An invited special session exploring how modern deep learning architectures and new input encodings accelerate multi-omics analysis and drug discovery.
Abstract
In the last decade, the state-of-the-art has seen rapid advancements in designing and training powerful deep Neural Networks (NNs), especially Convolutional NNs, Graph NNs, Variational Autoencoders, Diffusion Models, Transformers, and Large Language Models (LLMs). Applying these AI models to biomedical data has opened unprecedented opportunities in the field of computational biology, revolutionizing both the traditional drug development paradigm and the way to analyze the huge amount of biomedical data coming from multi-omic approaches, which are needed to unravel the underlying mechanisms of diseases.
In particular, drug discovery has become increasingly popular after AlphaFold. Drug discovery is a complex process aimed at identifying molecules that can interact with specific molecular targets to treat diseases. It is costly and time-consuming, often requiring over a decade and more than 2 billion dollars to bring a single medicinal product to the market. The main challenges to be addressed consist of (1) the complexity of biological systems, often requiring a multi-omics approach to be understood, (2) the vastness of the chemical space—estimated to be in the order of 1060 molecules, and (3) the iterative trial-and-error process on which experimental techniques rely.
Traditional analysis methods often struggle to capture the intricate patterns hidden within these multi-omic data, due to their high dimensionality, heterogeneity, and complex interrelationships. Thanks to their ability to learn and model complex representations and relationships, advanced NNs offer a powerful solution to tackle these analyses. In addition, by screening large libraries (e.g., predicting molecular properties, binding affinities, and potential toxicity profiles) with unprecedented accuracy, they lead to the early selection of promising candidates, significantly reducing the number of compounds needing slow and expensive experimental testing.
For instance, the use of generative models allows the design of new molecules with desired properties, offering novel compounds that might not have been considered using traditional methods and that could target diseases with limited or no existing treatments. However, the effectiveness of these models heavily relies on how the input data are represented and encoded. Thus, novel encoding schemes that can capture the underlying biological mechanisms and relationships are crucial for unlocking the full potential of these advanced architectures.
Improved representations + powerful deep models can reduce experimental load, prioritize candidates, and speed up the transition from data to therapies.
Generative models, property predictors and screening strategies can uncover novel molecules and shorten timelines.
Emphasis on encoding and interpretability increases trust and facilitates integration with laboratory workflows.