L.B. Zwep (1, 2), K.L.W. Duisters (2), T. Hankemeier (1), J.J. Meulman (2), P.J. Upadhyay (1), J.G.C. van Hasselt (1)
(1) Leiden Academic Centre for Drug Research, Leiden University, Leiden, The Netherlands. (2) Mathematical Institute, Leiden University, Leiden, The Netherlands.
Background:
The availability of patient-specific high-dimensional molecular profiling approaches (e.g. “omics”) can potentially be used to predict drug treatment response in patients is increasing. Machine learning approaches in combination with clinical outcomes are increasingly explored to predict optimal drug treatments [1]. However, these approaches are of limited utility to identify molecular predictors associated with specific pharmacological response metrics, nor can they be used for dose regimen optimization. Pharmacokinetic-pharmacodynamic (PKPD) tumor growth modelling has been used extensively to describe tumor growth kinetics in animal models and patients [2], which allows pharmacological characterization of treatment efficacy and resistance and prediction of clinical outcomes [3]. The integration of high-dimensional omics data with PKPD models would represent an important next step to increasing understanding of molecular determinants for treatment response, and to improve the use of such data to develop personalized treatments.
Objectives:
We propose two approaches to demonstrate how machine learning and PKPD modeling can be combined, by: i) selection of biological pathways associated with either drug effect or treatment resistance and ii) prediction of tumor growth kinetics based on high-dimensional omics data.
Methods:
Data: We used a previously published study that characterized tumor growth profiles in patient-derived xenografts (PDX) [4].
Pharmacometric modeling
A pharmacometric tumor growth model based on ordinary differential equations (ODEs) was developed in a nonlinear mixed effect modeling framework [3].
Group-lasso based identification of biological pathways
Transcriptomic and genomic from patient-specific tumors were then used to predict treatment-specific pharmacometric model parameters (i.e. KD, KR) using a group-lasso [5]. We linked the omics-data to pathway databases to define grouping of genes according to biological pathways.
Multivariate lasso-based prediction of tumor growth dynamics
Simultaneous prediction of tumor growth model parameters using tumor-specific omics-data was performed using a multivariate lasso model with genomic and transcriptomic data-associated genes as predictors. The tumor growth model parameters were predicted in a leave-one-out fashion. The expected tumor growth profiles were derived by solving the ODEs for both the estimated parameters and the predicted parameters. Performance was evaluated by a scaled area between the curves (sABC) at 2 months (56 days).
Results:
Standard model diagnostics showed the tumor growth model captured most tumor growth curves. In 18% of the growth curves, a resistance effect could be identified (LRT, p < 0.05).
Cross-validation showed improved predictive performance of omics data as compared to the null model for genomic and transcriptomic data for some treatments (35 and 26 treatments, respectively). For these treatments, the group-lasso identified 67 unique pathways associated with treatment response and resistance. For several treatments, similar pathways were identified indicating treatment-overarching mechanisms for treatment response and resistance.
Prediction of tumor growth-specific parameters using the multivariate lasso was evaluated using the sABC. For the genomic prediction we found 9% very good predictions (sABC < 0.1), another 9% of the curves was predicted good (0.1 < sABC < 0.2) and another 26% was predicted sufficiently (0.2 < sABC < 0.5). These results indicate tumor growth dynamics can be adequately recapitulated provided sufficient predictive information is available in the omics-data.
Conclusion:
We describe a novel practical methodology for integration of pharmacometric models with two ML-based lasso implementations, for the integration of longitudinal (tumor response) biomarkers and high-dimensional omics data.
Our methodology allows to derive novel pharmacological insights on biological pathways associated with variability in treatment response. Furthermore, we show how integration of pharmacometric modeling and the multivariate lasso enabled the prediction of expected tumor growth kinetics based on omics-data alone, which could be associated with a tumor growth-clinical outcome model in the future.
By using a two-step approach, we were able to utilize the developed tumor growth model to demonstrate two different applications for integration with lasso-based ML approaches.
References:
[1] Mizutani T, et al. J Radi Res. 2019;60(6):818–824.
[2] Ribba B, et al. CPT Pharmacometrics Syst Pharmacol. 2014;3(5):1-10.
[3] Claret L, et al. J Clin Oncol. 2009;27(25):4103-4108.
[4] Gao H, et al. Nat Med. 2015;21(11):1318-1325.
[5] Yuan M. & Lin Y. J R Stat Soc B. 2006;68(1):49-67.
Reference: PAGE () Abstr 9443 [www.page-meeting.org/?abstract=9443]
Poster: Oral: Methodology - New Tools