SMART (Small Molecular Analysis Reporting Tool)

Inspiration

In a world where pathogens are forming greater antibiotic resistance, diseases and viruses roam, the discovery of new drugs has become vital in clinical research to create a world that is safe from diseases. However, the process of finding a new drug through clinical trials while meeting efficiency and standards of health can be a perplexing task. ​​90% of clinical trials fail. 30% of the trial consists of drugs being overly toxic. 40% to 50% lacked its desired effect meanwhile another portion had unexpected harmful drug properties that were unable to continue. Billions of dollars are funded for these drugs and at the end, most of the drug candidates fail the trials.

What it does

SMART is a service that helps medical researchers confirm whether or not a potential drug is toxic to patients. Discovering new drugs poses many challenges, especially in creating clinical trials on humans. Many drugs like bromfenac can have toxic side effects that fail to meet the standards of health. Clinical researchers must work tirelessly to ensure that a drug will not have any side effects through scouring tens of databases and analysing their potential similarity, an effort that takes both time and money.

The Small Molecular Analysis Reporting Tool (SMART) is a solution that will help clinical researchers quickly examine and determine a drug’s toxicity through AI. We can eliminate the ambiguity, cost, and time that plagues this process and causes 90% of all clinical trials to fail. Currently, SMART has a working prototype; users are able to access and use this model to find the toxicity of a drug. SMART uses a massive database from the Tox21 Collaboration and creates a neural network model with an accuracy of more than 99%.

As of 2022, there are currently no commercial-wide solutions to identify drug toxicity. Although there are few research papers and articles theorizing the idea, this is the first practical functional model, already undergoing a training of 99% training accuracy.

How I built it

SMART uses a highly complex AI model to predict drug toxicity with more than 99% accuracy. The AI has 3 parts: The dataset, the model trainer, and the interface.

The database is sourced from Tox21, a massive collaboration between multiple agencies like the U.S. Environmental Protection Agency (EPA), National Toxicology Program (NTP), and Food and Drug Administration (FDA). Tox21 has a collection of drugs and a variety of assays which show whether a drug causes an adverse effect.

The neural network is a sequential tokenized model. The input layer takes in the chemical structures of each drug as tokens, then the hidden layers perform multiple complex calculations, and the output layer returns the expected value of the drug’s toxicity.

The interface is designed for clinical researchers to quickly analyze a potential drug. Users simply input the chemical structure (called SMILES) of the drug they want to be analyzed, and the AI will return whether or not it is toxic.

Challenges I ran into

It was extremely difficult to fuse together the ideas of converting chemical bonds, structures, and data into biological assay markers and SMILES that could be recognized by the AI.

Accomplishments that I'm proud of

SMART can: 1) Screen drugs before they enter the clinical trials and prevent the loss of research money from a drug that will not be successful. 2) Perform different biomarkers assays, creating a widespread and applicable tool for the research community. 3) Tailor experiments for drugs at a certain interaction, eliminating drugs that can negatively affect patients through health-induced risks. 4) Screen for multiple markers, optimizing the efficiency.

Additionally, I was able to meet with a R&D Researcher and Staff Machine Learning Engineer for validation and feedback.

What's next for SMART (Small Molecular Analysis Reporting Tool)

The ideal deliverable after the hackathon will be a fully working user interface for clinical researchers to interact with, allowing for full functionality. Additionally, the hackathon will be an opportunity to refine all sections of SMART, from the database used by the model to the timeline and budget, launch, and serve as a project accelerator. After the hackathon, and successful development, I plan to launch SMART as a commercially available product via startup accelerators.

Built With

• machine-learning
• python
• smiles
• streamlit
• tox21
• webapp