Use PLM Crystallization Prediction Online
Commercially Available PLM Crystallization Prediction No-Code Web Server
PLM Crystallization Prediction: Sequence-Based Protein Crystallization Propensity with Protein Language Models
PLM Crystallization Prediction is a set of protein language model (PLM) classifiers that predict from sequence alone whether a protein is likely to form diffraction-quality crystals. It comes from the study "Benchmarking protein language models for protein crystallization" by Raghvendra Mall, Rahul Kaushik, Zachary A. Martinez, Matt W. Thomson, Filippo Castiglione and colleagues, published open access in Scientific Reports in 2025. The authors built it on the open-source TRILL platform, and their code is on GitHub.
X-ray crystallography is the usual route to atomic protein structure, but crystallization is costly and often fails. Predicting crystallizability before committing to expression and screening can save time and material. The authors benchmark open PLMs against earlier predictors DeepCrystal, ATTCrys and CLPred, and report that ESM2 models give 3 to 5 percent gains on AUPR, AUC and F1 on independent test sets.
Feature | Detail |
|---|---|
Task | Binary classification: crystallizable versus non-crystallizable |
Training data | PepcDB, 23,333 proteins after filtering (3,846 crystallizable, 19,487 not) |
Embeddings | 14 PLM configurations, mean-pooled, including ESM2, Ankh, ProstT5, ProtT5-XL, SaProt and xTrimoPGLM |
Classifiers | LightGBM and XGBoost; CNN and LSTM on per-residue embeddings |
Length limit | Sequences up to 800 residues |
How PLM-Based Crystallization Prediction Works
Embed the sequence: A pre-trained PLM converts the protein sequence into a mean-pooled embedding without fine-tuning.
Classify: A LightGBM or XGBoost classifier, tuned with 10-fold cross-validation, scores crystallization propensity. The authors found LightGBM generalized better than XGBoost, which overfit.
Compare with deep heads: CNN and LSTM models trained on per-residue embeddings of the top three PLMs were also evaluated.
Generate candidates (optional workflow): The authors fine-tuned ProtGPT2 on crystallizable proteins, generated 3,000 sequences and filtered them through consensus classification, redundancy, secondary-structure, aggregation and homology screens down to 5 candidates.
Benchmark Results on Independent Test Sets
Test set | Best reported model | Result |
|---|---|---|
D2 balanced (1,787 proteins) | ESM2 T30-150M with LightGBM | Accuracy 85.7% (CLPred 85.1%), F1 0.854, MCC 0.715 |
SP_final, SwissProt (237) | ESM2 T36-3B with LightGBM | Accuracy 89%, F1 0.911, MCC 0.769 |
TR_final, TrEMBL (1,012) | ESM2 T30-150M with LightGBM | Accuracy 89.4% (CLPred 85.4%, DeepCrystal 84.1%), F1 0.862, MCC 0.778 |
On D2, ESM2 T30-150M with XGBoost reached AUPR 0.929 and AUC 0.936, which the authors report as 4.3 and 3.3 percent above DeepCrystal. On SP_final, ProstT5 with LightGBM had the best AUC (0.940) and AUPR (0.964) among PLM classifiers, and a ProstT5 LSTM had the highest recall at 0.935. All five generated candidates were predicted crystallizable by the PLM models, whereas DeepCrystal and CLPred each missed one.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with PLM Crystallization Prediction on Tamarind Bio
Construct triage: Rank candidate constructs or targets by predicted crystallizability before expression.
Structural biology planning: Decide which proteins are good candidates for X-ray crystallography.
Design screening: Filter generated or engineered sequences for crystallization potential, as the authors did with ProtGPT2 output.
Method comparison: Use ensemble PLM predictions alongside earlier predictors for a second opinion.
How to Use PLM Crystallization Prediction on Tamarind Bio
Log in: Sign in at tamarind.bio and open the PLM Crystallization Prediction tool.
Provide sequences: Enter or upload protein sequences of up to 800 residues, the length cap used in the study.
Choose the model: Select the PLM and classifier if exposed; ESM2 T30-150M with LightGBM was the strongest on two test sets.
Run: Submit the job.
Download outputs: Retrieve the predicted crystallization propensity for each sequence.
Interpret: Treat scores as a ranking aid rather than a guarantee, and compare across candidates.
Use downstream: Prioritize constructs for crystallization screening.
Things to Keep in Mind
All embeddings are zero-shot because the largest PLMs could not be fine-tuned on the available GPU.
The SP_final test set is small, with 237 proteins.
MSA-based predictors and some other tools were excluded from the comparison.
The five generated candidates are computational predictions without experimental validation in the paper.
Inputs were capped at 800 residues.
Source: Mall R, Kaushik R, Martinez ZA, Thomson MW, Castiglione F, et al. Benchmarking protein language models for protein crystallization. Scientific Reports 15, 2381 (2025). DOI: 10.1038/s41598-025-86519-5. Code: github.com/raghvendra5688/crystallization_benchmark.