fairfolio
Saved Projects
← Back to Explore

A Novel GenAI Approach To Generate De Novo Terpene Synthase Enzymes By Fine-Tuning ProtGPT2

JSHS · 2025

Hamsini Ramanathan, Roman Bushuiev · Seattle Academy of Arts and Sciences (SAAS), Seattle, WA

Overview

In the vast protein space of 1 billion functional proteins, about 750 million functional proteins remain unexplored today. De novo proteins in this vast space have numerous life -changing applications in personalized medicine, disease diagnosis, biocatalysis, and biotechnology. However, designing de novo proteins using traditional techniques like Directed Evolution is resource intensive and typically takes several months to years. Recent advances in GenAI technologies like protein language models (PLMs) offer a promising solution. We hypothesized that if we fine -tune a PLM like ProtGPT2 with a protein family like Terpene Synthase (TPS) enzymes, the fine -tuned model can generate de novo TPS protein sequences with valid structures. We fine -tuned a distilled tiny version of ProtGPT2 using a dataset of 79,000 TPS sequences mined from UniProt and used the fined-tuned model to generate an initial set of 28,000 new TPS sequences. From this set of 28,000 sequences, we finally filtered down to seven putative de novo TPS enzymes with low perplexity scores, high TPS detection scores, valid 3D structures, and relatively low sequence identity to our training set. The CLEAN model also classified the seven sequences as TPS enzymes. The InterPro model found at least one TPS domain in each of the seven sequences providing additional validation. Six of the seven sequences were classified as druggable proteins by the SPIDER model. Our novel GenAI approach to protein design offers a scalable in -silico method for discoveri ng valid de novo protein candidates in the vast protein search space, significantly accelerating protein design and drug discovery. ForeCAT: Advancing Clear Air Turbulence Prediction for Aviation Safety with Atmospheric Physics Informed Neural Networks and Spatiotemporal Weather Data Aditya Sengupta The Overlake School, Redmond, WA Unexpected turbulence, particularly Clear Air Turbulence (CAT), remains one of aviation’s most serious challenges, causing passenger injuries, operational disruptions, and costing airlines $500 million annually. CAT accounts for 70% of weather-related aviation incidents, with climate change expected to double its frequency in the coming decades. The recent Singapore Airlines turbulence event in May 2024, which resulted in fatalities and injuries, highlights the urgency for improved prediction. Detecting CAT is challenging due to its occurrence in clear air, making conventional algorithms operational in the aviation industry, like Graphical Turbulence Guidance (GTG), less effective due to their reliance on linear, thresholded metrics. ForeCAT offers a breakth rough approach by integrating Artificial Intelligence (AI) with atmospheric physics. Using partial differential equations (PDE) based turbulence diagnostics, ForeCAT’s neural network captures complex turbulence interactions for accurate predictions. ForeCAT achieves 95.5% classification accuracy, three times higher than GTG. Additionally, ForeCAT’s accurately predicts Eddy Dissipation Rate (EDR), the industry-standard turbulence intensity metric. ForeCAT successfully predicted severe turbulence at the Singa pore Airlines event location with 87% confidence, demonstrating its potential to mitigate such incidents. ForeCAT is integrated into LoCATe, an app for real -time CAT prediction, aiding pilots and air traffic controllers in proactive flight path adjustments. To further enhance model accuracy, CATalog, a low-cost turbulence measurement device, is developed to crowdsource and democratize access to turbulence data. By leveraging AI and physics -based diagnostics to significantly advance accurate CAT forecasting, ForeCAT provides a transformative solution for enhancing aviation safety and navigating the increasingly turbulent skies of the future.

Competition history

  • JSHS 2025 Category not listed

Resources

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Source: Junior Science and Humanities Symposium

Save projects to your library

Sign in with Google to keep track of projects you find interesting, organized into folders. An account also raises your daily allowance for “Has this been done?”, and lets you create a key for the MCP server with a much higher limit than anonymous use. Browsing stays public.

Continue with Google