Machine Learning Analysis of Individual Tumor Lesions in Four Metastatic Colorectal Cancer Clinical Studies: Linking Tumor Heterogeneity to Overall Survival

Organ-specific tumor dynamics predict survival of patients with metastatic colorectal cancer

BACKGROUND

We aim to compare the prognostic value of organ-specific dynamics with the sum of the longest diameter (SLD) dynamics in patients with metastatic colorectal cancer (mCRC).

METHODS

All datasets are accessible in Project Data Sphere, an open-access platform. The tumor growth inhibition models developed based on organ-level SLD and SLD were used to estimate the organ-specific tumor growth rates (KGs) and SLD KG. The early tumor shrinkage (ETS) from baseline to the first measurement after treatment was also evaluated. The relationship between organ-specific dynamics, SLD dynamics, and survival outcomes (overall survival, OS; progression-free survival, PFS) was quantified using Kaplan-Meier analysis and Cox regression.

RESULTS

This study included 3687 patients from 6 phase III mCRC trials. The liver emerged as the most frequent metastatic site (2901, 78.7 %), with variable KGs across different organs in individual patients (liver 0.0243 > lung 0.0202 > lymph node 0.0127 > other 0.0118 [week-1]). Notably, the dynamics for different organs did not equally contribute to predicting survival outcomes. In liver metastasis cases, liver KG proved to be a superior prognostic indicator for OS and surpasses the predictive performance of SLD, (C-index, liver KG 0.610 vs SLD KG 0.606). A similar result can be found for PFS. Moreover, liver ETS also outperforms SLD ETS in predicting survival. Cox regression analysis confirmed liver KG is the most significant variable in survival prediction.

CONCLUSIONS

In mCRC patients with liver metastasis, liver dynamics is the primary prognostic indicator for both PFS and OS. In future drug development for mCRC, greater emphasis should be directed towards understanding the dynamics of liver metastasis development.

Realizing the promise of Project Optimus: Challenges and emerging opportunities for dose optimization in oncology drug development

Project Optimus is a US Food and Drug Administration Oncology Center of Excellence initiative aimed at reforming the dose selection and optimization paradigm in oncology drug development. This project seeks to bring together pharmaceutical companies, international regulatory agencies, academic institutions, patient advocates, and other stakeholders. Although there is much promise in this initiative, there are several challenges that need to be addressed, including multidimensionality of the dose optimization problem in oncology, the heterogeneity of cancer and patients, importance of evaluating long-term tolerability beyond dose-limiting toxicities, and the lack of reliable biomarkers for long-term efficacy. Through the lens of Totality of Evidence and with the mindset of model-informed drug development, we offer insights into dose optimization by building a quantitative knowledge base integrating diverse sources of data and leveraging quantitative modeling tools to build evidence for drug dosage considering exposure, disease biology, efficacy, toxicity, and patient factors. We believe that rational dose optimization can be achieved in oncology drug development, improving patient outcomes by maximizing therapeutic benefit while minimizing toxicity.

Machine Learning in Modeling Disease Trajectory and Treatment Outcomes: An Emerging Enabler for Model-Informed Precision Medicine

The increasing breadth and depth of resolution in biological and clinical data, including -omics and real-world data, requires advanced analytical techniques like artificial intelligence (AI) and machine learning (ML) to fully appreciate the impact of multi-dimensional population variability in intrinsic and extrinsic factors on disease progression and treatment outcomes. Integration of advanced data analytics in Quantitative Pharmacology is crucial for drug-disease knowledge management, enabling precise, efficient and inclusive drug development and utilization - an application we refer to as model-informed precision medicine. AI/ML enables characterization of the molecular and clinical sources of heterogeneity in disease trajectory, advancing end point qualification and biomarker discovery, and informing patient enrichment for proof-of-concept studies as well as trial designs for efficient evidence generation incorporating digital twins and virtual control arms. Explainable ML methods are valuable in elucidating predictors of efficacy and safety of pharmacological treatments, thereby informing response monitoring and risk mitigation strategies. In oncology, emerging opportunities exist for development of the next generation of disease models via ML-assisted joint longitudinal modeling of high-dimensional biomarker data such as circulating tumor DNA and radiomics profiles as predictors of survival outcomes. Finally, mining real-world data leveraging ML algorithms enables understanding of the impact of exclusion criteria on clinical outcomes, thereby informing rational design of appropriately inclusive clinical trials through data-driven broadening of eligibility criteria. Herein, we provide an overview of the aforementioned contexts of use of ML in drug-disease modeling based on examples across multiple therapeutic areas including neurology, rare diseases, autoimmune diseases, oncology and immuno-oncology.

Virtual patients, digital twins and causal disease models: paving the ground for in silico clinical trials

Computational models are being explored to simulate in silico the efficacy and safety of drug candidates and medical devices. Disease models that are based on patients' profiling data are being produced to represent interactomes of genes or proteins and to infer causality in the pathophysiology {AuQ: Edit OK?}, which makes it possible to mimic the impact of drugs on relevant targets. Virtual patients designed from medical records as well as digital twins were generated to simulate specific organs and to predict treatment efficacy at the individual patient level {AuQ: Edit OK?}. As the acceptance of digital evidence by regulators grows, predictive artificial intelligence (AI)-based models will support the design of confirmatory trials in humans and will accelerate the development of efficient drugs and medical devices.

Machine Learning in Metastatic Cancer Research: Potentials, Possibilities, and Prospects

Cancer has received extensive recognition for its high mortality rate, with metastatic cancer being the top cause of cancer-related deaths. Metastatic cancer involves the spread of the primary tumor to other body organs. As much as the early detection of cancer is essential, the timely detection of metastasis, the identification of biomarkers, and treatment choice are valuable for improving the quality of life for metastatic cancer patients. This study reviews the existing studies on classical machine learning (ML) and deep learning (DL) in metastatic cancer research. Since the majority of metastatic cancer research data are collected in the formats of PET/CT and MRI image data, deep learning techniques are heavily involved. However, its black-box nature and expensive computational cost are notable concerns. Furthermore, existing models could be overestimated for their generality due to the non-diverse population in clinical trial datasets. Therefore, research gaps are itemized; follow-up studies should be carried out on metastatic cancer using machine learning and deep learning tools with data in a symmetric manner.

Opportunities and Challenges of Disease Progression Modeling in Drug Development - An IQ Perspective

Disease progression modeling (DPM) represents an important model-informed drug development framework. The scientific communities support the use of DPM to accelerate and increase efficiency in drug development. This article summarizes International Consortium for Innovation & Quality (IQ) in Pharmaceutical Development mediated survey conducted across multiple biopharmaceutical companies on challenges and opportunities for DPM. Additionally, this summary highlights the viewpoints of IQ from the 2021 workshop hosted by the US Food and Drug Administration (FDA). Sixteen pharmaceutical companies participated in the IQ survey with 36 main questions. The types of questions included single/multiple choice, dichotomous, rank questions, and open-ended or free text. The key results show that DPM has different representation, it encompasses natural disease history, placebo response, standard of care as background therapy, and can even be interpreted as pharmacokinetic/pharmacodynamic modeling. The most common reasons for not implementing DPM as frequently seem to be difficulties in internal cross-functional alignment, lack of knowledge of disease/data, and time constraints. If successfully implemented, DPM can have an impact on dose selection, reduction of sample size, trial read-out support, patient selection/stratification, and supportive evidence for regulatory interactions. The key success factors and key challenges of disease progression models were highlighted in the survey and about 24 case studies across different therapeutic areas were submitted from various survey sponsors. Although DPM is still evolving, its current impact is limited but promising. The success of such models in the future will depend on collaboration, advanced analytics, availability of and access to relevant and adequate-quality data, collaborative regulatory guidance, and published examples of impact.

Artificial Intelligence Predictive Models of Response to Cytotoxic Chemotherapy Alone or Combined to Targeted Therapy for Metastatic Colorectal Cancer Patients: A Systematic Review and Meta-Analysis

Tailored treatments for metastatic colorectal cancer (mCRC) have not yet completely evolved due to the variety in response to drugs. Therefore, artificial intelligence has been recently used to develop prognostic and predictive models of treatment response (either activity/efficacy or toxicity) to aid in clinical decision making. In this systematic review, we have examined the ability of learning methods to predict response to chemotherapy alone or combined with targeted therapy in mCRC patients by targeting specific narrative publications in Medline up to April 2022 to identify appropriate original scientific articles. After the literature search, 26 original articles met inclusion and exclusion criteria and were included in the study. Our results show that all investigations conducted on this field have provided generally promising results in predicting the response to therapy or toxic side-effects. By a meta-analytic approach we found that the overall weighted means of the area under the receiver operating characteristic (ROC) curve (AUC) were 0.90, 95% C.I. 0.80-0.95 and 0.83, 95% C.I. 0.74-0.89 in training and validation sets, respectively, indicating a good classification performance in discriminating response vs. non-response. The calculation of overall HR indicates that learning models have strong ability to predict improved survival. Lastly, the delta-radiomics and the 74 gene signatures were able to discriminate response vs. non-response by correctly identifying up to 99% of mCRC patients who were responders and up to 100% of patients who were non-responders. Specifically, when we evaluated the predictive models with tests reaching 80% sensitivity (SE) and 90% specificity (SP), the delta radiomics showed an SE of 99% and an SP of 94% in the training set and an SE of 85% and SP of 92 in the test set, whereas for the 74 gene signatures the SE was 97.6% and the SP 100% in the training set.

Gene-Mutation-Based Algorithm for Prediction of Treatment Response in Colorectal Cancer Patients

Purpose: Despite the high mortality of metastatic colorectal cancer (mCRC), no new biomarker tools are available for predicting treatment response. We developed gene-mutation-based algorithms as a biomarker classifier to predict treatment response with better precision than the current predictive factors. Methods: Random forest machine learning (ML) was applied to identify the candidate algorithms using the MSK Cohort (n = 471) as a training set and validated in the TCGA Cohort (n = 221). Logistic regression, progression-free survival (PFS), and univariate/multivariate Cox proportional hazard analyses were performed and the performance of the candidate algorithms was compared with the established risk parameters. Results: A novel 7-Gene Algorithm based on mutation profiles of seven KRAS-associated genes was identified. The algorithm was able to distinguish non-progressed (responder) vs. progressed (non-responder) patients with AUC of 0.97 and had predictive power for PFS with a hazard ratio (HR) of 16.9 (p < 0.001) in the MSK cohort. The predictive power of this algorithm for PFS was more pronounced in mCRC (HR = 16.9, p < 0.001, n = 388). Similarly, in the TCGA validation cohort, the algorithm had AUC of 0.98 and a significant predictive power for PFS (p < 0.001). Conclusion: The novel 7-Gene Algorithm can be further developed as a biomarker model for prediction of treatment response in mCRC patients to improve personalized therapies.

Assessing the impact of organ-specific lesion dynamics on survival in patients with recurrent urothelial carcinoma treated with atezolizumab or chemotherapy

Background

Tumor dynamics typically rely on the sum of the longest diameters (SLD) of target lesions, and ignore heterogeneity in individual lesion dynamics located in different organs.

Patients and methods

Here we evaluated the benefit of analyzing lesion dynamics in different organs to predict survival in 900 patients with metastatic urothelial carcinoma treated with atezolizumab or chemotherapy (IMvigor211 trial).

Results

Lesion dynamics varied largely across organs, with lymph nodes and lung lesions showing on average a better response to both treatments than those located in the liver and locoregionally. A benefit of atezolizumab was observed on lung and liver lesion dynamics that was attributed to a longer duration of treatment effect as compared to chemotherapy (P value = 0.043 and 0.001, respectively). The impact of lesion dynamics on survival, assessed by a joint model, varied greatly across organs, irrespective of treatment. Liver and locoregional lesion dynamics had a large impact on survival, with an increase of 10 mm of the lesion size increasing the instantaneous risk of death by 12% and 10%, respectively. In comparison, lymph nodes and lung lesions had a lower impact, with a 10-mm increase in the lesion size increasing the instantaneous risk of death by 7% and 5%, respectively. Using our model, we could anticipate the benefit of atezolizumab over chemotherapy as early as 6 months before the end of the study, which is 3 months earlier than a similar model only relying on SLD.

Conclusion

We showed the interest of organ-level tumor follow-up to better understand and anticipate the treatment effect on survival.

•

Nine hundred metastatic urothelial carcinoma patients from the IMvigor211 phase III trial were treated with atezolizumab versus chemotherapy.

•

A total of 4489 organ-specific measurements were made: 1544 measurements in the lymph, 999 in the lung, 691 in the liver, and 559 locoregionally.

•

Longer treatment effect was observed in the lung (P value = 0.043) and liver (P = 0.001) lesions under atezolizumab compared to chemotherapy.

•

Those with a 10-mm growth of liver lesion had their instantaneous risk of death increased by 12%, compared to 5% in the lung.

•

Treatment effect on overall survival could be predicted based on early organ-specific tumor data 6 months after last patient inclusion.

Regulation of signal transduction pathways in colorectal cancer: implications for therapeutic resistance

Resistance to anti-cancer treatments is a critical and widespread health issue that has brought serious impacts on lives, the economy and public policies. Mounting research has suggested that a selected spectrum of patients with advanced colorectal cancer (CRC) tend to respond poorly to both chemotherapeutic and targeted therapeutic regimens. Drug resistance in tumours can occur in an intrinsic or acquired manner, rendering cancer cells insensitive to the treatment of anti-cancer therapies. Multiple factors have been associated with drug resistance. The most well-established factors are the emergence of cancer stem cell-like properties and overexpression of ABC transporters that mediate drug efflux. Besides, there is emerging evidence that signalling pathways that modulate cell survival and drug metabolism play major roles in the maintenance of multidrug resistance in CRC. This article reviews drug resistance in CRC as a result of alterations in the MAPK, PI3K/PKB, Wnt/β-catenin and Notch pathways.

Relevance of primary lesion location, tumour heterogeneity and genetic mutation demonstrated through tumour growth inhibition and overall survival modelling in metastatic colorectal cancer

AIMS

The aims of this work were to build a semi-mechanistic tumour growth inhibition (TGI) model for metastatic colorectal cancer (mCRC) patients receiving either cetuximab + chemotherapy or chemotherapy alone and to identify early predictors of overall survival (OS).

METHODS

A total of 1716 patients from 4 mCRC clinical studies were included in the analysis. The TGI model was built with 8973 tumour size measurements where the probability of drop-out was also included and modelled as a time-to-event variable using parametric survival models, as it was the case in the OS analysis. The effects of patient- and tumour-related covariates on model parameters were explored.

RESULTS

Chemotherapy and cetuximab effects were included in an additive form in the TGI model. Development of resistance was found to be faster for chemotherapy (drug effect halved at wk 8) compared to cetuximab (drug effect halved at wk 12). KRAS wild-type status and presenting a right-sided primary lesion were related to a 3.5-fold increase in cetuximab drug effect and a 4.7× larger cetuximab resistance, respectively. The early appearance of a new lesion (HR = 4.14), a large tumour size at baseline (HR = 1.62) and tumour heterogeneity (HR = 1.36) were the main predictors of OS.

CONCLUSIONS

Semi-mechanistic TGI and OS models have been developed in a large population of mCRC patients receiving chemotherapy in combination or not with cetuximab. Tumour-related predictors, including a machine learning derived-index of tumour heterogeneity, were linked to changes in drug effect, resistance to treatment or OS, contributing to the understanding of the variability in clinical response.

Fast screening of covariates in population models empowered by machine learning

One of the objectives of Pharmacometry (PMX) population modeling is the identification of significant and clinically relevant relationships between parameters and covariates. Here, we demonstrate how this complex selection task could benefit from supervised learning algorithms using importance scores. We compare various classical methods with three machine learning (ML) methods applied to NONMEM empirical Bayes estimates: random forest, neural networks (NNs), and support vector regression (SVR). The performance of the ML models is assessed using receiver operating characteristic (ROC) curves. The F1 score, which measures test accuracy, is used to compare ML and PMX approaches. Methods are applied to different scenarios of covariate influence based on simulated pharmacokinetics data. ML achieved similar or better F1 scores than stepwise covariate modeling (SCM) and conditional sampling for stepwise approach based on correlation tests (COSSAC). Correlations between covariates and the number of false covariates does not affect the performance of any method, but effect size has an impact. Methods are not equivalent with respect to computational speed; SCM is 30 and 100-times slower than NN and SVR, respectively. The results are validated in an additional scenario involving 100 covariates. Taken together, the results indicate that ML methods can greatly increase the efficiency of population covariate model building in the case of large datasets or complex models that require long run-times. This can provide fast initial covariate screening, which can be followed by more conventional PMX approaches to assess the clinical relevance of selected covariates and build the final model.

Application of Machine Learning in Translational Medicine: Current Status and Future Opportunities

The exponential increase in our ability to harness multi-dimensional biological and clinical data from experimental to real-world settings has transformed pharmaceutical research and development in recent years, with increasing applications of artificial intelligence (AI) and machine learning (ML). Patient-centered iterative forward and reverse translation is at the heart of precision medicine discovery and development across the continuum from target validation to optimization of pharmacotherapy. Integration of advanced analytics into the practice of Translational Medicine is now a fundamental enabler to fully exploit information contained in diverse sources of big data sets such as “omics” data, as illustrated by deep characterizations of the genome, transcriptome, proteome, metabolome, microbiome, and exposome. In this commentary, we provide an overview of ML applications in drug discovery and development, aligned with the three strategic pillars of Translational Medicine (target, patient, dose) and offer perspectives on their potential to transform the science and practice of the discipline. Opportunities for integrating ML approaches into the discipline of Pharmacometrics are discussed and will revolutionize the practice of model-informed drug discovery and development. Finally, we posit that joint efforts of Clinical Pharmacology, Bioinformatics, and Biomarker Technology experts are vital in cross-functional team settings to realize the promise of AI/ML-enabled Translational and Precision Medicine.

Spatiotemporal Heterogeneity across Metastases and Organ-Specific Response Informs Drug Efficacy and Patient Survival in Colorectal Cancer

The sum of target lesions is routinely used to evaluate patient objective responses to treatment in the RECIST criteria, but it fails to address response heterogeneity across metastases. This study argues that spatiotemporal heterogeneity across metastases and organ-specific response is informative for drug efficacy and patient survival. We analyzed the longitudinal data of 11,404 metastatic lesions in 2,802 colorectal cancer patients from five phase III clinical trials. Initially, a metric Gower distance was applied to quantify response heterogeneity across metastases. Next, the spatiotemporal response heterogeneity across anatomic sites, therapies, and KRAS mutation status was assessed and examined for its association with drug efficacy and long-term patient survival. The response of metastatic lesions broadly differed across anatomic sites and therapies. About 60% of patients had at least one lesion respond contrarily from total tumor size. High interlesion heterogeneity was associated with shorter progression-free survival and overall survival. Targeted therapies (bevacizumab or panitumumab) combined with standard chemotherapy reduced interlesion heterogeneity and elicited more favorable effects from liver lesions (P < 0.001) than chemotherapy alone. Moreover, the favorable responses in liver metastases (> 30% shrinkage) were associated with extended patient overall survival (P < 0.001), in contrast to lesions in the lungs and lymph nodes. Altogether, the spatiotemporal response heterogeneity across metastases informed drug efficacy and patient survival, which could improve the current methods for treatment evaluation and patient prognosis. SIGNIFICANCE: These findings support the modification of RECIST criteria to include individual lesion response to improve assessments of drug efficacy.

Comparison of Various Phase I Combination Therapy Designs in Oncology for Evaluation of Early Tumor Shrinkage Using Simulations

There is still a lack of efficient designs for identifying the dose response in oncology combination therapies in early clinical trials. The concentration response relationship can be identified using the early tumor shrinkage time course, which has been shown to be a good early response marker of clinical efficacy. The performance of various designs using an exposure–tumor growth inhibition model was explored using simulations. Different combination effects of new drug M and cetuximab (reference therapy) were explored first assuming no effect of M on cetuximab (to investigate the type I error (α)), and subsequently assuming additivity or synergy between cetuximab and M. One‐arm, two‐arm, and four‐arm designs were evaluated. In the one‐arm design, 60 patients received cetuximab + M. In the two‐arm design, 30 patients received cetuximab and 30 received cetuximab + M. In the four‐arm design, in addition to cetuximab and cetuximab + M as standard doses, combination arms with lower doses of cetuximab were evaluated (15 patients/arm). Model‐based predictions or “simulated observations” of early tumor shrinkage at week 8 (ETS8) were compared between the different arms. With the same number of individuals, the one‐arm design showed better statistical power than other designs but led to strong inflation of α in case of misestimated reference for ETS8 value. The two‐arm design protected against this misestimation and, with the same total number of subjects, would provide higher statistical power than a four‐arm design. However, a four‐arm design would be helpful for exploring more doses of cetuximab in combination with M to better understand the interaction.