God_Axis
AETHER
"Let me not die without a fight, without true glory, without some deed that men unborn may hear"
Submitted
May 22, 2026, 9:06 AM
Last Updated
May 22, 2026, 9:06 AM
Project Links
Problem Statement
Over 8 billion tons of plastic waste have accumulated on Earth. Only 9% is recycled. Natural enzymes like PETase can break down PET plastic but are too slow, taking months to years. Recent discoveries from 2024 to 2025 show that engineered variants can work faster, but finding optimal mutations requires screening millions of candidates, which is a costly and slow lab process. The challenge is to build an AI model that runs on consumer laptops with no cloud GPU, to generate and rank novel PETase enzyme sequences with predicted degradation rates greater than 10x wild-type. The solution must produce at least 10 viable enzyme candidates with reasoning for each
Solution & AI Usage
Causal Inference (DoWhy/OLS): Runs offline to identify exact, high-impact amino acid positions, mathematically shrinking the search space. Evolutionary Search (GA + PSO): A custom-built swarm algorithm that intelligently navigates mutation combinations instead of searching blindly. Protein Language Model (ESM-2 150M INT8): Quantized for i3 CPUs. Acts strictly as a fitness judge and physics filter (using attention entropy), not a sequence generator. Reasoning Engine: Translates mathematical outputs and mutations into plain-English biophysical explanation cards.
Full Description
LaptopEnzyme (AETHER) is a decentralized, laptop-powered AI engine designed to accelerate the discovery of plastic-degrading PETase enzymes. Built to run entirely on low-end consumer hardware (like an Intel i3 without a GPU), it democratizes bioremediation research by eliminating the need for expensive cloud compute. Instead of relying on black-box sequence generators, AETHER uses an offline causal inference graph (via DoWhy) to pinpoint the exact amino acid positions that drive degradation. It then deploys a custom-built GA+PSO swarm algorithm to intelligently explore mutation combinations. A locally running, INT8-quantized ESM-2 language model (150M) acts strictly as a judge, scoring each candidate's fitness and filtering for physical stability using attention entropy and pre-cached Rosetta energy deltas. The system outputs 10 viable enzyme candidates with predicted degradation rates exceeding 10x the wild-type baseline, complete with plain-English biochemical reasoning cards. Finally, to ensure transparent and open scientific provenance, every discovery payload is uploaded to IPFS and batched into a single, tamper-proof transaction on the Stellar blockchain.
Tech Stack
Screenshots
5 attached
