MolCure

What it does -

Our web app is a drug-target interaction prediction platform that generates optimized antibiotic combinations based on bacterial RNA analysis. It allows users to input bacterial RNA sequences and, through a series of bioinformatics and AI-driven processes, predicts the most effective molecular structures of antibiotics tailored to attack the given bacteria.

RNA Sequencing & Gene Expression Analysis: The platform maps the bacterial genome from RNA sequences and analyzes protein expression levels to determine the composition of the bacterial cell wall.
AI-Powered Drug Prediction: Using a machine learning model trained on curated datasets, the system generates different drug combinations optimized to target the bacteria’s specific cell wall properties.
Visualization of Antibiotic Structures: The final output includes both 2D and 3D molecular models of the predicted antibiotic, allowing for a detailed structural analysis.

Inspiration

AI isn’t just for software development—it can transform biology and chemistry too. Antibiotic resistance is a growing threat; once-treatable infections are becoming life-threatening due to resistant bacteria. We saw an opportunity to use AI to accelerate the discovery of new antibiotics and address this urgent need. Our goal was also to bridge the gap between AI and bioinformatics, showing how these tools can be accessible and impactful in drug discovery.

AI-driven drug discovery can help identify new antibiotics to combat resistant bacteria.

What We Learned

Diving into bioinformatics and cheminformatics was challenging but rewarding. We learned the basics of bacterial cell wall structures and how antibiotics interact with these components. Understanding mechanisms of antibiotic action and how bacteria develop resistance gave us crucial context. We also explored molecular docking principles and how a drug’s molecular structure influences its efficacy. On the AI side, we learned to fine-tune models for chemical data, work with molecular datasets, and seamlessly integrate different AI frameworks.

How We Built It

Frontend: Developed an interactive UI with React and TypeScript for a smooth user experience.
Backend: Used Python (Flask/FastAPI) to handle requests and process molecular data.
AI Models: We integrated Chemprop and MolGPT. Chemprop is a message-passing neural network for molecular property prediction, which has been used to predict antibiotic activity (e.g., identifying the novel antibiotic halicin from a chemical library). MolGPT is a GPT-based model for generating new molecular structures; we used it to propose novel antibiotic candidates.
Dataset: Curated a specialized novel-made pipeline that predicts the bacterial cell wall composition molecule structures and known antibiotics, pulling from ChEMBL (a large database of bioactive drug-like molecules) and other sources.
Cloud & Deployment: Deployed the frontend via AWS Amplify and handled AI model inference with AWS Bedrock. (Amazon Bedrock is a fully managed service providing access to high-performing foundation models). We also integrated models from Hugging Face, a leading platform for sharing and deploying machine learning models.
Visualization: Employed RDKit and a 3D molecular viewer to display antibiotic structures dynamically. RDKit is an open-source cheminformatics toolkit that supports 2D/3D molecular operations and visualization, allowing users to see the molecular structure of each antibiotic prediction.

After connecting all these components, we achieved a cohesive pipeline where a user input (bacterial cell wall molecular profile) flows through our AI models to predict an effective antibiotic, which is then visualized in the browser.

System architecture illustrating how the frontend, AI models, and cloud services work together in the antibiotic discovery platform.

Challenges We Faced

Understanding Bioinformatics: Coming from a computer science background, we had to quickly learn how molecules behave in biological systems. Grasping bacterial cell wall composition and antibiotic mechanisms was a steep learning curve.
Data Acquisition & Preprocessing: Finding the right data was tough. We needed molecular data on various bacterial cell wall components and known antibiotics. Curating and cleaning this data (from sources like ChEMBL and scientific literature) to fit our model’s format was time-consuming but essential.
Algorithm Development: We devised an RNA sequencing-inspired approach to improve how we predict antibiotic interactions. This novel method required blending knowledge from genomics with machine learning, and tuning it involved a lot of trial and error.
Framework Pivot: Initially, we considered using SyntheMol for molecule generation. SyntheMol is a generative AI method for designing novel antibiotic candidates. However, after some testing, we realized it wasn’t the best fit for our pipeline’s needs. This meant pivoting to a new strategy (using MolGPT and Chemprop together), which was a late but necessary change.

Achievements

AI-Driven Matching: Built an AI-powered system that predicts and recommends the best antibiotic based on the molecular structure of a bacterium’s cell wall. This personalized matching could help target bacteria more effectively.
Integrated Models: Successfully combined multiple AI models in one seamless pipeline. The generative model proposes new antibiotic structures, and the predictive model evaluates their efficacy, all behind the scenes.
Interactive Visualization: Created an interactive 3D visualization tool to showcase each antibiotic’s molecular structure. Users can see the shape and features of the predicted antibiotic, making the results more tangible and interpretable.
User-Friendly Platform: Developed a modern web interface that makes this complex AI technology accessible. Even users without a background in AI or bioinformatics can input data and understand the results, helping to democratize AI-driven drug discovery.

Our platform visualizes predicted antibiotic molecules in 3D, allowing users to inspect the structure and understand how the drug might interact with bacterial cells.

What's Next?

We have big plans to extend this project. First, we want to include more bacterial strains – expanding beyond the initial set of pathogens so the tool can recommend antibiotics for a wider range of bacteria. We’re also looking to improve the molecule generation process by incorporating additional generative AI models, which could propose even more diverse antibiotic structures. Our RNA sequencing-inspired prediction approach will be refined for greater accuracy, aligning the model’s predictions more closely with real-world biological outcomes. Finally, we aim to optimize the platform for real-world use, possibly collaborating with biologists and pharmacologists to validate our AI-generated antibiotic suggestions in laboratory settings. With these enhancements, we hope our project can move from a prototype towards a practical tool in the fight against antibiotic resistance.