Membrane Geometry Reveals Strengths and Limitations of AlphaFold-Predicted Transmembrane Protein Complexes

High-confidence protein complex predictions do not necessarily represent physically or biologically plausible assemblies. This limitation is particularly relevant for transmembrane (TM) proteins, whose interactions are constrained by the lipid bilayer. Here, we systematically evaluated membrane compatibility in the recently expanded AlphaFold Protein Structure Database, analysing 189,469 high-confidence AlphaFold-Multimer TM-TM complexes and additional TM-soluble assemblies. Most predictions exhibited plausible membrane geometry: among 125,657 evaluable complexes with TM proteins possessing more than two TM helices, 95.4% had joint hydrophobic thicknesses within the expected range of 15-35 angstroms. Nevertheless, some complexes displayed substantial inconsistencies in membrane positioning or protomer orientation despite favourable prediction confidence. Interactions between TM proteins, particularly those containing only one or two TM helices, were frequently dominated by extramembrane contacts. Among TM-soluble complexes, 77.4% showed no penetration of the soluble partner into the hydrophobic membrane core. We further investigated whether AlphaFold3 (AF3) could overcome limitations of AlphaFold-Multimer. Although AF3 recovered an experimentally supported SERCA-phospholamban interaction, it failed to reproduce the cholesterol-mediated norepinephrine transporter homodimer. A benchmark of 173 cholesterol-containing membrane protein targets further revealed limited accuracy in predicting cholesterol-binding poses, particularly for structures released after the AF3 training cutoff. Together, these findings demonstrate that membrane-specific geometric constraints provide a valuable complement to conventional confidence metrics and highlight the persistent challenges of predicting assemblies involving lipids.