Project

Standigm currently focuses on predicting new indications for existing drugs, called drug repositioning through deeply trained AI model with molecular features of drug responses and drug uses.

Project 1: Predicting synergetic drug combinations
The goal of the project is to predict synergetic drug combinations that effectively kill cancer cells. The disease context, the characteristics of drugs, and the interaction between drugs and diseases are described using various chemical properties of drugs, target proteins/signaling pathways of drugs, monotherapy drug response data and genomic data for cell lines, among others. Based on these data, we designed three types of input features (cell line-specific, drug-specific, and drug-cell line interaction features). With various combinations of these features and different learning parameters, we built an ensemble of gradient boosting classifiers, which ranked 3rd among 71 teams worldwide in "AstraZeneca-Sanger Drug Combination Prediction Dream Challenge (2015~2016)". The methods will be applied to precision medicine, patient stratification, and therapeutic/diagnostic design.
Status: Completed in May 2016
Project 2: Learning drug-perturbed data
The goal of the project is to predict new indications for existing drugs, called drug repositioning, from the features of drug responses and drug uses discovered via deep learning. Here, the drug responses and uses are described by the data of drug-perturbed gene expressions and of therapeutic usage labels, respectively. We have been developing various multimodal learning methods (XGBoost, Compact Bilinear Pooling, etc.) for robust feature learning. These models suggest 20 hits with new drug indications. For two of the 20 drug repositioning hits, literature has shown their experimental validations. For the remaining hits, Standigm is designing in-vivo and in-vitro experimental validations to demonstrate Standigm.AI’s prediction power. Our drug repositioning AI will discover hidden relational patterns of drugs and indications, ultimately saving the cost of drug development.

Status: Completed in January 2017
Project 3: Building & Learning biological knowledge
The goal of the project is to predict missing relationships between proteins, drugs, and diseases by learning from a biological knowledge GraphDB. We constructed the biological GraphDB consisting of nodes (32,373 proteins, 7815 drugs, 3721 diseases, etc.) and edges (55,618 drug-disease edges, 125,686 drug-protein edges, etc.) using knowledge from various open and private sources. Users can easily access and visualize the GraphDB via web browsers and navigate the DB using the Cypher query language. We are developing models to predict unknown links between the nodes (proteins, drugs, disease etc.) in the GraphDB. The models will be applied to finding new drug indications, new drug targets, and new disease biomarkers.
Status: In progress (September 2017)
Project 4: Drug Target Prediction
The goal of the project is to select reliable targets of drugs by prediction of binding affinity of drugs to proteins based on machine learning approaches. 13,308 protein-ligand complex structures with binding affinity values are learned using the combination of various featurization methods and learning algorithms. The inputs of models are experimental protein-drug complex structures or structures generated by docking simulation. The best targets are selected by the ensemble of scoring functions of different models. The methods will be applied to finding new targets(proteins) or the best drug for the protein.

Status: In progress (September 2017)
Project 5: EMR data representation
The goal of the project is to develop intelligent methods for patient and medical record representation learning, and ultimately predict clinical events such as patient diagnosis, prognosis, or medication categories. We collaborate with Ajou University School of Medicine, Korea. This research is supported by a grant of the Korea Health Technology R&D Project, “Development of the Artificial Intelligence supporting a Clinical Trial System (AI-CTS)”
Status: In progress (September 2017)