Water channel pore size determines exclusion properties but not solute selectivity

Aquaporins (AQPs) are a ubiquitous family of transmembrane water channel proteins. A subgroup of AQP water channels also facilitates transmembrane diffusion of small, polar solutes. A constriction within the pore, the aromatic/arginine (ar/R) selectivity filter, is thought to control solute permeability: previous studies on single representative water channel proteins suggest narrow channels conduct water, whilst wider channels permit passage of solutes. To assess this model of selectivity, we used mutagenesis, permeability measurements and in silico comparisons of water-specific as well as glycerol-permeable human AQPs. Our studies show that single amino acid substitutions in the selectivity filters of AQP1, AQP4 and AQP3 differentially affect glycerol and urea permeability in an AQP-specific manner. Comparison between in silico-calculated channel cross-sectional areas and in vitro permeability measurements suggests that selectivity filter cross-sectional area predicts urea but not glycerol permeability. Our data show that substrate discrimination in water channels depends on a complex interplay between the solute, pore size, and polarity, and that using single water channel proteins as representative models has led to an underestimation of this complexity.

. Structure of the AQP ar/R region. (A) Location of ar/R filter and NPA motifs in AQP structure, exemplified by AQP4 (PDB entry 3GD8). Red spheres indicate oxygen atoms of co-crystallised water molecules. (B) Primary sequence of human AQP4. Residues making up the ar/R selectivity filter are highlighted in red and the conserved NPA motifs in blue. Structures of (C) a GLP and (D) a wAQP are exemplified by AQP3 (shown is a homology model to the GlpF structure; PDB code 1FX8) and AQP4 (PDB code 3GD8 26 ). Numbers in brackets refer to the positions defined in the sequence alignment in (E). (E) Sequence alignment of human AQPs and two E. coli AQPs comparing the four regions contributing to the ar/R region. GLPs are highlighted in green. The conserved residues are highlighted in blue; deviations from this are highlighted in red. Panels B-E are reproduced from P. Kitchen PhD thesis 35 .

Results
Mutagenesis of the ar/R region of AQP4, but not AQP1, creates channels that are selective for either urea or glycerol. Previous studies of rat AQP1 showed that increasing the diameter of the rat AQP1 pore through substitution of H180 of the ar/R motif to alanine allows the passage of urea. Increasing the diameter further (through the double substitution F56A/H180A) allows passage of both urea and glycerol, with the urea permeability approximately two-fold higher than the glycerol permeability, whilst the water permeability was unchanged 7 .
To investigate whether substitution of the analogous residues in human AQP4 (F77, H201 and R216) has the same effect, we generated six AQP4 selectivity filter single substitution mutants, F77A, H201A, H201G, H201E, H201F, R216A, and four double substitution mutants, F77A/H201A, F77A/H201G, F77A/R216A and H201A/ R216A, using site-directed mutagenesis. These mutants were transiently transfected into HEK293 cells, chosen for their high transfection efficiency and low intrinsic glycerol/urea permeability. Cell swelling with iso-osmotic glycerol or urea solutions was measured using an adaptation of the plate-reader-based calcein fluorescence quenching method 12,13 to quantify glycerol and urea permeability in live mammalian cells ( Fig. 2A). Surface expression was measured by cell-surface biotinylation (Fig. 2B). Of the 10 mutants that were studied, only AQP4 H201F had reduced protein expression compared to WT AQP4, which was further confirmed by Western blot (data not shown). This may be due to an interruption of protein folding by the introduction of steric clashes in the pore between several bulky hydrophobic amino acid side-chains.
Mutagenesis of H201 to different amino acid residues selectively created channels with a preference for either urea or glycerol but not both ( Fig. 2A,C,D,E). Substitution to alanine (H201A) conferred glycerol but not urea permeability (Fig. 2C,D). This contrasts with the urea permeability reported for the analogous mutation in rat AQP1 (H180A) expressed in Xenopus oocytes, although we could not reproduce this with human AQP1 H180A in mammalian cells (Fig. 2F). Conversely, glycine substitution in AQP4 (H201G) conferred urea permeability but not glycerol permeability (Fig. 2C,D). We have previously demonstrated that these substitutions do not alter AQP4 water permeability 13 .
Substitution of the selectivity filter arginine residue with alanine (R216A) conferred both urea and glycerol permeability to AQP4 (Fig. 2C), in contrast to data for the analogous AQP1 R195V mutant reported by others 7 , which was impermeable to both urea and glycerol. Our data for AQP1 R195A also showed that the channel was impermeable to both urea and glycerol (Fig. 2D). Interestingly, the AQP4 double mutant H201A/R216A lost the effects of either single substitution despite robust surface expression. We previously demonstrated that the R216A mutation modestly increases AQP4 water permeability by approximately 50%, by preventing the arginine sidechain from transiently occluding the pore 13 .
In contrast to the arginine and histidine mutations, the single substitution of the phenyalanine in position 1 with alanine (AQP4 F77A and AQP1 F56A) had no effect on urea or glycerol permeability of either AQP4 or AQP1 (Fig. 2C,D). The AQP4 double mutants F77A/H201A and F77A/H201G were permeable to both glycerol and urea, with a preference for urea and a P u /P g ratio of 4.6 ± 1.1 and 3.7 ± 0.4 respectively. This is similar to the equivalent AQP1 mutant (F56A/H180A), which in our hands had a P u /P g ratio of 3.6 ± 1.4 that agrees qualitatively with the value of 1.7 ± 0.6 reported for the analagous mutation to rat AQP1 7 . The double F77A/R216A mutant was also permeable to both solutes, with a similar P u /P g ratio of 3.3 ± 0.6. Surprisingly, the H201A/R216A mutant was permeable to neither urea nor glycerol despite the permeability of the single H201A and R216A mutants, and despite robust surface expression measured by cell-surface biotinylation (see Fig. 2B), suggesting the the protein is correctly folded and inserted into the plasma membrane.
GLP-mimetic mutants of AQP4 are not solute permeable. Having established that the urea and glycerol permeabilities of human AQP4 can be altered by point mutations in the ar/R-motif, we explored whether substituting the ar/R-motif residues found in AQP4 for those found in GLPs would give a channel with GLP-like behaviour. As described above, the main difference between the ar/R-region in wAQPs and GLPs is the histidine in position 2, which is replaced in GLPs by a small residue such as glycine or alanine (Fig. 1). In addition, GLPs typically have an aromatic residue (usually tyrosine or phenylalanine) in position 3, the side-chain of which packs in front of this small residue (as seen in the GlpF crystal structure). The creation of this structural landscape in AQP4 (H201G/A210Y, H201G/A210F, H201A/A201Y, H201A/A210F) did not create a glycerol-permeable channel when expressed in HEK293 cells (representative calcein fluorescence quenching timeseries in Fig. 3A), or a urea-permeable channel (not shown) despite robust surface expression determined by cell surface biotinylation (Fig. 3B). AQPs 3, 9 and 10 have different relative permeabilities for glycerol and urea. Next, we measured the glycerol and urea permeability of three human GLPs: human AQP3, 9 and 10. All three displayed different, biased selectivity for urea and glycerol (Fig. 4). For AQP3, there is some debate in the literature concerning the urea permeability (see 14 for a detailed discussion), however in our hands, wild-type human AQP3 expressed in HEK293 cells was not measurably urea-permeable (Fig. 4B). AQP10 had a glycerol permeability of 0.94 ± 0.11 when normalised to AQP3 permeability, whereas the permeability of AQP9 to glycerol was almost two-fold higher at 1.8 ± 0.2 (Fig. 4D). In contrast to AQP3, AQP9 and AQP10 were permeable to both glycerol and urea but were oppositely biased, with ratios of urea permeability to glycerol permeability (P u /P g ) of 0.84 ± 0.01 and 1.29 ± 0.03, respectively (Fig. 4E). These were both significantly different from an unbiased permeability ratio of (2019) 9:20369 | https://doi.org/10.1038/s41598-019-56814-z www.nature.com/scientificreports www.nature.com/scientificreports/ 1, determined by one-sample t-tests followed by Bonferroni correction for multiple comparisons (p = 0.02 and p = 0.01, respectively).
The urea and glycerol permeabilities of AQP3 can be altered independently. To investigate whether mutations in the ar/R-motif of AQP3 affect its ability to discriminate between glycerol and urea, we constructed several point mutants and measured glycerol and urea permeability. Substitution of the tyrosine residue in position 3 for alanine (Y212A) conferred measurable urea permeability, whilst concomitantly reducing glycerol www.nature.com/scientificreports www.nature.com/scientificreports/ permeability to approximately half that of the wild-type (Fig. 5A,B). Combining the Y212A mutation with G203H (G203H/Y212A), to mimic the histidine residue present in wAQPs in position 2, produced a channel impermeable to both urea and glycerol (Fig. 3A). The single G203H mutation caused a reduction in protein expression and surface expression comparable to the AQP4 H201F mutant (data not shown). Again, this was suspected to be due to interrupted protein folding caused by steric clashes in the pore.
Urea but not glycerol permeability correlates with pore geometry. To study the effect of mutations on the structure and pore size of the AQP4 at the ar/R selectivity filter region, we generated models of all of our AQP4 mutants by in silico mutagenesis using Swiss PDB Viewer and the crystal structure of human AQP4 (PDB code 3GD8) as a template. The cross-sectional radii and area of the pore at the ar/R-motif was evaluated using the programs HOLE 15 and SYBYL (Tripos Inc, St. Louis, MO, USA) respectively ( Fig. 6 and Table 1). At values greater than ~14 Å 2 , the channel cross-sectional area at the selectivity filter correlated linearly with urea permeability for all AQP4 mutants (R 2 = 0.83, p = 0.006), with the exception of the urea-impermeable H201A/R216A mutant (circled) (Fig. 7A). Using the AQP4 data as a training set, we then attempted to predict the urea permeability of AQP3, 9 and 10 as well as the AQP3 ar/R-motif mutants. For this purpose, homology models of AQP3, AQP9 and AQP10 were created using Swiss PDB Viewer and the crystal structure of E. coli GlpF as the template (PDB code 1FX8). The AQP3 homology model was further used to generate structures of ar/R-motif mutants as described above. As seen in Fig. 7B, the urea permeabilities of 4 of the 5 GLP constructs could be predicted www.nature.com/scientificreports www.nature.com/scientificreports/ using the AQP4-based linear correlation between urea permeability and cross-sectional area. In contrast, neither the AQP4 mutants nor the GLP constructs showed a clear correlation between cross-sectional area and glycerol permeability. Our data thus suggests that urea permeability, but not glycerol permeability, can be explained by the pore geometry.
An unexpected result from our in vitro permeability experiments of AQP4 mutants was that the mutation H201A created a glycerol channel, whilst the mutation H201G created a urea channel. We hypothesized that the H201A mutation, together with F77, forms a "hydrophobic corner", analogous to what is seen in GlpF where the planes of two aromatic residues forms a hydrophobic corner that is suggested to mediate van der Waals contacts with the glycerol alkyl chain (Fig. 8) 16 . In the AQP4 H201G mutant, this corner may be disrupted since the loss of the alanine side-chain could make the V197 backbone carbonyl group solvent accessible and therefore available for hydrogen bonding with water or solute molecules in the pore. This may allow conversion of the H201A glycerol channel to the H201G urea channel, as urea may be able to satisfy the V197 hydrogen bond whereas glycerol cannot. To test this hypothesis in silico, we generated 50 ns molecular dynamics trajectories of H201A, H201G and wild-type AQP4 tetramers using Gromacs. In wild-type AQP4 and the H201A mutant, no hydrogen bonds were observed between the V197 backbone and water molecules in the pore, using the Hbonds plugin for VMD with a 3 Å and 20° hydrogen bond cut-off. In contrast, in the H201G mutant, we found hydrogen bonds between V197 www.nature.com/scientificreports www.nature.com/scientificreports/ and water molecules in all four monomers, with an average occupancy of 24.7 ± 5.4% (using the same 3 Å and 20° cut-off), with the error estimated as the standard deviation over the four monomers (Fig. S1A). For comparison, we measured the hydrogen bond occupancy of water molecules with the two asparagine residues forming the NPA motifs. These were 55.7 ± 7.0% for N97 and 58.2 ± 10.2% for N213. A representative simulation snapshot in which the V197 side-chain hydrogen bond is occupied is shown in Supplementary Fig. 1B.

Discussion
Current understanding of water channel structure and function is informed by data from model family members, primarily AQP1 and GlpF. For large protein families with high sequence, structural and functional homologies, the characteristics of model proteins may often be generalised to the whole family. In the case of the highly-homologous family of water channels, the proposed mechanism for solute selectivity is largely derived from comparing structural details of the water-specific channel AQP1 to those of GlpF from E. coli, which is also permeable to larger solutes such as glycerol and urea. Based on these studies, it is now widely-accepted that any neutral polar solute will pass through the pore of a water channel as long as the selectivity filter is wide enough.
The work presented here clearly demonstrates that this generalization is incorrect. Instead, the location of specific residues in the selectivity filter of mammalian AQPs and the interplay between them affects solute exclusion and specificity, with the pore size being only one of several contributing factors. Using permeability studies in transfected mammalian cells, we show that in human AQP4, both the histidine (position 2) and arginine (position 4) residues of the ar/R selectivity filter region are crucial for neutral solute exclusion while the phenylalanine (position 1) is not. Specifically, AQP4 H201A and H201G formed glycerol and urea selective channels respectively while AQP4 R216A was permeable to both (Fig. 2C). This is in striking contrast to human AQP1 for which the analogous mutations (R195A and H180A/G) were impermeable to both glycerol and urea (Fig. 2D). Similar results were observed in a previous study in Xenopus oocytes in which neither the H180A nor R195V mutations to rat AQP1 generated a glycerol or urea permeable channel 7 . Taken together, these data show that analogous single amino acid substitutions in AQP1 and AQP4 give channels with different exclusion properties. This suggests that the molecular details of how these two water-specific AQPs are able to exclude neutral solutes such as glycerol and urea differ, and does not only depend on the exact composition of the ar/R-motif or the pore size. Consequently, relying on AQP1 as a model of water-selective AQPs has led to an incomplete understanding of how the selectivity filter aids substrate discrimination across the whole family. www.nature.com/scientificreports www.nature.com/scientificreports/ Figure 6. Pore-lining residues of AQP4 mutants at the selectivity filter. Models were constructed by in silico mutagenesis of the AQP4 crystal structure and using Swiss PDBviewer. Cross-sectional areas (bold) were calculated using SYBL. This figure is reproduced from P. Kitchen PhD thesis 35  www.nature.com/scientificreports www.nature.com/scientificreports/ Of particular note is the surprising result that mutating AQP4 H201 to alanine or glycine formed neutral solute channels with opposite glycerol or urea permeabilities. In the crystal structure of the E. coli glycerol facilitator, GlpF, a co-crystallised glycerol molecule had its carbon backbone packed into a corner created by two hydrophobic residues (W48 and F200, Fig. 8). It was suggested that this hydrophobic corner facilitates the preference of GlpF for glycerol over urea because glycerol can have a hydrophobic "face" that can pack into the corner, whereas urea does not. This necessitates energetically unfavourable breaking of urea-protein or urea-water hydrogen bonds when entering the selectivity filter 16 . Indeed, molecular dynamics (MD) simulations of GlpF suggest that the gauche-gauche isomer of glycerol, in which the hydroxyl groups are lined up along one "face" of the molecule leaving an opposing hydrophobic face, is strongly favoured (~80%) in the selectivity filter despite making up <10% of the population in bulk solution 17 . Furthermore, in steered MD, a smaller pulling force was required to make gauche-gauche glycerol permeate the GlpF channel compared to other isomers 18 . We propose that the AQP4 H201A mutation, in combination with F77, mimics this hydrophobic effect, whereas the glycine mutant, lacking the hydrophobic methyl group side-chain of alanine, does not (Fig. 8). This is supported by our MD simulations, which suggest that this hydrophobic corner is further disrupted by the exposure of the V197 backbone carbonyl group to the pore in the H201G mutant (Fig. S1A). This would generate an extra hydrogen-bonding site within the pore that may provide an energetic penalty disfavouring the packing of the glycerol backbone (which cannot satisfy the bond) into this region of the selectivity filter.
The structural features of the H201G mutant ar/R-region are similar to the selectivity filter region of the Urea Transporter family of dedicated urea channels (see Supplementary Fig. 2A,B). In these channels, the selectivity filter is elongated and composed of three regions, an inner (S i ) middle (S m ) and outer (S o ) filter region, all of which share the same features with hydrophobic residues on facing sides and backbone carbonyls that provide opposing hydrogen bonds to permeating urea molecules (19, 20. Interestingly, the S m -region displays strong structural similarities to the conserved NPA-region in AQPs, with two threonine residues in analogous positions as the asparagines residues. In the H201G mutant ar/R-region, the exposure of the backbone carbonyl of V197 breaks the hydrophobic corner associated with glycerol transport, thus creating a selectivity filter with two opposing hydrophobic sides, mimicking those of the urea channels (Fig. 8D). These structural differences provides a plausible explanation for increased urea selectivity of the H201G mutant.
Alternatively, it is also possible that the observed permeability differences are caused by knock-on structural effects of the mutations, leading to proteins with different channel sizes and/or shapes. If this is the case, the www.nature.com/scientificreports www.nature.com/scientificreports/ structural change is likely to be subtle, given that both of the AQP4 H201 mutants still form functional membrane channels that are expressed at the cell surface and that we have previously shown have identical single-channel water permeability to wild-type AQP4 13 .
The importance of the hydrophobic corner for glycerol permeability is further supported by our studies of human AQP3. In our homology model of AQP3, a tyrosine residue, Y212, occupied a similar position to the phenylalanine (F200) in GlpF, producing a similar hydrophobic corner as that seen in the GlpF crystal structure (Fig. 8). As this corner was previously suggested to facilitate glycerol selectivity in GlpF 16 , we mutated the tyrosine residue to alanine. This mutation reduced the glycerol permeability of AQP3 to 51 ± 15% of the wild-type permeability. In contrast, the urea permeability of the channel was increased from zero (or at least undetectably low) to 47 ± 20% of the wild-type glycerol permeability, giving a P u /P g ratio of 0.93 ± 0.16. This provides experimental www.nature.com/scientificreports www.nature.com/scientificreports/ support for the idea of a glycerol-selecting hydrophobic corner in GLPs that preferentially conduct glycerol over urea. Together with the results for the H201A and H201G mutants discussed above, this provides the first experimental evidence that disruption of this hydrophobic corner can have a large effect on the glycerol:urea bias of the AQP pore.
Although the studies above support the role of the hydrophobic corner in selecting for glycerol, mutations in AQP4 aimed to mimic the ar/R-motif of GLPs (including the hydrophobic corner) failed to generate glycerol or urea permeable channels (Fig. 4). This, along with the differences between our AQP1 and AQP4 mutants, suggests that the exact residues of the ar/R region are not the only molecular determinant of solute permeability, but that the surrounding structural context is also important. This is in agreement with crystallographic analyses of AqpZ and GlpF, which suggested that positioning of the selectivity filter residues by the extracellular loops C and E may also contribute to AQP solute permeability/exclusion 10 .
We have, to our knowledge, made the first comparison of glycerol and urea permeability of human GLPs (AQPs 3, 9 and 10) in live mammalian cells, whilst controlling for relative surface expression. For AQP3, there are conflicting reports in the literature about whether it functions as a urea channel 19,20 and it is not clear under what circumstances, if any, urea may permeate the channel. We have discussed this apparent discrepancy in detail in a recent review 14 . In our hands, human AQP3 was glycerol permeable, but was not measurably urea permeable when transiently expressed in HEK293 cells. To exclude artefacts associated with the use of GFP fusion proteins, we repeated measurements of AQP3 urea permeability in HEK293 cells expressing untagged AQP3 with the same result (data not shown). AQP10 had similar glycerol permeability to AQP3, whereas the permeability of AQP9 to glycerol was almost two-fold higher (Fig. 5). Interestingly, AQPs 9 and 10 had opposite biases towards glycerol and urea permeability with both having a P u /P g ratio significantly different from an unbiased score of 1. It may be that these differences in the P u /P g ratio represent a physiological mechanism by which the relative solute permeabilities of a membrane can be fine-tuned by altering the relative expression of different GLPs.
Plotting the urea permeability of the AQP4 mutants against the channel cross-sectional area revealed a linear relationship when the area was above ~14 Å2. This correlated well with the measured urea permeabilities of AQPs 3, 9 and 10, and the urea permeable AQP1 H180A/F56A mutant. Interestingly, the selectivity filter cross-sectional area of our modelled wild-type AQP3 structure is right on the threshold of permeability predicted by this model. We would therefore expect that if AQP3 is urea permeable, the permeability is very low. This may help to explain the conflicting data in the literature on AQP3 as a urea channel. In our hands it was not measurably permeable to urea, but this may be due to differences in timescale between our experiments (~1 min), and alternative approaches to measuring solute permeability, such as radiolabelled solute uptake. In contrast, no clear correlation could be found between glycerol permeability and cross-sectional area. This suggests that the selectivity filter geometrical dimensions are a key determinant for urea permeability but not for glycerol.
During preparation of this manuscript, an X-ray crystal structure of human AQP10 was published 21 . To estimate the reliability of our homology models, we therefore aligned our homology model based on GlpF to this structure (PDB code 6F7H). We find good agreement of the position of selectivity filter residues between our model and the crystal structure (selectivity filter residues heavy atom RMSD = 0.71 Å, see Fig. S1C) and the discrepancy between the selectivity filter radii calculated by HOLE and cross-sectional areas calculated by SYBL are <10% in both cases; we therefore proceeded with our homology model for in silico analysis. The quality of our prediction of the AQP10 channel geometry also lends weight to our homology models of AQP3 and AQP9.
Taken together, our data show that solute permeability of mammalian GLPs and solute exclusion by wAQPs depend on a complex interplay between the exact residues that form the ar/R region, the physical size and chemical properties of the filter created by these residues, and the structural context in which they are situated. For large protein families with high sequence, structural and functional homologies, the characteristics of model proteins may often be generalised to the whole family. Our study demonstrates that in the case of the AQP water channel protein family, this approach has led us to underestimate the complexity of the problem of water channel solute permeability. In order to develop a detailed understanding of water channel solute permeation, data on multiple members of the family should be compared and contrasted.

Methods
DnA constructs and mutagenesis. Human AQP cDNAs cloned into the C-terminal GFP expression vector pDEST47 were generated as previously described 22 . Untagged constructs were made by mutating the first two codons of the AQP-GFP linker peptide to two in-frame stop codons (TAGTGA). Success was confirmed by a ~25 kDa shift of the protein product in SDS-PAGE corresponding to loss of the GFP and lack of fluorescence after transient transfection into HEK293 cells. All site-directed mutagenesis was done using a modified QuikChange protocol, as previously described 23 .
Calcein fluorescence quenching. 10  On the day of the experiments, cells were washed once with growth medium (DMEM + 10% FBS, antibiotic-free) and incubated for 90 minutes at 37 °C with 5 μM calcein-AM (Molecular Probes, Life Technologies, diluted from a 5 mM DMSO stock) and 1 mM probenecid (an organic anion transporter inhibitor used to minimize calcein leakage from cells) in growth medium. After 90 minutes, cells were washed twice with HEPES-buffered growth medium with 1 mM probenecid to remove the DMSO used as a vehicle for calcein. Cells were then returned to the incubator for 10 minutes in HEPES-buffered growth medium (plus probenecid) to equilibrate with the new medium. The plate was moved to a pre-heated (37 °C) Biotek Synergy HT plate reader and calcein fluorescence was measured at 50 ms intervals for 10 s before and 50 s after injection of isotonic urea or glycerol solution, to a cell surface biotinylation. Cell surface proteins from transiently transfected HEK293 cells were biotinylated and AQP surface expression was quantified using a neutravidin-based ELISA, as previously described 24 . For comparison between different AQPs, GFP-tagged constructs were used and GFP was detected in the ELISA using an anti-GFP antibody (Abcam ab6556, diluted 1:5,000). permeability calculations. Normalised fluorescence data was converted to cell volume using a calibration curve derived from fluorescence intensity measurements and cell volume measurements (using a coulter counter) after equilibration with different concentrations of extracellular glycerol, adapted from Fenton et al. 12 . Exponential decay functions of the form C-Ae −kt (with A the amplitude, k time constant and C = 1 + A), were fitted to the normalised cell volume data. Glycerol permeability was then calculated as P gly = k(V 0 /A), with V 0 the initial cell volume (estimated from the coulter counter calibration curve), and A the cell surface area (estimated as 9-fold higher than the surface area of a spherical cell of volume V 0 to account for membrane folding, following Nemeth-Cahalan et al. 25 ). By this method, we find that HEK293 cells expressing AQP3-GFP have a membrane glycerol permeability P gly = 7.6 × 10 −6 cm s −1 , whereas GFP-transfected HEK293 cell have no measurable glycerol permeability (see Supplementary Fig. 2, panel C). For comparison between different aquaporins or different mutants, to account for variation in surface density we normalised P gly to the surface density estimated from cell surface biotinylation ELISAs. All data is represented as surface density normalised P gly (or P urea ), as a proportion of the surface density normalised AQP3 P gly . In silico analysis. AQP4 mutant structures were generated based on the 1.8 Å AQP4 crystal structure 26 .

Statistical analysis.
Simulations were done using Gromacs software, version 4.5.5 27 . The GROMOS53A6 28,29 force field was used and modified to include lipid parameters 30 . An AQP4 tetramer was generated according to the biological assembly entry in the AQP4 PDB file 3GD8 26 . Simulations were done as previously described 13 . Hydrogen bonds were identified using the Hbonds plugin of Visual Molecular Dynamics 31 using 3 Å and 20° cut-offs. Hydrogen bond occupancy was calculated according to these cut-offs at 100 ps intervals along the trajectories and averaged over the four monomers.
A homology model of human AQP3 was generated using E. coli GlpF (PDB 1FX8) as template, using Swiss-Model 32 . In silico mutagenesis was done using Swiss PDB Viewer 33 and the lowest energy amino acid sidechain conformer(s) were chosen in each case.
Channel radii of all structures were calculated using HOLE 34 and channel cross-sectional areas using SYBL (Tripos Inc, St. Louis, MO, USA).

Data availability
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.