Artificial intelligence and machine learning: just a hype or a new opportunity for pharmacometrics?
R. Bartolucci, S. Grandoni, N. Melillo, G. Nicora, E. Sauta, E.M.Tosca, P. Magni
Department of Electrical, Computer and Biomedical Engineering, University of Pavia, via Ferrata 5, Pavia, I-27100, Italy.
Objectives:
Artificial Intelligence (AI) in drug development has attracted a growing interest [1]. AI deals with computer systems able to perform human-like tasks or solve complex problems. Machine Learning (ML) is a subfield of AI whose aim is to learn from data in order to find hidden patterns that could be exploited for classification or clustering purpose. ML is further categorized into supervised, unsupervised and reinforcement learning (RL), according to the type of learning procedure on labelled (i.e., with known classes), unlabelled data or with a reward/penalty schema. Among the plethora of ML techniques, Deep Learning (DL), which exploits deep neural networks architectures, has recently outperformed previously developed methods in different tasks. The adoption of AI/ML techniques is well established during both drug discovery [2,3] and patient recruitment process [4,5]. However, ML and AI are also spreading in other drug development phases, historically supported by pharmacometrics and modelling approaches.
In this work, we provide an overview of the AI/ML applications in pharmacometrics, with the aim of understanding how AI/ML can support, substitute or be integrated with model-based approaches and trying to clarify their effective role in this field.
Methods:
We performed a literature review and we tried to classify the different contributions. Four main pharmacometric tasks, that have been approached by AI/ML, were identified in the literature: i) model building and covariate selection; ii) PK/PD parameters prediction; iii) Time-To-Event (TTE) analysis; iv) therapy optimization.
For each topic, the most significant works are reported.
Results:
1) Genetic Algorithm approaches and Stochastic Approximation for Model Building Algorithm (SAMBA) were proposed to select the best combination of structural features, covariates effects and random effects in [6-11] and [12], respectively. Moreover, automatic covariate selection has be performed exploiting Gene Expression Programming [13] or Multivariate Adaptive Regression Splines [14].
2) Artificial neural networks (ANN) were used to make PK/PD predictions: in [15] an ANN was trained to estimate the plasma concentrations of a drug in a given population, with performances comparable to those obtaining by classical NLME models with NONMEM.
3) In TTE analysis several ML techniques, such as Random Survival Forest, Support-Vector Machine, one-layer ANN and DL were used in place of the standard Cox Proportional Hazard model to discover non-linear relationships between covariates and risk [16, 17].
4) RL methods are successfully applied to therapy optimization. In [18-20], they are used to find an optimized treatment strategy that can balance drug efficacy and toxicity for a single patient, or the entire trial group, and the dose amount is decided by an agent that evaluates the patient status (i.e., tumor size, neutropenia level), obtained from model simulations.
Conclusions:
Despite AI and ML methodologies are not novel approaches and are well established in other fields or phases of drug development, they have recently raised interest in pharmacometrics. From our analysis, it emerges the contribution that the adoption of AI/ML methods could have to support different tasks. For example, in the automation of the model building process they could be a more efficient and valid alternative to the standard forward-addition/backward-elimination approaches, even if an user (subjective) evaluation of the selection steps remains essential. They are also promising for therapy optimization, especially in the perspective of personalized medicine. In other circumstances, however, they cannot substitute a model-based strategy and their data-driven approach is in contrast to the current pharmacometrician efforts towards the building of more mechanistic models. In conclusion, AI/ML methods could be successfully exploited in pharmacometrics and their capabilities should be assessed for further tasks such model-based meta-analysis, but only after a careful and critical evaluation of the characteristic of the investigated problem and an assessment of their real applicability.
References:
[1] Mak, K. K., & Pichika, M. R. (2018). Artificial intelligence in drug development: present status and future prospects. Drug discovery today
[2] Duch, W., Swaminathan, K., & Meller, J. (2007). Artificial intelligence approaches for rational drug design and discovery. Current pharmaceutical design, 13(14), 1497-1508.
[3] Popova, M., Isayev, O., & Tropsha, A. (2018). Deep reinforcement learning for de novo drug design. Science advances, 4(7), eaap7885.
[4] http://www.clinicalinformaticsnews.com/2017/09/29/the-intelligent-trial-ai-comes-to-clinical-trials.aspx
[5] https://emerj.com/ai-sector-overviews/ai-machine-learning-clinical-trials-examining-x-current-applications/
[6] Sale, M., & Sherer, E. A. (2015). A genetic algorithm based global search strategy for population pharmacokinetic/pharmacodynamic model selection. British journal of clinical pharmacology, 79(1), 28-39.
[7] Bies, R. R., Muldoon, M. F., Pollock, B. G., Manuck, S., Smith, G., & Sale, M. E. (2006). A genetic algorithm-based, hybrid machine learning approach to model selection. Journal of Pharmacokinetics and Pharmacodynamics, 33(2), 195-221.
[8] Sherer, E. A., Sale, M. E., Pollock, B. G., Belani, C. P., Egorin, M. J., Ivy, P. S., ... & Scher, H. I. (2012). Application of a single-objective, hybrid genetic algorithm approach to pharmacokinetic model building. Journal of pharmacokinetics and pharmacodynamics, 39(4), 393-414.
[9] Sale, M., Pollock, B. G., Bies, R. R, Sherer, E. A., Identification of optimal NONMEM models using a multi-objective genetic algorithm. PAGE 21 (2012) Abstr 2336 [www.page-meeting.org /?abstract= 2336]
[10] Sherer, E. A., Sale, M., Manuck S., Muldoon, M., Pollock, B. G., Bies, R. R, Three case studies of pharmacokinetic model building using a multi-objective genetic algorithm. PAGE 21 (2012) Abstr 2454 [www.page-meeting.org/?abstract=2454]
[11] Chaturvedula, A., Sale, M. E., & Lee, H. (2014). Genetic algorithm guided population pharmacokinetic model development for simvastatin, concurrently or non‐concurrently co‐administered with amlodipine. The Journal of Clinical Pharmacology, 54(2), 141-149.
[12] http://rsmlx.webpopix.org/userguide/buildmlx/
[13] Yamashita, F., Fujita, A., Sasa, Y., Higuchi, Y., Tsuda, M., & Hashida, M. (2017). An Evolutionary Search Algorithm for Covariate Models in Population Pharmacokinetic Analysis. Journal of pharmaceutical sciences, 106(9), 2407-2411.
[14] Hall, R. G., Pasipanodya, J. G., Swancutt, M. A., Meek, C., Leff, R., & Gumbo, T. (2017). Supervised Machine‐Learning Reveals That Old and Obese People Achieve Low Dapsone Concentrations. CPT: pharmacometrics & systems pharmacology, 6(8), 552-559.
[15] Chow, H. H., Tolle, K. M., Roe, D. J., Elsberry, V., & Chen, H. (1997). Application of neural networks to population pharmacokinetic data analysis. Journal of pharmaceutical sciences, 86(7), 840-845.
[16] Gong, X., Hu, M., & Zhao, L. (2018). Big Data Toolsets to Pharmacometrics: Application of Machine Learning for Time‐to‐Event Analysis. Clinical and translational science, 11(3), 305-311.
[17] Katzman, J. L., Shaham, U., Cloninger, A., Bates, J., Jiang, T., & Kluger, Y. (2018). DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC medical research methodology, 18(1), 24.
[18] Gregory Yauney, Pratik Shah (2018). Reinforcement Learning with Action-Derived Rewards for Chemotherapy and Clinical Trial Dosing Regimen Selection. Proceedings of Machine Learning Research 85 2018
[19] Houy N, Le Grand F (2018). Optimal dynamic regimens with artificial intelligence: The case of temozolomide. PLoS ONE 13(6): e0199076.
[20] Yufan Zhao et Al. Reinforcement learning design for cancer clinical trials. Stat Med. 2009 November 20; 28(26): 3294–3315.-45
