Nuria Folguera-Blasco1, Florencia A. T. Boshier1, Aydar Uatay1, Cesar Pichardo-Almarza1, Massimo Lai1, Jacopo Biasetti2, Richard Dearden3, Megan Gibbs4, Holly Kimko5
1Systems Medicine, Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Cambridge, United Kingdom 2Systems Medicine, Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Gothenburg, Sweden 3 Machine Learning and AI, Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Cambridge, United Kingdom 4 Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Waltham, MA, United States 5 Systems Medicine, Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Gaithersburg, MD, United States
Introduction/Objectives:
We are currently living an era where ML/AI approaches seem to be (greatly) improving more traditional methods, replacing in many instances ‘old school’ techniques, also in the drug discovery field. However, is a ML/AI-driven approach always the best solution? It is well-known that to apply ML/AI approaches one of the main requirements is to have a ‘good’ dataset, where good needs to be understood in the sense of complete —the dataset captures the relevant features— and large enough —so that the dataset can be split, allowing for algorithms to be trained and tested there. Therefore, when the dataset does not fulfil those characteristics, it is clear that a ML/AI approach is not the most suitable tool to use. In fact, even when we have a ‘good’ dataset, symbiotic approaches, where ML/AI is combined with another tool, may bring a greater benefit than the use of ML/AI alone. In this poster, we will present opportunities (gAIns) and flag challenges (pAIns) that a symbiotic approach between Quantitative Systems Pharmacology (QSP) and ML/AI may bring to the pharmaceutical industry.
Methods:
In order to understand when ML/AI and QSP can work together, we will first review when each approach is suitable on its own (e.g. when there is a clear understanding of the underlying biological process, a QSP approach may be better suited). Then, we will highlight some of the opportunities for the symbiosis to be beneficial by describing some examples where the combination of QSP and ML/AI brings gAIns, i.e. the combination allows to go one step further than any of the single methodologies. We will also focus on showcasing the risks of the symbiotic approach, by enumerating some pAIns that the combination of QSP and ML/AI yields, and how those should be addressed for its mitigation.
Results:
We will present several case scenarios where the coupling of QSP models and AI/ML has a clear beneficial impact (gAIns), and we will also show some of the pAIns appearing when integrating both modelling techniques.
gAIns: Understanding clinical measurements. Clinical data is very heterogeneous, as it may contain biomarker results (e.g. blood samples), clinical records (e.g. medication history) and imaging data (e.g. X-rays). Not only this, but heterogeneity is also shown by including a mixture of quantitative and qualitative measurements. The symbiotic QSP and ML/AI approach offers the ability to gain a better understanding of such records. QSP models can be used for the dynamics of quantitative measurements (such as the cell blood counts), and a ML/AI approach can link those time-course results with clinical endpoints. Thus, the symbiotic approach represents a clear gAIn compared to more simple strategies, and seems a promising tool for treatment planning, e.g. by understanding the appearance of side effects.
pAIns: Error propagation. When applying a ML/AI technique, we know that algorithms are not 100% accurate, e.g. there are classification errors in the testing dataset. If a ML/AI tool is applied prior to a QSP model, how that error or uncertainty is integrated in the QSP modelling output? By highlighting this issue that appears when combining both modelling tools, we want to emphasise the lack of techniques developed in this area, where some theory similar to error propagation for more conventional tools, should be developed and widely accepted and adapted by the community so that its use becomes standardised.
Conclusions: Despite the arising pAIns appearing when using the symbiotic QSP and ML/AI approach, which should not be underestimated, it seems that the symbiosis is feasible and ready for success, as illustrated with the described gAIns. For its full application in drug development, QSP modellers and data scientists in the pharmaceutical industry and in academia should work together with drug regulatory agencies to further highlight gAIns and make a concerted effort to identify and overcome pAIns arising when applying the symbiotic approach, with the ultimate goal of developing useful drugs efficiently and treating patients effectively. Such symbiotic methodologies could thus take even further the applicability of ML/AI tools in drug discovery.
References:
[1] Folguera-Blasco et al. (2024). Coupling Quantitative Systems Pharmacology Modelling to Machine Learning and Artificial Intelligence: its pAIns and gAIns. Under revision.
Reference: PAGE 32 (2024) Abstr 11270 [www.page-meeting.org/?abstract=11270]
Poster: Methodology - New Modelling Approaches