DeepTMpred

DeepTMpred predicts the topology of alpha-helical transmembrane proteins (TMPs), identifying transmembrane helices (TMH) and their orientation to support structural characterization.

Key Features:

• ESM embeddings: Uses embeddings from a pre-trained self-supervised ESM language model trained on unlabeled protein sequences to derive latent feature representations for TMPs.
• Convolutional Neural Networks (CNNs): Employs CNNs to capture spatial hierarchies in sequence-derived representations.
• Attentive Neural Networks: Applies attentive neural networks to focus on relevant sequence regions dynamically.
• Conditional Random Fields (CRFs): Uses CRFs to model dependencies between predictions, improving topology consistency.
• Reduced reliance on evolutionary information: Enables topology prediction when evolutionary information is limited or unavailable.
• TMH-level performance: Demonstrates superior predictive performance at the transmembrane helix (TMH) level compared to existing state-of-the-art methods.
• Prediction speed: Generates topology predictions rapidly, producing per-protein results in seconds.

Scientific Applications:

• Structural biology: Provides TMH and orientation annotations to assist structural modeling and interpretation of alpha-helical TMPs.
• Drug design: Informs target selection and ligand accessibility by mapping TMP topology relevant to drug development.
• Membrane proteome annotation: Enables large-scale topology annotation of membrane proteomes, including proteins with scarce evolutionary information.

Methodology:

Derives latent features using a pre-trained self-supervised ESM language model on protein sequences, then applies convolutional neural networks and attentive neural networks with conditional random fields to predict TMP topology.