PEZy-miner: An artificial intelligence driven approach for the discovery of plastic-degrading enzyme candidates

Based on the article “PEZy-miner: An artificial intelligence driven approach for the discovery of plastic-degrading enzyme candidates,” here is a summary of the key findings and methodologies:

The Challenge Plastic waste presents a severe environmental crisis. While biocatalytic depolymerization using enzymes is a promising and sustainable approach for plastic treatment and recycling, discovering these novel enzymes is notoriously difficult and time-consuming when relying on conventional cultivation-based or omics methods. There is a growing need for computational methods capable of mining massive protein sequence databases to identify potential candidates, yet relying solely on sequence similarity often results in incorrect predictions.

The Solution: PEZy-Miner The researchers developed an innovative machine learning (ML) framework named PEZy-Miner to identify and recommend highly promising plastic-degrading enzyme candidates. This AI-driven tool addresses the limitations of homology-based searches by employing large protein language models (pLMs) alongside confidence and uncertainty estimation.

The PEZy-Miner framework incorporates three key modules:

• Protein Language Model (pLM): Encodes protein sequences into numerical vectors. The study evaluated three state-of-the-art pLMs: ProtBERT, ESM-2, and RoBERTa.

• Classification Module: Predicts the degradability of an enzyme/plastic pair by processing the encoded sequence vector, biophysical features, and one-hot encoded plastic types. Multilayer perceptron (MLP) and prototype classifiers were evaluated.

• Confidence and Uncertainty Estimation: Calculates the reliability of predictions across multiple runs, ranking and filtering out low-confidence/high-uncertainty candidates to ensure a robust top list.

Methodology and Model Evaluation The researchers constructed two datasets: an “experimental dataset” featuring 236 enzyme/plastic pairs across 11 plastic types with known ground-truth labels, and a “homologous dataset” containing putative sequences.

Upon evaluating the six combinations of pLMs and classifiers, the ProtBERT_MLP and ProtBERT_proto models proved superior in generating stable, accurate predictions across multiple random seed runs. The researchers attribute ProtBERT’s success to its massive pre-training dataset (216 million sequences) and large number of parameters (420 million).

Mining and Biological Validation To test PEZy-Miner’s capability to mine unknown databases, researchers blended 60 experimentally verified (“tagged”) enzyme/plastic pairs into a dataset of over 40,000 putative homologous sequences. The ProtBERT-based models successfully identified and concentrated the verified entries by 14 to 33 times within the top 0.5% and 1% of their predicted lists.

By combining the top 1% predictions of both ProtBERT models, PEZy-Miner confidently shortlisted 75 enzyme/plastic pairs, uncovering 27 unique homologous sequences spanning six plastic types (PLA, PCL, PBSA, PBS, PET, and PU). Biological analysis of these shortlisted candidates revealed that many were cutinases originating from plant-infecting fungi (e.g., Fusarium), which aligns with existing biological knowledge regarding enzymes capable of degrading complex, hydrophobic polyesters.

Ultimately, PEZy-Miner serves as an effective computational tool to significantly narrow down massive protein databases into a concentrated list of highly promising, non-obvious plastic-degrading enzyme candidates for future experimental validation.