Abstract The costly process of bringing new therapeutics to market and high attrition rates have motivated the search for new frameworks in drug research and development (R&D). These challenges extend to the clinical space with the need for better patient stratification and therapy matching. Moreover, despite significant leaps in deep learning, even the most sophisticated methods rely on static analytical structures. Thus, unaddressed needs in therapy development and applications call for unconventional thinking to capture dynamic processes across preclinical and clinical spaces. With this review, we trace how crucial algorithmic pieces have been coming together over the past decades for the next generation of deep learning, which we define as adaptive learning. This new class of robust analytical architectures will be enabled through self-organised models where inputs changing over time can guide deep networks to adapt to biases and optimise learning. There have already been glimpses of such groundbreaking solutions in liquid neural networks (LNNs), graph attention algorithms, digital twins, and engineering research. As we review multiple examples and applications, we want to highlight the emerging fundamental shifts in discovery and analytical paradigms. Only by continuing to develop new frameworks can we capture complex disease interactomes and identify or improve therapeutic avenues. Insight Box Our work underscores the emerging shifts in research and development (R&D) and drug discovery from a deep learning perspective. First, we identify and discuss the missing link in drug discovery that affects multiple areas in translational research. We then demonstrate how critical algorithmic, analytical, and technological pieces have been coming together to address these challenges. Furthermore, we illustrate how applied AI and deep learning frameworks will need to change and what solutions are already available. Consequently, we employ examples of novel deep learning architectures, digital twins, and clinical research. Notably, there has been very little discussion to acknowledge current limitations in deep learning from a translational perspective. Thus, our work offers not only new insights but also a direction of travel for future developments, which we define as adaptive learning.
Kanapeckaitė et al. (Thu,) studied this question.