Fig. 3: A neural network allows to predict binding of transmembrane domains to the EMC.

a Architecture of the EMC binding predictor ipred^EMC. The neural network was trained with a dataset derived from 44 ConMem sequences (Fig. 2 and Supplementary Fig. 3), which were split into 9 subsequences each, from which features from the AAindex database were extracted. The predictor was then applied to design new sequences with different EMC affinities as well as to discover TMDs with high affinity within a list of known EMC clients. Experimental data from these two validation strategies served as a basis for a second iteration of training. This refined version of the predictor was ultimately applied to all TMDs of the human proteome to uncover new EMC chaperone clients. b Prediction and experimental validation of sequences with defined EMC interaction strength (strong/moderate/weak) integrated into a poly-leucine TMD. In a model protein (see Fig. 1d) containing a TMD composed of 24 leucines, flanked by proline and lysine as in ConMem (Fig. 2a), amino acid exchanges were performed at the positions indicated in the schematic based on predictions by ipred^EMC. The chosen weak binders were predicted to be the three most weakly-binding sequences, while the chosen strong binders were the three sequences with the highest predicted binding score. The three moderate binders were chosen from sequences with intermediate predicted binding scores. Binding of the EMC to these constructs was assessed by co-IP experiments in at least five independent replicates and normalized to the poly-leucine sequence (top) (mean  ±  SD, *P-value  <  0.05, **P-value  <  0.01, ***P-value  <  0.001, ****P-value  <  0.0001, two-tailed Student’s t tests). c Application of the predictor to previously reported EMC clients (Supplementary Data 2) to predict binding affinities for the individual transmembrane domains. Schematics of the selected proteins depict the number and orientation of their TMDs and are colored according to the predicted EMC affinity in shades of yellow (low) and red (high). Experimental validation of the predicted binding for a selection of client transmembrane domains is shown as black data points. Binding was assessed in at least five independent replicates and normalized to the highest obtained value. For comparison, the predicted values were also normalized on a 0 to 1 scale and plotted as red crosses. Please note that the topology of SLC4A2 is not well defined, and we show the one consistent with Uniprot annotation here, yet we cannot exclude a different orientation of the TMDs, which would affect our findings. d The ipred^EMC predictor was applied to the whole human membrane proteome to reveal the distribution of EMC binding signatures in all human membrane proteins. For multipass TM proteins, the highest-scoring TMD within the protein was used as the score for a given protein. Predicted affinities were compared between single-pass and multipass membrane proteins, or between N_cyto and N_exo membrane proteins. The number of proteins was normalized to the total number of proteins in their respective subset, and the affinity values were normalized to 0–100. e Analysis of the top 500 hits of the ipred^EMC predictor according to protein class using the PANTHER tool⁶². The shares of different types of transporters were added together, and these classes are labeled in shades of blue.