Membrane protein prediction methods

Abstract

We survey computational approaches that tackle membrane protein structure and function prediction. While describing the main ideas that have led to the development of the most relevant and novel methods, we also discuss pitfalls, provide practical hints and highlight the challenges that remain. The methods covered include: sequence alignment, motif search, functional residue identification, transmembrane segment and protein topology predictions, homology and ab initio modeling. In general, predictions of functional and structural features of membrane proteins are improving, although progress is hampered by the limited amount of high-resolution experimental information available. While predictions of transmembrane segments and protein topology rank among the most accurate methods in computational biology, more attention and effort will be required in the future to ameliorate database search, homology and ab initio modeling.

Fig. 1. Predicting structure and function for a protein experimentally known to be an IMP.

First, the target sequence will be searched against the PDB (6), looking for homologs of known 3D structure. If this search returns at least one good match (template) with a high-resolution structure, it will be possible to apply homology modeling techniques to obtain a model for the target proteins whose resolution will in general depend on the similarity with the template. If the search is either unsuccessful or returns a low-resolution structure, further analysis will instead be needed. Prediction of TM segments, kinks, functional residues and motifs can all help in elucidating the target structural and functional features, either in combination with low-resolution structural information (e.g. from cryo-electron microscopy) or with ab initio modeling techniques. When predicting function it will be useful to search not only the PDB but also other databases, such as SWISS-PROT (147), Interpro (148) and Pfam (50). Note that database searches are performed through alignments methods using either substitution matrices or HMMs.

Fig. 2. Timeline of TMH prediction methods.

Early methods used per-residues hydropathy scales and a window around each residue to produce a smoothed sequence profile. A threshold, usually manually-adjusted, was then introduced to predict TMHs. Modern approaches use machine-learning algorithms such as NNs (illustrated), HMMs or SVMs to predict each residue in e.g. two states, TM or non-TM. In this figure, we schematically show a common NN architecture used for predicting TMHs. As discussed in more detail in the text, there are two NN levels: the first (also called sequence-to-structure NN) produces a per-residue score (analogous to the per-residue value from hydropathy scales); the second (structure-to-structure) takes the output of the first NN and smoothes its values by taking into account the first level predictions for the neighboring residues (somehow similar to the window used in early approaches). Finally, ensemble approaches combine several different methods (in this example, one NN and two HMMs) to produce a consensus prediction.

Fig. 2. Timeline of TMH prediction methods.

Early methods used per-residues hydropathy scales and a window around each residue to produce a smoothed sequence profile. A threshold, usually manually-adjusted, was then introduced to predict TMHs. Modern approaches use machine-learning algorithms such as NNs (illustrated), HMMs or SVMs to predict each residue in e.g. two states, TM or non-TM. In this figure, we schematically show a common NN architecture used for predicting TMHs. As discussed in more detail in the text, there are two NN levels: the first (also called sequence-to-structure NN) produces a per-residue score (analogous to the per-residue value from hydropathy scales); the second (structure-to-structure) takes the output of the first NN and smoothes its values by taking into account the first level predictions for the neighboring residues (somehow similar to the window used in early approaches). Finally, ensemble approaches combine several different methods (in this example, one NN and two HMMs) to produce a consensus prediction.

Fig. 3

Genomics and Structural Genomics of helical IMPs.