and estimate the amount and sizes of (visible and invisible) BMs, as well as their future behavior. These two key computational markers are [Formula: see text], the proliferation rate of a single tumor cell; and [Formula: see text], the per day, per cell, probability to metastasize. The predictive value of these individual computational biomarkers was evaluated. The model was able to correctly describe the number and size of metastases at [Formula: see text] for 20 patients. Parameters [Formula: see text] and [Formula: see text] were significantly associated with overall survival (OS) (HR 1.65 (1.07-2.53) p = 0.0029 and HR 1.95 (1.31-2.91) p = 0.0109, respectively). Adding the computational markers to the clinical ones significantly improved the predictive value of OS (c-index increased from 0.585 (95% CI 0.569-0.602) to 0.713 (95% CI 0.700-0.726), p < 0.0001). We demonstrated that our model was applicable to brain oligoprogressive patients in NSCLC and that the resulting computational markers had predictive potential. This may help lung cancer physicians to guide and personalize the management of NSCLC patients with intracranial oligoprogression.

A senescence-mimicking (senomimetic) VEGFR TKI side-effect primes tumor immune responses via IFN/STING signaling.

Tyrosine kinase inhibitors (TKIs) that block the vascular endothelial growth factor receptors (VEGFRs) disrupt tumor angiogenesis but also have many unexpected side-effects that impact tumor cells directly. This includes the induction of molecular markers associated with senescence, a form of cellular aging that typically involves growth arrest. We have shown that VEGFR TKIs can hijack these aging programs by transiently inducting senescence-markers (SMs) in tumor cells to activate senescence-associated secretory programs that fuel drug resistance. Here we show that these same senescence-mimicking ('senomimetic') VEGFR TKI effects drive an enhanced immunogenic signaling that, in turn, can alter tumor response to immunotherapy. Using a live-cell sorting method to detect beta-galactosidase, a commonly used SM, we found that subpopulations of SM-expressing (SM+) tumor cells have heightened interferon (IFN) signaling and increased expression of IFN-stimulated genes (ISGs). These ISG increases were under the control of the STimulator of INterferon Gene (STING) signaling pathway, which we found could be directly activated by several VEGFR TKIs. TKI-induced SM+ cells could stimulate or suppress CD8 T-cell activation depending on host:tumor cell contact while tumors grown from SM+ cells were more sensitive to PD-L1 inhibition in vivo, suggesting that offsetting immune-suppressive functions of SM+ cells can improve TKI efficacy overall. Our findings may explain why some (but not all) VEGFR TKIs improve outcomes when combined with immunotherapy and suggest that exploiting senomimetic drug side-effects may help identify TKIs that uniquely 'prime' tumors for enhanced sensitivity to PD-L1 targeted agents.

A Senescence-Mimicking (Senomimetic) VEGFR TKI Side Effect Primes Tumor Immune Responses via IFN/STING Signaling.

Tyrosine kinase inhibitors (TKIs) that block the vascular endothelial growth factor receptors (VEGFRs) not only disrupt tumor angiogenesis but also have many unexpected side effects that impact tumor cells directly. This includes the induction of molecular markers associated with senescence, a form of cellular aging that typically involves growth arrest. We have shown that VEGFR TKIs can hijack these aging programs by transiently inducting senescence markers (SMs) in tumor cells to activate senescence-associated secretory programs that fuel drug resistance. Here we show that these same senescence-mimicking ("senomimetic") VEGFR TKI effects drive an enhanced immunogenic signaling that, in turn, can alter tumor response to immunotherapy. By using a live cell sorting method to detect ß-galactosidase, a commonly used SM, we found that subpopulations of SM-expressing (SM+) tumor cells have heightened IFN signaling and increased expression of IFN-stimulated genes (ISGs). These ISGs increase under the control of the STimulator of the INterferon Gene (STING) signaling pathway, which we found could be directly activated by several VEGFR TKIs. TKI-induced SM+ cells could stimulate or suppress CD8 T-cell activation depending on host-tumor cell contact while tumors grown from SM+ cells were more sensitive to PDL1 inhibition in vivo, suggesting that offsetting immune-suppressive functions of SM+ cells can improve TKI efficacy overall. Our findings may explain why some (but not all) VEGFR TKIs improve outcomes when combined with immunotherapy and suggest that exploiting senomimetic drug side effects may help identify TKIs that uniquely "prime" tumors for enhanced sensitivity to PDL1-targeted agents.

Predicting Survival in Patients with Advanced NSCLC Treated with Atezolizumab Using Pre- and on-Treatment Prognostic Biomarkers.

Existing survival prediction models rely only on baseline or tumor kinetics data and lack machine learning integration. We introduce a novel kinetics-machine learning (kML) model that integrates baseline markers, tumor kinetics, and four on-treatment simple blood markers (albumin, C-reactive protein, lactate dehydrogenase, and neutrophils). Developed for immune-checkpoint inhibition (ICI) in non-small cell lung cancer on three phase II trials (533 patients), kML was validated on the two arms of a phase III trial (ICI and chemotherapy, 377 and 354 patients). It outperformed the current state-of-the-art for individual predictions with a test set C-index of 0.790, 12-months survival accuracy of 78.7% and hazard ratio of 25.2 (95% CI: 10.4-61.3, P < 0.0001) to identify long-term survivors. Critically, kML predicted the success of the phase III trial using only 25 weeks of on-study data (predicted HR = 0.814 (0.64-0.994) vs. final study HR = 0.778 (0.65-0.931)). Modeling on-treatment blood markers combined with predictive machine learning constitutes a valuable approach to support personalized medicine and drug development. The code is publicly available at https://gitlab.inria.fr/benzekry/nlml_onco.

Maximum tolerated dose versus metronomic scheduling in the treatment of metastatic cancers.

Although optimal control theory has been used for the theoretical study of anti-cancerous drugs scheduling optimization, with the aim of reducing the primary tumor volume, the effect on metastases is often ignored. Here, we use a previously published model for metastatic development to define an optimal control problem at the scale of the entire organism of the patient. In silico study of the impact of different scheduling strategies for anti-angiogenic and cytotoxic agents (either in monotherapy or in combination) is performed to compare a low-dose, continuous, metronomic administration scheme with a more classical maximum tolerated dose schedule. Simulation results reveal differences between primary tumor reduction and control of metastases but overall suggest use of the metronomic protocol.

Mechanistic modeling of metastatic relapse in early breast cancer to investigate the biological impact of prognostic biomarkers.

BACKGROUND AND OBJECTIVE: Estimating the risk of metastatic relapse is a major challenge to decide adjuvant treatment options in early-stage breast cancer (eBC). To date, distant metastasis-free survival (DMFS) analysis mainly relies on classical, agnostic, statistical models (e.g., Cox regression). Instead, we propose here to derive mechanistic models of DMFS. METHODS: The present series consisted of eBC patients who did not receive adjuvant systemic therapy from three datasets, composed respectively of 692 (Bergonié Institute), 591 (Paoli-Calmettes Institute, IPC), and 163 (Public Hospital Marseille, AP-HM) patients with routine clinical annotations. The last dataset also contained expression of three non-routine biomarkers. Our mechanistic model of DMFS relies on two mathematical parameters that represent growth (α) and dissemination (µ). We identified their population distributions using mixed-effects modeling. Critically, we propose a novel variable selection procedure allowing to: (i) identify the association of biological parameters with either α, µ or both, and (ii) generate an optimal candidate model for DMFS prediction. RESULTS: We found that Ki67 and Thymidine Kinase-1 were associated with α, and nodal status and Plasminogen Activator Inhibitor-1 with µ. The predictive performances of the model were excellent in calibration but moderate in discrimination, with c-indices of 0.72 (95% CI [0.48, 0.95], AP-HM), 0.63 ([0.44, 0.83], Bergonié) and 0.60 (95% CI [0.54, 0.80], IPC). CONCLUSIONS: Overall, we demonstrate that our novel method combining mechanistic and advanced statistical modeling is able to unravel the biological roles of clinicopathological parameters from DMFS data.

Reuniting philosophy and science to advance cancer research.

Cancers rely on multiple, heterogeneous processes at different scales, pertaining to many biomedical fields. Therefore, understanding cancer is necessarily an interdisciplinary task that requires placing specialised experimental and clinical research into a broader conceptual, theoretical, and methodological framework. Without such a framework, oncology will collect piecemeal results, with scant dialogue between the different scientific communities studying cancer. We argue that one important way forward in service of a more successful dialogue is through greater integration of applied sciences (experimental and clinical) with conceptual and theoretical approaches, informed by philosophical methods. By way of illustration, we explore six central themes: (i) the role of mutations in cancer; (ii) the clonal evolution of cancer cells; (iii) the relationship between cancer and multicellularity; (iv) the tumour microenvironment; (v) the immune system; and (vi) stem cells. In each case, we examine open questions in the scientific literature through a philosophical methodology and show the benefit of such a synergy for the scientific and medical understanding of cancer.

Tumor growth monitoring in breast cancer xenografts: A good technique for a strong ethic.

PURPOSE: Although recent regulations improved conditions of laboratory animals, their use remains essential in cancer research to determine treatment efficacy. In most cases, such experiments are performed on xenografted animals for which tumor volume is mostly estimated from caliper measurements. However, many formulas have been employed for this estimation and no standardization is available yet. METHODS: Using previous animal studies, we compared all formulas used by the scientific community in 2019. Data were collected from 93 mice orthotopically xenografted with human breast cancer cells. All formulas were evaluated and ranked based on correlation and lower mean relative error. They were then used in a Gompertz quantitative model of tumor growth. RESULTS: Seven formulas for tumor volume estimation were identified and a statistically significant difference was observed among them (ANOVA test, p < 2.10-16), with the ellipsoid formula (1/6 π × L × W × (L + W)/2) being the most accurate (mean relative error = 0.272 ± 0.201). This was confirmed by the mathematical modeling analysis where this formula resulted in the smallest estimated residual variability. Interestingly, such result was no longer valid for tumors over 1968 ± 425 mg, for which a cubic formula (L x W x H) should be preferred. MAIN FINDINGS: When considering that tumor volume remains under 1500mm3, to limit animal stress, improve tumor growth monitoring and go toward mathematic models, the following formula 1/6 π × L × W x (L + W)/2 should be preferred.

Development and Validation of a Prediction Model of Overall Survival in High-Risk Neuroblastoma Using Mechanistic Modeling of Metastasis.

Prognosis of high-risk neuroblastoma (HRNB) remains poor despite multimodal therapies. Better prediction of survival could help to refine patient stratification and better tailor treatments. We established a mechanistic model of metastasis in HRNB relying on two processes: growth and dissemination relying on two patient-specific parameters: the dissemination rate µ and the minimal visible lesion size Svis. This model was calibrated using diagnosis values of primary tumor size, lactate dehydrogenase circulating levels, and the meta-iodobenzylguanidine International Society for Paediatric Oncology European (SIOPEN) score from nuclear imaging, using data from 49 metastatic patients. It was able to describe the data of total tumor mass (lactate dehydrogenase, R2 > 0.99) and number of visible metastases (SIOPEN, R2 = 0.96). A prediction model of overall survival (OS) was then developed using Cox regression. Clinical variables alone were not able to generate a model with sufficient OS prognosis ability (P = .507). The parameter µ was found to be independent of the clinical variables and positively associated with OS (P = .0739 in multivariable analysis). Critically, addition of this computational biomarker significantly improved prediction of OS with a concordance index increasing from 0.675 (95% CI, 0.663 to 0.688) to 0.733 (95% CI, 0.722 to 0.744, P < .0001), resulting in significant OS prognosis ability (P = .0422).

Machine Learning for Prediction of Immunotherapy Efficacy in Non-Small Cell Lung Cancer from Simple Clinical and Biological Data.

BACKGROUND: Immune checkpoint inhibitors (ICIs) are now a therapeutic standard in advanced non-small cell lung cancer (NSCLC), but strong predictive markers for ICIs efficacy are still lacking. We evaluated machine learning models built on simple clinical and biological data to individually predict response to ICIs. METHODS: Patients with metastatic NSCLC who received ICI in second line or later were included. We collected clinical and hematological data and studied the association of this data with disease control rate (DCR), progression free survival (PFS) and overall survival (OS). Multiple machine learning (ML) algorithms were assessed for their ability to predict response. RESULTS: Overall, 298 patients were enrolled. The overall response rate and DCR were 15.3% and 53%, respectively. Median PFS and OS were 3.3 and 11.4 months, respectively. In multivariable analysis, DCR was significantly associated with performance status (PS) and hemoglobin level (OR 0.58, p < 0.0001; OR 1.8, p < 0.001). These variables were also associated with PFS and OS and ranked top in random forest-based feature importance. Neutrophil-to-lymphocyte ratio was also associated with DCR, PFS and OS. The best ML algorithm was a random forest. It could predict DCR with satisfactory efficacy based on these three variables. Ten-fold cross-validated performances were: accuracy 0.68 ± 0.04, sensitivity 0.58 ± 0.08; specificity 0.78 ± 0.06; positive predictive value 0.70 ± 0.08; negative predictive value 0.68 ± 0.06; AUC 0.74 ± 0.03. CONCLUSION: Combination of simple clinical and biological data could accurately predict disease control rate at the individual level.