Unsupervised protein embeddings outperform hand-crafted sequence and structure features at predicting molecular...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Protein function prediction is a challenging bioinformatics problem.
• Deep supervised models require extensive labeled data, which is scarce for this task.
• Large amounts of unlabeled protein sequences are available.

Purpose of the Study:

• To evaluate an unsupervised deep sequence model for protein molecular function prediction.
• To compare its performance against traditional feature engineering methods.
• To assess the contribution of 3D structure information to the prediction task.

Main Methods:

• Applied a pre-trained deep sequence model in an unsupervised setting to supervised protein function prediction.
• Compared performance against hand-crafted features (one-hot encoding, k-mer counts, etc.).
• Evaluated the impact of incorporating 3D protein structure data.

Main Results:

• The unsupervised deep sequence model achieved competitive performance, outperforming hand-crafted features.
• A simple two-layer perceptron was sufficient for high performance.
• Combining sequence representation with 3D structure did not improve results, suggesting 3D information is implicitly learned.

Conclusions:

• Unsupervised pretraining of deep sequence models is effective for protein function prediction.
• This approach reduces the need for complex prediction architectures.
• 3D structure information may not be necessary when using powerful sequence representations learned through unsupervised pretraining.