According to the US Food and Drug Administration (FDA), every year, there are more than two million serious adverse drug reactions (ADR) (i.e., injuries as a result of taking a particular dose of medication or prolong administration of a drug or a mixed of two or more drugs), taking away 100,000 lives. ADR is now the leading cause of death and morbidity, ahead of pulmonary disease, diabetes, AIDS, pneumonia and accidents. Hence, recognizing potential ADR has become an important step in drug design.
Existing AI models are not perfect
Deep learning models tapping on massive biomedical data had emerged as a new alternative of predicting ADR but most of which worked on the basis that drug with similar chemical structures tend to share similar properties. Besides, some of these models were trained with limited labelled data and their performances may be in question when new drug is concerned.
Most importantly, often, only a handful of sub-components within the drug’s molecular structure are responsible for ADR and the rest are irrelevant. So, even if the drugs do not result in an ADR, it’s still possible for their irrelevant parts to overlap. However, some models indiscriminately take into consideration the whole molecular structure, causing them to be biased towards the irrelevant part and directly undermined the ADR predictions or making them non-interpretable.
To overcome the challenges, researchers from Harvard T. H. Chan School of Public Health; IQVIA, MIT-IBM Watson AI Lab, and Georgia Institute of Technology had created a new neural network model called ChemicAl SubstrucTurE Representation (CASTER) which will perform ADR predictions based on chemical structure data of various medications. Basically, the model will consider two different medications and creates a prediction of whether the duo will interact, how they will interact and the impact of the interaction.
CASTER, the new neural network
To do so, researchers have to rewrite 3D chemical structures of various drugs into SMILES, a character format that can be read by a neural network model. CASTER will then be trained using both labelled (i.e., drugs that known to result in ADR) and non-labelled (i.e., drugs that have no reported adverse reactions) data and look into the functional sub-substructures of different pairing of drugs. As CASTER targets on chemical structures, this means it can look beyond ADR and search for possible adverse drug-food interactions (DFI) too.
In addition, CASTER comes with a deep auto-encoding module that memorizes all the ADR combinations found within the labelled and non-labelled data and generalized them onto the development of new drugs to minimize bias towards irrelevant parts and other noises. CASTER also has a dictionary learning module that will access and measure how relevant each input of sub-components is to an ADR; giving human practitioners a clearer indication.
CASTER was tested on two common drug interaction data sets and researchers found it to be more accurate and interpretable than most of other existing AI models that are used to make ADR predictions. Overall, the research team believes the model will contribute tremendously to future drug design as they continue to improve the dataset. These findings were also presented at the proceedings of the Association for the Advancement of Artificial Intelligence (AAAI) took place last month.