The genotype-phenotype landscape of an allosteric protein

Key results

1. ShownA random-mutagenesis library of lacI was phenotyped in one pooled experiment. It spans a wide range of dose-response behaviour, with EC50 varying over more than three orders of magnitude.

   62 472 distinct coding sequences encoding 60 398 amino acid sequences, averaging 4.4 amino acid substitutions per sensor. EC50 ranged from under 1 to over 1000 μmol/L, and G0 and G∞ covered more than a 35-fold range. Synonymous SNPs had no measurable effect (Kolmogorov-Smirnov p = 0.17 to 0.71 for the Hill parameters). · Mapping the genotype-phenotype landscape; Precise engineering of genetic sensors; Comparison of synonymous mutations

2. ShownA recurrent neural network with variational Bayesian uncertainty predicts held-out EC50 and G∞ well and G0 less well. It adds estimates for sensors whose EC50 lies near or above the highest ligand concentration measured.

   Trained on 47 462 quality-filtered variants (80% training split). The recurrent model generally gave higher R² than a feed-forward model for each Hill parameter across ten cross-validation tests. The text gives no R² values. Error and uncertainty grow with distance from wild type. The Hill parameter n was not modelled. · DNN predicts allosteric dose-response; Deep neural network (DNN) modeling; Figure legends

3. ShownSingle-substitution effects were mapped onto the LacI structure, and EC50 effects combine almost log-additively across substitutions. The authors read this as a free-energy balance with additive contributions from many residues.

   Dose-response curves were obtained for 1991 of 2110 SNP-accessible substitutions (964 measured, 1027 DNN-predicted). 38% measurably affect the curve. About 9% and 27% shift EC50 more than 5-fold or 2-fold. Only 0.57% (12 of 2101) of double substitutions deviate more than 2.5-fold from log-additivity. · Impact of substitutions on allostery; Analysis of single-substitution data; Figure legends

4. ShownChoosing library members by target dose-response parameters yielded sensors whose flow-cytometry curves matched the targets. EC50 was hit within 1.25-fold, while accuracy for G0 and G∞ ranged from 1.12-fold to 4.3-fold and depended on growth conditions.

   For 31 sensors with near-wild-type G0 and G∞, the RMS difference in log10(EC50) between sequencing and cytometry was 0.16 (1.45-fold) for EC50 above 1 μmol/L, and 0.097 (1.25-fold) after bias correction. Accuracy for targeting G0 and G∞ ranged from 1.12-fold (near the wild-type G∞ level) up to 4.3-fold (near the wild-type G0 level). About 120 sensors were verified. · Precise engineering of genetic sensors; Dose-response curve verification; Figure legends

5. ShownThe landscape contains inverted sensors (G0 > G∞) and sensors with band-stop or band-pass curves. The sets are genetically diverse. A band-stop phenotype arises from two substitutions that are individually sigmoidal, so such phenotypes can be readily accessible.

   About 230 inverted sensors (0.35% of the library). 43 strongly inverted sensors yielded 10 associated substitutions, yet 51% carry none of them. Band-stop sensors make up nearly 0.2% of the library. 13 band-stop and 3 band-pass sensors were verified by flow cytometry. R195H/G265D gives band-stop while each single substitution is sigmoidal. · Precise engineering of genetic sensors; Novel sensor phenotypes; Identification of high-frequency substitutions…; Figure legends