Minisymposium

The intersection of precision medicine and machine learning (ML) offers a wide range of problems and possible approaches. The prediction of tumor response to single and combination drug agents is an active area of ML application development, as tens of deep learning models are currently available from the community and are under active development. Comparing the behavior of these models is very difficult, and is not a well-studied area, as different projects differ wildly in their problem assumptions and approach to the problem. A range of other problems must also be addressed, including assessing the robustness of the models across a range of health science use cases, hardware resources, and other situations. In this presentation, we will describe our approach to build infrastructure to support the studies outlined above. We are developing a scalable workflow framework to manage, curate, and execute community models in varying scientific problems. Typical use cases include hyperparameter optimization to tune models for general or specific use cases, comparison to study the differences across models, and cross-study analyses that compare training datasets. This presentation will cover the internal design of the system, how it may be extended and used, and results from supercomputers Polaris and Aurora.

In the evolving field of computational drug design and discovery, the accurate prediction of drug responses through deep learning models remains a significant challenge due to varying methodologies in model implementation and validation. This inconsistency hampers the objective assessment of model capabilities across different drug representation methods, architectures, and datasets. As models become more complex and datasets more diverse, the necessity for standardized model comparison methodologies becomes imperative. Traditional comparison approaches, which typically rely on performance scores from disparate studies, lead to incomparable and inconsistent results, obstructing the understanding of factors critical to predictive performance. Addressing this issue, I will discuss our results based on foundaitonal models and large scale CMP-CV workflow, an automated cross-validation framework designed for the consistent training and evaluation of multiple deep-learning models. By employing standardized datasets, preprocessing techniques, and performance metrics, CMP-CV fosters controlled experimentation while allowing systematic variation in model hyperparameters and architectures. Additionally, the framework supports custom analytical functions, enabling a more profound investigation into model representations and associated uncertainties, thereby establishing a more standardized and comprehensive approach to model comparison in drug response prediction.

Fueled by pervasive supercomputing and growing availability of response data have led to a dramatic increase in the availability of predictive drug response models. Over the past several years, the collaboration between the US Department of Energy and the US National Cancer Institute have resulted in multiple advances towards an Open and FAIR ecosystem for creating, evaluating and sharing drug response models. The presentation will provide an overview of several of these projects providing publicly available resources including IMPROVE (Innovative Methods and data for Predictive Oncology Validation and Evaluation), ATOM (Accelerating Therapeutics for Opportunities in Medicine) and MODAC (the NCI Predictive Oncology Model and Data Clearinghouse). Future directions and priorities will also be shared in terms of ethical AI and guidance for transparency to build trust and reuse of trained models.

This panel discussion will explore the landscape of artificial intelligence in personalized medicine, focusing on the development, integration, and regulatory challenges of AI models designed to predict tumor responses, treatment toxicities, and other related responses. Experts will discuss the sustainability of model portability, transparency, and the incorporation of digital twins, addressing the technological and ethical hurdles in scaling these models for clinical impact. The session aims to explore the complexities of using AI to tailor patient-specific treatments, supporting innovation while guaranteeing patient safety in the rapidly evolving field of medical AI.