Benchmarking protein language models for protein crystallization.

Area of Science:

• Structural Biology
• Computational Biology
• Machine Learning

Background:

• Protein structure determination is crucial for understanding function, often relying on X-ray crystallography.
• In silico methods, particularly deep learning, are being developed to predict protein crystallization propensity from sequences, addressing experimental limitations.
• Protein Language Models (PLMs) offer a powerful approach for learning sequence-based representations.

Purpose of the Study:

• To benchmark the performance of open protein language models (PLMs) for predicting protein crystallization propensity.
• To identify the most effective PLMs and machine learning classifiers for this task.
• To explore the utility of PLMs in generating novel, potentially crystallizable proteins.

Main Methods:

• Benchmarking various open PLMs (ESM2, Ankh, ProtT5-XL, ProstT5, xTrimoPGLM, SaProt) using LightGBM/XGBoost classifiers on average protein embeddings.
• Comparing PLM-based methods against state-of-the-art sequence-based predictors (DeepCrystal, ATTCrys, CLPred).
• Fine-tuning the ProtGPT2 model for de novo protein generation, followed by multi-stage filtration.

Main Results:

• LightGBM classifiers using ESM2 embeddings (30/36 layers, 150M/3B parameters) demonstrated significant performance gains (up to 3x) across multiple metrics (AUPR, AUC, F1) compared to all other tested models.
• The study identified the most effective PLMs for predicting protein crystallization outcomes.
• Five novel, potentially crystallizable proteins were successfully designed and identified through the fine-tuned ProtGPT2 model and rigorous filtration.

Conclusions:

• Open protein language models, particularly ESM2, show superior performance in predicting protein crystallization propensity.
• The TRILL platform effectively democratizes the use of PLMs for this critical task.
• PLM-based generative approaches hold promise for accelerating the discovery of novel proteins with desired properties, such as crystallizability.