The Chicken Dinner Dilemma: Sequence Alignment for AMR

🍗Inspiration

Everyday meals could be hiding a deadly secret: antibiotic-resistant (AMR) bacteria, looking back at its victim, you, from the chicken on the dinner table.

And so, it all started with a shocking realization: the next superbug could already be on our dinner plates. Antibiotic-resistant bacteria, often associated with undercooked poultry, have now become hardly treatable. When we learned that antimicrobial resistance (AMR) was predicted to cause 49 million deaths by 2050, we knew we had to act. Inspired by the real-world impact of AI in healthcare, we decided to work on sequence alignment for AMR —before it was too late.

🛠What it does and how we built it

The Chicken Dinner Dilemma uses AI-powered sequence alignment to track and predict AMR gene evolution, preventing the rise of untreatable infections by recognizing antibiotic-resistant sequences in bacteria.

To build this model, we worked on 2 different transformers to maximize efficiency and confirm data. We used a FLAN-T5 model as well as a QWEN transformer to acquire optimal scoring. We first split the data 80/20, where 80% was used for training and 20% for evaluation. Then, to use Transformer T5 or QWEN, we first converted raw data into tokenized inputs before feeding the data in string format to the model for training. After that, the model learned to apply the steps to align the sequences. Next, to evaluate the training, we checked for different possible failures such as overfitting or high loss. If the training failed we tweaked parameters to optimize the learning rate and batch size, and we debugged using tensorBoard; not testing again to avoid overfitting.

Otherwise, if the training worked, performance was validated.

In the final step, both models were tested on test data to measure actual performance.

With this method, we optimized the score using machine learning.

🔍Challenges we ran into

Our first challenge was that MSA is an NP-hard problem, meaning the complexity grows exponentially with more sequences. We used ML to optimize and accelerate the process. Then, we had to decide where to insert the gaps (-). We separately used both the Transformer (FLAN-T5) model and Transformer Decoder QWEN to recognize optimal gap placement. Furthermore, to prevent overfitting of training sequences, we stopped the model after its straining loss entered a plateau. Additionally, faster algorithms could sacrifice alignment accuracy while more accurate models (deep learning-based) would be too slow for practical use. In response, we optimized models for parallel computing GPU. Lastly, handling noisy, incomplete, and crowdsourced data was quite a challenge; we applied preprocessing filters to reduce null values.

🔬What we learned

During this challenge, we discovered the evolutionary presence of AI in healthcare and medicine, and how the concept of citizen science could be used in an incredibly effective way to approach real-world problems with minimal resources.

Computationally, we learned how to tokenize data for the T5 and the QWEN transformers, before we learned how to effectively model both with increased accuracy and decreased time consumption such that the machine was trained to recognize patterns and remove irrelevant data.

Working with 2 different transformer models in a limited time taught us to maximize individual skills of the team members as well as communicating properly to divide the workload efficiently while reevaluating and confirming the results altogether.

🔍What's next for The Chicken Dinner Dilemma: Sequence Alignment for AMR

Although finding the AMR sequence rapidly and efficiently is an evolutionary step toward preventing the antimicrobial resistance phenomenon, finding the right treatment for this new family of antigens is also crucial to solving this problem. Using an AI model to record the effectiveness of different substances on AMR bacteria could be key for finding the treatments necessary to prevent a pandemic!

If AI could help detect the resistant gene early enough, we could have the time to step forward and prevent the next AMR outbreak, and eat chicken safely!