Beyond the Black Box: Tools and Techniques for Explaining Causal ML Models in Different Phases of Dr

Mouse models have long been used to characterize human diseases and drug signatures. However, due to differences between organisms, translating genomic changes can be quite challenging. Therefore, understanding the linkages between human and mouse remains crucial.
In our study, we utilized PBMC mouse and human single-cell RNA-seq time course data under trauma conditions to conduct transfer learning analysis. We carefully examined the characteristics of genes with good translatability and explored their linkage to 'causal' genes. Collectively, our work aims to enhance the understanding of disease mechanisms and drug actions based on animal models. 

The automated extraction of patterns and structures to predict trial outcome from vast repositories of data is a fundamental benefit of new AI/ML methods. Depending on the specific domain of application, distinct challenges manifest themselves, necessitating tailored approaches and methodologies.

In the biopharmaceutical domain, data is inherently complex, stemming from the wide range of sources, formats, and structures involved throughout the drug development process. These include clinical trial design specifications, pharmacokinetic and pharmacodynamic profiles, and disease-specific biological and clinical information. This diversity introduces significant challenges in data integration and interpretation. Additionally, the rapid advancement of data generation technologies and the increasing public availability of biomedical datasets continue to expand the data landscape. As a result, there is a growing need for sophisticated analytical methods capable of extracting meaningful insights across the research and development lifecycle.

In this presentation, we will examine the Hierarchical Interaction Network (HINT), a graph neural network architecture introduced by Fu et al. (2022), which is designed to model complex relationships across heterogeneous data types—including structured variables, unstructured text, and graph-based representations. We proposed a new and improved HINT. The presentation will cover the core principles of the HINT framework, along with a discussion of datasets specifically curated to reflect the unique challenges of biopharmaceutical research.
 

Individualized treatment regimen (ITR) and subgroup identification are active research areas in clinical trial biostatistics. Recent advances have leveraged modern machine learning methods to identify patient characteristics that predict a stronger benefit of experimental therapies compared to the current standard of care. However, many of these methods produce black-box solutions. While some researchers have proposed simplifying treatment recommendation rules using tree-based methods, the results have been less than satisfactory. This presentation will discuss interpretability tools for treatment recommendations based on Phase III clinical trials. We will reflect on the feasibility of achieving the holy grail of individualized medicine and explore possible paths leading to it. 

In this talk (based on our recent tutorial in Statistics in Medicine) I review recent advances in statistical methods for identification and evaluation of Heterogeneous Treatment Effects (HTE), including subgroup identification, estimation of conditional average treatment effects (CATE), and individualized treatment regimens (ITR) using data from randomized clinical trials and observational studies. Several classes of methods including indirect approaches based on modeling response surface as a function of treatment and covariates, and direct approaches targeting causal estimands of interest. Selected approaches will be evaluated using simulated data mimicking randomized clinical trials and observational studies with non-random treatment assignment. 

Heart failure (HF) is a complex clinical syndrome with a significant global burden. Recent clinical trials and observational studies have demonstrated that guideline-directed medical therapy (GDMT) for heart failure, encompassing both preserved and reduced ejection fraction phenotypes, can significantly reduce cardiovascular mortality, heart failure hospitalization, and all-cause mortality.
The current GDMT recommendations include four medication classes: β-blockers, renin-angiotensin-aldosterone system inhibitors (angiotensin receptor neprilysin inhibitors or angiotensin-converting enzyme inhibitors/angiotensin receptor blockers), sodium-glucose cotransporter-2 inhibitors, and mineralocorticoid receptor antagonists. However, real-world studies have revealed challenges in implementing GDMT due to increased polypharmacy and treatment complexity, leading to suboptimal initiation and titration of medications.
Several implementation strategies have been developed to improve the usage of GDMT, but most of them focus on treatment for various stages of heart failure with limited sample sizes. There is a critical need for a personalized GDMT model that can assess different treatment plans based on individual patient characteristics to maximize improvements in patient outcomes.
In this talk, we propose a causal Machine Learning (causalML) method-based predictive modeling approach. This innovative data-driven method aims to evaluate individual treatment effectiveness, potentially supporting clinical decision-making for more tailored patient care. The model is designed to generate timely treatment suggestions that could serve as an additional resource for clinicians to develop personalized approaches for HF patients. Additionally, this work shares learning experiences and lessons learned regarding the application of causal machine learning methods to real-world evidence studies.
Keywords: Causal Machine Learning, Personalized Medicine, Guideline-Directed Medical Therapy, Heart Failure, Real-World Evidence, Real-World Data