Machine-learning model predicts anti-TNF nonresponse in RA patients

A machine-learning model that uses clinical profiles and genetic information has shown promise in predicting which rheumatoid arthritis patients respond to anti–tumor necrosis factor drugs in a patient population of European descent.

The model can “help up to 40% of European-descent anti–tumor necrosis factor [TNF] nonresponders avoid ineffective treatments” when compared with the usual “trial-and-error practice,” according to the authors led by Yuanfang Guan, PhD, of the department of computational medicine and bioinformatics at the University of Michigan, Ann Arbor.

The ability to accurately predict rheumatoid arthritis patients’ response to treatments would provide valuable information for optimal drug selection and would help potential nonresponders avoid drug expenses and side effects, such as an increased risk of infections, Dr. Guan and coauthors noted in Arthritis & Rheumatology.

The investigators used a modeling technique called Gaussian process regression (GPR) to predict anti-TNF drug responses. “GPR is designed to predict the unknown dependent variable for any given independent variables based on known but noisy observations of the dependent and independent variables,” they explained.

The model they used won first place in the Dialogue on Reverse Engineering Assessment and Methods: Rheumatoid Arthritis Responder Challenge, which used a crowd-based competition framework to develop a validated molecular predictor of anti-TNF response in RA.

The model was developed and cross-validated using 1,892 patients randomly selected from a training data set of 2,706 individuals of European ancestry compiled from 13 patient cohorts. All patients met 1987 American College of Rheumatology criteria for RA or were diagnosed by a board-certified rheumatologist. In addition, patients were required to have at least moderate disease activity at baseline, based on a 28-joint Disease Activity Score (DAS28) greater than 3.2.

The research team also evaluated the model using an independent dataset of 680 patients from the CERTAIN (Comparative Effectiveness Registry to study Therapies for Arthritis and Inflammatory Conditions) study.

The model combined demographic, clinical, and genetic markers to predict patients’ changes in DAS28 24 months after their baseline assessment, and identify nonresponders to anti-TNF treatments, the authors explained.

“Specifically, the [model] predicts the changes in [DAS28] of patients who have taken 12 months of anti-TNF treatments, and also classifies the patients’ responses based on the EULAR response metric,” they wrote.

Results showed that, in cross-validation tests, the model predicted changes in DAS28 with a correlation coefficient of 0.406, correctly classifying responses of 78% of subjects, with an area under the receiver operating characteristic curve (AUROC) of about 0.66.

In the independent test, the method achieved a Pearson correlation coefficient of 0.393 in predicting the change in DAS28.

Genetic SNP biomarkers provided a small additional contribution to the prediction on top of the clinical models, the authors noted.

“Compared to traditional trial-and-error practice, our model can help up to 40% of European-descent anti-TNF nonresponders avoid ineffective treatments. The model performance is even comparable to some published models utilizing additional biomarker data, whose AUROC ranges from 55% to 74% over various testing sets,” they wrote.

The GPR model has practical advantages in clinical application, unlike many sophisticated machine-learning algorithms, according to the authors. For example, GPR is a well-studied statistical model, its similarity-modeling approach is intuitive, and its results are easy to interpret.

“Our GPR model can predict subpopulations that do not respond to the treatment. This can help physicians tailor treatments for individual patients based on their conditions. ... The model can also estimate confidence intervals for its predictions, allowing physicians to judge how confident the predictions are,” the study authors wrote.

However, they cautioned that because the model was built using patients of European descent they did not expect it to achieve a similar performance in other populations. “Extension of the model over other populations requires new patient data and separate feature selection.”

The research was supported by the National Science Foundation and the National Natural Science Foundation of China. Several of the researchers reported financial relationships with pharmaceutical or technology companies.

SOURCE: Guan Y et al. Arthritis Rheumatol. 2019 Jul 24. doi: 10.1002/art.41056.

A machine-learning model that uses clinical profiles and genetic information has shown promise in predicting which rheumatoid arthritis patients respond to anti–tumor necrosis factor drugs in a patient population of European descent.

The model can “help up to 40% of European-descent anti–tumor necrosis factor [TNF] nonresponders avoid ineffective treatments” when compared with the usual “trial-and-error practice,” according to the authors led by Yuanfang Guan, PhD, of the department of computational medicine and bioinformatics at the University of Michigan, Ann Arbor.

The ability to accurately predict rheumatoid arthritis patients’ response to treatments would provide valuable information for optimal drug selection and would help potential nonresponders avoid drug expenses and side effects, such as an increased risk of infections, Dr. Guan and coauthors noted in Arthritis & Rheumatology.

The investigators used a modeling technique called Gaussian process regression (GPR) to predict anti-TNF drug responses. “GPR is designed to predict the unknown dependent variable for any given independent variables based on known but noisy observations of the dependent and independent variables,” they explained.

The model they used won first place in the Dialogue on Reverse Engineering Assessment and Methods: Rheumatoid Arthritis Responder Challenge, which used a crowd-based competition framework to develop a validated molecular predictor of anti-TNF response in RA.

The model was developed and cross-validated using 1,892 patients randomly selected from a training data set of 2,706 individuals of European ancestry compiled from 13 patient cohorts. All patients met 1987 American College of Rheumatology criteria for RA or were diagnosed by a board-certified rheumatologist. In addition, patients were required to have at least moderate disease activity at baseline, based on a 28-joint Disease Activity Score (DAS28) greater than 3.2.

The research team also evaluated the model using an independent dataset of 680 patients from the CERTAIN (Comparative Effectiveness Registry to study Therapies for Arthritis and Inflammatory Conditions) study.

The model combined demographic, clinical, and genetic markers to predict patients’ changes in DAS28 24 months after their baseline assessment, and identify nonresponders to anti-TNF treatments, the authors explained.

“Specifically, the [model] predicts the changes in [DAS28] of patients who have taken 12 months of anti-TNF treatments, and also classifies the patients’ responses based on the EULAR response metric,” they wrote.

Results showed that, in cross-validation tests, the model predicted changes in DAS28 with a correlation coefficient of 0.406, correctly classifying responses of 78% of subjects, with an area under the receiver operating characteristic curve (AUROC) of about 0.66.

In the independent test, the method achieved a Pearson correlation coefficient of 0.393 in predicting the change in DAS28.

Genetic SNP biomarkers provided a small additional contribution to the prediction on top of the clinical models, the authors noted.

“Compared to traditional trial-and-error practice, our model can help up to 40% of European-descent anti-TNF nonresponders avoid ineffective treatments. The model performance is even comparable to some published models utilizing additional biomarker data, whose AUROC ranges from 55% to 74% over various testing sets,” they wrote.

The GPR model has practical advantages in clinical application, unlike many sophisticated machine-learning algorithms, according to the authors. For example, GPR is a well-studied statistical model, its similarity-modeling approach is intuitive, and its results are easy to interpret.

“Our GPR model can predict subpopulations that do not respond to the treatment. This can help physicians tailor treatments for individual patients based on their conditions. ... The model can also estimate confidence intervals for its predictions, allowing physicians to judge how confident the predictions are,” the study authors wrote.

However, they cautioned that because the model was built using patients of European descent they did not expect it to achieve a similar performance in other populations. “Extension of the model over other populations requires new patient data and separate feature selection.”

The research was supported by the National Science Foundation and the National Natural Science Foundation of China. Several of the researchers reported financial relationships with pharmaceutical or technology companies.

SOURCE: Guan Y et al. Arthritis Rheumatol. 2019 Jul 24. doi: 10.1002/art.41056.

A machine-learning model that uses clinical profiles and genetic information has shown promise in predicting which rheumatoid arthritis patients respond to anti–tumor necrosis factor drugs in a patient population of European descent.

The model can “help up to 40% of European-descent anti–tumor necrosis factor [TNF] nonresponders avoid ineffective treatments” when compared with the usual “trial-and-error practice,” according to the authors led by Yuanfang Guan, PhD, of the department of computational medicine and bioinformatics at the University of Michigan, Ann Arbor.

The ability to accurately predict rheumatoid arthritis patients’ response to treatments would provide valuable information for optimal drug selection and would help potential nonresponders avoid drug expenses and side effects, such as an increased risk of infections, Dr. Guan and coauthors noted in Arthritis & Rheumatology.

The investigators used a modeling technique called Gaussian process regression (GPR) to predict anti-TNF drug responses. “GPR is designed to predict the unknown dependent variable for any given independent variables based on known but noisy observations of the dependent and independent variables,” they explained.

The model they used won first place in the Dialogue on Reverse Engineering Assessment and Methods: Rheumatoid Arthritis Responder Challenge, which used a crowd-based competition framework to develop a validated molecular predictor of anti-TNF response in RA.

The model was developed and cross-validated using 1,892 patients randomly selected from a training data set of 2,706 individuals of European ancestry compiled from 13 patient cohorts. All patients met 1987 American College of Rheumatology criteria for RA or were diagnosed by a board-certified rheumatologist. In addition, patients were required to have at least moderate disease activity at baseline, based on a 28-joint Disease Activity Score (DAS28) greater than 3.2.

The research team also evaluated the model using an independent dataset of 680 patients from the CERTAIN (Comparative Effectiveness Registry to study Therapies for Arthritis and Inflammatory Conditions) study.

The model combined demographic, clinical, and genetic markers to predict patients’ changes in DAS28 24 months after their baseline assessment, and identify nonresponders to anti-TNF treatments, the authors explained.

“Specifically, the [model] predicts the changes in [DAS28] of patients who have taken 12 months of anti-TNF treatments, and also classifies the patients’ responses based on the EULAR response metric,” they wrote.

Results showed that, in cross-validation tests, the model predicted changes in DAS28 with a correlation coefficient of 0.406, correctly classifying responses of 78% of subjects, with an area under the receiver operating characteristic curve (AUROC) of about 0.66.

In the independent test, the method achieved a Pearson correlation coefficient of 0.393 in predicting the change in DAS28.

Genetic SNP biomarkers provided a small additional contribution to the prediction on top of the clinical models, the authors noted.

“Compared to traditional trial-and-error practice, our model can help up to 40% of European-descent anti-TNF nonresponders avoid ineffective treatments. The model performance is even comparable to some published models utilizing additional biomarker data, whose AUROC ranges from 55% to 74% over various testing sets,” they wrote.

The GPR model has practical advantages in clinical application, unlike many sophisticated machine-learning algorithms, according to the authors. For example, GPR is a well-studied statistical model, its similarity-modeling approach is intuitive, and its results are easy to interpret.

“Our GPR model can predict subpopulations that do not respond to the treatment. This can help physicians tailor treatments for individual patients based on their conditions. ... The model can also estimate confidence intervals for its predictions, allowing physicians to judge how confident the predictions are,” the study authors wrote.

However, they cautioned that because the model was built using patients of European descent they did not expect it to achieve a similar performance in other populations. “Extension of the model over other populations requires new patient data and separate feature selection.”

The research was supported by the National Science Foundation and the National Natural Science Foundation of China. Several of the researchers reported financial relationships with pharmaceutical or technology companies.

SOURCE: Guan Y et al. Arthritis Rheumatol. 2019 Jul 24. doi: 10.1002/art.41056.