https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9109907/
Writing – original draft, Writing – review & editing, 1 ,* Jennifer L. Wilson, Conceptualization, Resources, Supervision, Writing – original draft, Writing – review & editing, 2 Davide Bacciu, Conceptualization, Supervision, Writing – original draft, Writing – review & editing, 1 Kevin J. Grimes, Conceptualization, Resources, Supervision, Writing – original draft, Writing – review & editing, 2 and Corrado Priami, Conceptualization, Supervision, Writing – original draft, Writing – review & editing 1
Schizophrenia is a debilitating psychiatric disorder, leading to both physical and social morbidity. Worldwide 1% of the population is struggling with the disease, with 100,000 new cases annually only in the United States. Despite its importance, the goal of finding effective treatments for schizophrenia remains a challenging task, and previous work conducted expensive large-scale phenotypic screens. This work investigates the benefits of Machine Learning for graphs to optimize drug phenotypic screens and predict compounds that mitigate abnormal brain reduction induced by excessive glial phagocytic activity in schizophrenia subjects. Given a compound and its concentration as input, we propose a method that predicts a score associated with three possible compound effects, i.e., reduce, increase, or not influence phagocytosis. We leverage a high-throughput screening to prove experimentally that our method achieves good generalization capabilities. The screening involves 2218 compounds at five different concentrations. Then, we analyze the usability of our approach in a practical setting, i.e., prioritizing the selection of compounds in the SWEETLEAD library. We provide a list of 64 compounds from the library that have the most potential clinical utility for glial phagocytosis mitigation. Lastly, we propose a novel approach to computationally validate their utility as possible therapies for schizophrenia.
Phagocytosis is a fundamental biological process to protect biological organisms from exogenous infectious particles as well as to preserve equilibrium and efficiency of the host by removing its unwanted cells. A dysregulation of the phagocytic activity can lead to severe consequences for the host. In this study, we focus on a recent theory that relates an excessive phagocytic activity in brain cells, and a consequent abnormal reduction in brain volume, to the development of schizophrenia. Our working hypothesis is that pharmaceutical compounds that can reduce excessive of phagocytic activity might prove effective as a schizophrenia treatment. Rather than attempting to develop ex-novo such a chemical compound, we rely on a more cost-effective and efficient approach that seeks candidate therapies in a set of approved chemical compounds. To achieve this, we train a machine learning model capable of predicting, with good accuracy, the ability of a molecular compound to increase or decrease phagocytosis in the target brain cells. Our approach leverages learning models capable of directly processing the molecular graph of the compound, leading to the identification of 64 candidate drugs of potential clinical utility.
This is a PLOS Computational Biology Methods paper.
Schizophrenia is a chronic and severe mental disorder that affects how a person thinks, feels, and behaves. It is expressed as a combination of symptoms, such as recurrent psychosis, social withdrawal, anhedonia, and cognitive dysfunctions. Worldwide about 1% of the population is diagnosed with schizophrenia, with 100,000 new cases annually only in the United States [1].
A recent study [2] states that brain volumes, measured on Magnetic Resonance Imaging (MRI) scans, are abnormal in patients with schizophrenia compared to unaffected individuals, with a reduction in both grey and white matter. In particular, the decreased density of dendritic spines in schizophrenia subjects has been supposed by MacDonald et al. [3] as the result of an excessive pruning activity against synapses. This action is assumed to be performed by glial cells, which are non-neuronal cells with multiple functions in the central nervous system to support and remove neurons. This assumption is supported by the evidence that glial phagocytic activity may be directly associated with the prevalence of various neurodegenerative diseases due to hyperactivation of phagocytic pathways [4, 5]. In addition, the novel PET tracer binds to synaptic vesicle glycoprotein 2A (SV2A) and shows diminished uptake in the frontal and anterior cingulate cortex in individuals with schizophrenia [6].
Towards the goal of discovering novel treatments for schizophrenia, previous work conducted a large-scale phenotypic screen to discover compounds with the ability to alter glial cell phagocytosis. However, understanding structure activity relationships is a challenge in these screens. Further, generating accurate models is difficult because it is not known what chemical information is most associated with predicting chemical function and how to best represent this information for predictive models. Additionally, further experiments remain the gold standard for validating model predictions. Yet, it is not always possible to conduct additional high-throughput screens and we require alternative methods for testing the utility of model predictions.
The compound property/activity prediction problem is a task faced by pharmaceutical companies and academia to improve the comprehension of diseases, discover new drugs, or identify new indications of existing drugs. It is standard practice to scan large libraries of compounds to test their biological activity. However, this operation can be costly and time-consuming, and Machine Learning (ML) methods can be helpful to reduce the effort needed to run experiments. For that reason, in the last decades, several computational approaches have been proposed to determine compound properties, or as filter to select the most promising compounds for clinical and biological experiments [7, 8]. A pioneering work is that by Bianucci et al. [9], where the authors employed Cascade Correlation Networks for structures to predict the boiling point of Alkenes and to predict the affinity towards the Benzodiazepine/GABAA receptor by a group of Benzodiazepines. More recently, Banerjee et al. [10] have developed a ML model to discriminate between sweet and bitter taste of molecules. Specifically, the model leverages a static fingerprint of the molecule to predict the property through a Random Forest. Similarly, Lind et al. [11] feed a static fingerprint and oncogene mutation status to a Random Forest to predict the activity versus inactivity of drugs against cancer cell lines. These results demonstrate that it is possible to associate chemical information to biological outcomes. Yet, static fingerprints are not sufficient in all applications.
Explicitly for schizophrenia, Zhao et al. [12] explored five different ML approaches to repurpose drugs for schizophrenia, depression, and anxiety disorders. In particular, they considered Deep Neural Networks, Support Vector Machines, Elastic Net regression, Random Forest, and Gradient Boosted Trees. Models were trained to predict whether a drug is a known treatment for the disease or not, using drug expression profiles as inputs. Those profiles capture transcriptomic changes when HL60, PC3, and MCF7 cell lines were treated with a chemical. Xu et al. [13] proposed PhenoPredict, a ranking algorithm for schizophrenia drug repurposing. PhenoPredict infers drug treatments from diseases that are phenotypically related to schizophrenia. These models demonstrate the ability to connect chemical information to biological information, yet they are limited to predicting molecular changes (such as gene expression) and are not suited to predicting phagocytic activity from phenotypic screens.
This work studies the use of ML techniques to predict the effects of compounds on glial phagocytic activity that cause abnormal brain reduction in schizophrenia subjects. This work has been done in collaboration with SPARK at Stanford University [14, 15], the main node of a partnership network between university and industry experts in chemistry, biology, and medicine to advance academic biomedical research discoveries into promising new treatments for patients. The objective of this work is to propose a ML method apt to optimize drug phenotypic screens. Specifically, our method identify compounds that reduce glial phagocytic activity for the treatment of schizophrenia, which, to the best of our knowledge, has not been proposed before. Our contribution can be summarized as follows. First, we introduce a ML method based on Deep Graph Networks to predict if a compound can influence the glial phagocytic activity in the brain tissue. Then, we evaluate our method on a real high-throughput screening experiment provided by SPARK. The proposed model achieves a macro Area Under the ROC curve (AUROC) of 0.68 when predicting if a compound inhibits, intensifies, or does not affect the phagocytic activity. Afterwards, we perform an analysis to understand the potential benefits of our approach in a practical scenario. Specifically, we leverage our method to prioritize the selection of a new set of compounds in the SWEETLEAD library, leading to the identification of 64 potential candidates. Lastly, we propose a novel approach to understand the relevance of compounds to biological use case. That approach allows us to compare our results with the more than 287,000 references in the literature. With this analysis we highlight the effectiveness of the model in identify compounds that are already studied in relation to brain-related diseases.