From Sequence to Structure: pros and cons of AlphaFold-Based Modeling

This talk presents a machine learning workflow for investigating structure–function relationships in membrane proteins, with a focus on pharmacologically relevant families such as SLC transporters, GPCRs, aquaporins, ATPases, and ABC transporters. The main objective is to assess how sequence- and structure-derived descriptors can support fold-class classification, protein–protein interaction prediction, and protein–ligand binding prediction. In the first task, protein embeddings derived from sequence language models, specifically ESM2, were used to classify membrane proteins into folding classes. A feed-forward neural network, optimized through cross-validation and regularized by dropout and early stopping, achieved 94% accuracy on human proteins and 74% accuracy on an external Mus musculus dataset. This indicates strong predictive performance, while also highlighting reduced generalization across species. In the second task, the same strategy was extended to binary classification of interacting and non-interacting protein pairs, with particular emphasis on dimerization and recognition processes involving transmembrane proteins. In the third task, the model was applied to protein–ligand binding prediction, integrating true positives, expected negatives based on cross-ligands, and true negatives derived from protein–protein interaction inhibitors. A central part of the talk critically examines the use of 3D structural descriptors derived from AlphaFold models. Molecular surfaces were represented using MaSIF, which encodes local geometric and physicochemical properties relevant to molecular recognition. However, including MaSIF descriptors extracted from AlphaFold structures reduced performance in the protein–protein and protein–ligand interaction tasks. This suggests that, although AlphaFold provides high-quality structural hypotheses, its models do not necessarily correspond to functionally relevant conformational states. This limitation is particularly important for membrane proteins, especially SLC transporters. AlphaFold models often resemble occluded or metastable intermediate conformations, in which the transport channel is closed on both sides of the membrane. In such states, surface features may become poorly informative for ligand recognition, substrate accessibility, or partner binding. Similar issues can affect GPCRs and other transmembrane families, where function, oligomerization, and binding depend on conformational equilibria, lipid environment, and ligand- or partner-induced rearrangements. The talk concludes that sequence embeddings provide robust descriptors for predictive tasks on membrane proteins, whereas structural descriptors derived from static models must be used with caution. AlphaFold models are valuable structural priors, but they should not be treated as functionally resolved conformations without validation, membrane-aware refinement, or conformational sampling.