Deep Generative Modelling in Systems Medicine: From Transcriptomics Data to Drug Development

Abstract

Artificial intelligence (AI) has been widely applied across many domains, including the biomedical research field. These AI models mostly include discriminative models such as classifiers and regressors that predict for example diseases, disease progression or susceptibility to drugs based on patient data. Outside of the medical field, however, there has been a recent surge of another type of AI: generative models that allow to create data rather than discriminate it. In this work, I illustrate the opportunities offered by generative models in the context of systems medicine on two application scenarios. <br /> Firstly, I investigate the utility of several types of generative models on the task of generating synthetic transcriptomics data. To unlock the full potential of transcriptomics data, making it widely available to other researchers is essential. However, its highly personal nature requires measures to protect patient privacy. In this work, I investigate generative AI for the generation of private synthetic patient cohorts with the goal to enable sharing of transcriptomics data without privacy-violations. I evaluate the synthetic cohorts generated by models that do or do not protect patient privacy for biological soundness. This includes the data’s utility to train classifiers for real data, as well as the preservation of differential expression and co-expression of genes. <br /> Secondly, I address generative modelling for the property-based de novo design of compounds. I firstly focus on the problem of AI-learnable representations of molecules and particularly consider the requirements of such a representation and investigate how well they preserve chemical information content. Specifically, I research the impact of the model architecture of autoencoders on the embedding quality of encoded molecules. From that, I devise an optimisation strategy that not only enhances embedding quality, but also reduces the resources required for training the autoencoder. In a time where global warming progresses at an alarming speed and where resources are waning, measuring the utility of new inventions needs to include the consideration of its environmental impact. I then consider the task of devising new compounds with generative AI. Here, my focus lies especially on the guided generation based on desired properties. During the model design, I additionally concentrate on devising an architecture that is flexible, allowing easy addition of further properties without retraining the model. I demonstrate the effectiveness of the approach for single-property as well as multi-property conditioning.