Detecting anomalous proteins using deep representations.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Identifying unique proteins and genes drives biomedical advances.
• Manual inspection of protein properties is becoming unfeasible with large datasets.

Purpose of the Study:

• To develop an automated method for identifying unusual protein properties using anomaly detection.
• To leverage deep learning models for generating protein representations without labeled data.

Main Methods:

• Utilized a state-of-the-art anomaly detection paradigm adapted from computer vision.
• Employed pretrained deep neural network models to create protein language models (pLM).
• Applied pLM anomaly detection to identify unusual functions, phylogenetic families, and sequence segmentation.

Main Results:

• Successfully highlighted human prion-like proteins and distinguished viral from host proteins.
• Identified non-classical ion/metal binding proteins and enzymes, and segmented protein sequences.
• Demonstrated the utility of anomaly scores in 3D folding-related segmentation and identified candidates for rare functionalities.

Conclusions:

• The combination of protein language models and anomaly detection is effective for discovering protein characteristics.
• This novel method shows improved performance over existing baselines across diverse tasks.
• The approach offers a scalable solution for identifying unusual proteins in large-scale biological data.