Abstract
We investigate to what extent yeast complementation assays, which in principle can provide large amounts of training data for machine-learning models, yield quantitative correlations between growth rescue and single-channel recordings. If this were the case, yeast complementation results could be used as surrogate data for machine-learning-based channel design. Therefore, we mutated position L94 at the cavity entry of the model K(+) channel Kcv(PBCV1) to all proteinogenic amino acids. The function of the wild-type channel and its mutants was investigated by reconstituting them in planar lipid bilayers and by their ability to rescue the growth of a yeast strain deficient in K(+) uptake. The single-channel data show a distinct effect of mutations in this critical position on unitary conductance and open probability, with no apparent causal relationship between the two functional parameters. We also found that even conservative amino acid replacements can alter the unitary conductance and/or open probability and that most functional changes show no systematic relationship with the physicochemical nature of the amino acids. This emphasizes that the functional influence of an amino acid on channel function cannot be reduced to a single chemical property. Mutual comparison of single-channel data and yeast complementation results exhibit only a partial correlation between their electrical parameters and their potency of rescuing growth. Hence, complementation data alone are not sufficient for enabling functional channel design; they need to be complemented by additional parameters such as the number of channels in the plasma membrane.