Recent mutational analyses of ligand-gated ion channels (LGICs) have demonstrated a plausible site of anesthetic action within their transmembrane domains. Although there is a consensus that the transmembrane domain is formed from four membrane-spanning segments, the secondary structure of these segments is not known. We utilized 10 state-of-the-art bioinformatics techniques to predict the transmembrane topology of the tetrameric regions within six members of the LGIC family that are relevant to anesthetic action. They are the human forms of the GABA alpha 1 receptor, the glycine alpha 1 receptor, the 5HT3 serotonin receptor, the nicotinic AChR alpha 4 and alpha 7 receptors and the Torpedo nAChR alpha 1 receptor. The algorithms utilized were HMMTOP, TMHMM, TMPred, PHDhtm, DAS, TMFinder, SOSUI, TMAP, MEMSAT and TOPPred2. The resulting predictions were superimposed on to a multiple sequence alignment of the six amino acid sequences created using the CLUSTAL W algorithm. There was a clear statistical consensus for the presence of four alpha helices in those regions experimentally thought to span the membrane. The consensus of 10 topology prediction techniques supports the hypothesis that the transmembrane subunits of the LGICs are tetrameric bundles of alpha helices.
The family of proteins known as ligand-gated ion channels (LGICs) is composed of a group of proteins with diverse functional significance but homologous physical structure. Within this family are channels for the conductance of both cations and anions. Their ion conductance properties form the basis for the process of biological information transfer at neuronal synaptic junctions within a variety of biological organisms. The currents generated summate to form many neurological processes within higher organisms. These functions include, but are not limited to, such fundamental processes as memory, pain sensation, movement and awareness.
Most proteins of biological interest are studied after their structure has been determined via X-ray crystallography. Unfortunately, the difficulty in studying the LGICs is that they are transmembrane proteins composed of five large subunits arranged in a pentamer (Unwin, 1993). The pseudosymmetric pentameric arrangement of subunits forms a central ion pore. Each subunit is thought to contribute one alpha helix to the lining of this pore (Unwin, 1993). In addition, each subunit is composed of four transmembrane segments. The goal of the present study was to predict the secondary structure of these four transmembrane segments.
While several groups have performed experiments that shed light on the LGIC secondary structure (Akabas et al., 1994; Blanton et al., 1998; Tamamizu et al., 1999; Methot et al., 2001), these channels are at present impossible to study by more definitive X-ray crystallography. As a first step in molecular modeling of these ion channels, we used 10 state-of-the-art bioinformatics techniques to predict the secondary structure of the transmembrane tetrameric region within six LGICs. Once determined, we will use these assignments of secondary structure to the amino acid sequences of the LGICs subunits in the search for appropriate protein templates for further three-dimensional tertiary structure modeling of these proteins (Bertaccini and Trudell, 2001; Yamakura et al., 2001).
Materials and methods
We obtained the amino acid sequences of six LGICs from the Entrez protein database at the US National Library of Medicine: Torpedo nicotinic acetylcholine receptor alpha (Torpedo nAChRa1) precursor (gi 113076, sp P02710); human neuronal nicotinic acetylcholine receptor alpha 4 (nAChRa4) precursor (gi 1351848, sp P43681); human neuronal nicotinic acetylcholine receptor alpha 7 (nAChRa7) precursor (gi 2506127, sp P36544); human GABA alpha 1 receptor (GABAaRa1) precursor (gi4503859, ref NP_000797.1); human glycine alpha 1 receptor (GlyRa1) precursor (gi 121576, sp P23415); and human 5-hydroxytryptamine receptor (5HT3) precursor (gi 9715820, emb CAA06442.3). The signal portion of each precursor sequence was manually removed to create six sequence files, each of which contained the sequence of its respective mature protein. As a negative control we also obtained the sequence of porin, a transmembrane protein of known beta sheet structure.
Each mature protein sequence was submitted in total to each of 10 transmembrane protein topology prediction algorithms using default input parameters. The techniques used were as follows: DAS (Dense Alignment Surface method) (Cserzo et al., 1997); PHDhtm (dynamic programming) (Rost et al., 1996); SOSUI (transmembrane helix propensity scale) (Hirokawa et al., 1998); TMAP (residue conservation information) (Milpetz et al., 1995); TMFinder (dual prediction by segment hydrophobicity and non-polar phase helicity) (Deber et al., 2001); TMHMM v2.0 (hidden Markov model) (Krogh et al., 2001); TMPred (transmembrane helix propensity scale) (Hofmann, 1993); HMMTOP (hidden Markov model with special architecture) (Tusnady and Simon, 2001a); MEMSAT (log likelihoods) (Jones et al., 1994; McGuffin et al., 2000); and TopPred 2 (hydrophobicity analysis) (Claros and von Heijne, 1994).
The predicted first and last amino acids of each transmembrane segment from each of 10 prediction algorithms run on each of six amino acid sequences were tabulated in an Excel spreadsheet. A multiple sequence alignment of all six LGICs sequences was obtained via the CLUSTAL W algorithm (Thompson et al., 1994). CLUSTAL W was first run on all six mature sequences as a whole and then on each region thought to correspond to a transmembrane segment. There was little or no difference in the multiple sequence alignment of regions thought to contain TM segments 1, 2 and 3 whether performed as a whole or individually. The general terminal region thought to contain TM4 showed significant difference due to the long variable loop region between TMs 3 and 4. It was corrected according to its individual alignment among the six LGIC sequences. The predicted ends for each transmembrane helical segment from each prediction algorithm were then superimposed on to the multiple sequence alignment (Figure 1). The relative centers of each transmembrane segment from each prediction were calculated, averaged across prediction algorithms for a given transmembrane segment within a given sequence and the standard deviations were calculated (Tables I and II). Outliers were defined as those segments that had predicted centers beyond ±2SD from the mean of the center prediction for a given transmembrane segment within a specific LGIC sequence (i.e. across all 10 predictions for the TM1 segment of Torpedo nAChRa1). These outliers were removed from further consensus calculations. We then calculated an average beginning and ending amino acid residue for each TM within each LGIC.
Owing to differences in the number of residues in the receptor domain and in the 3–4 loop region, the numbering of the residues differs throughout the superfamily. Therefore, we had to translate the amino acid numbering system from each LGIC sequence to that of a reference sequence before we could average starting, ending or mid-points of the transmembrane domains. We chose Torpedo nAChR because it is the most studied. The averaged ends of each TM segment for each LGIC sequence were translated into a common numbering scheme based upon their homologous alignment with the Torpedo nAChRa1. Once translated into the same relative numbering scheme, the first and last residues of each corresponding TM segment could be averaged to derive its consensus location across LGICs (black boxes in Figure 1 and Tables I and II).
When the predictions from all 10 transmembrane topology prediction techniques are superimposed on the multiple sequence alignment of all six sequences, there is a clear statistical consensus for the presence of four alpha helices in those regions experimentally thought to span the membrane (Figure 1). After translating the amino acid sequence numbers of each sequence in question into their corresponding numbers in the Torpedo nAChRa1 based on the multiple sequence alignment, the average beginning, ending and length of the individual transmembrane segments were calculated for all six LGICs. They are expressed as follows in the coordinate system of the Torpedo nAChR: TM1 extends from 213 ± 1.2 to 234 ± 1, TM2 extends from 244 ± 1 to 261 ± 2, TM3 extends from 275 ± 1 to 296 ± 1 and TM4 extends from 408 ± 1 to 428 ± 2. This makes their respective lengths 22, 18, 22 and 21 amino acids. The corresponding regions in each of the other amino acid sequences analyzed can be seen in Tables I and II and Figure 1.
While the consensus of the predictions is relatively clear, each technique also made predictions that were outside the consensus. These included the prediction of regions outside the membrane-spanning domain (false positives), not predicting regions thought to be in the membrane-spanning region (false negatives) and the additional production of false-positive results when run against a transmembrane porin of know beta sheet structure. These results are described by technique as follows:
1. DAS: This technique inaccurately predicted small transmembrane helices in the very early parts of the 5HT3, the Torpedo nAChRa1, the nAChRa4 and the nAChRa7 receptors. These regions are all about 180 amino acid residues N-terminal from the consensus membrane-spanning region and they may involve a region of non-membrane alpha helical character. DAS correctly predicted all four transmembrane regions in all six LGICs except for TM3 of nAChRa4. Its prediction for TM2 of nAChRa4 had to be excluded from further consensus calculations since its center was >2SD from the average predicted center for the segment. It did not predict any transmembrane helices in the porin control of known beta sheet structure.
2. PHDhtm: This technique successfully predicted all four transmembrane alpha helices in all six LGICs without suggesting any additional possible regions, except for its prediction of TM2 in GABARa1. This latter prediction had to be excluded from further consensus calculations since its center was >2SD from the average predicted center for the segment. It also did not predict any transmembrane helices in porin.
3. SOSUI: This technique inaccurately predicted small transmembrane helices in the very early parts of the 5HT3, the Torpedo nAChRa1, the nAChRa4 and the nAChRa7 receptors. It correctly predicted all four transmembrane regions in all six LGICs, except for TM2 in nAChRa4, TM2 in GABARa1, TM2 in GlyRa1 and TM4 in GlyRa1. Its prediction for TM3 of nAChRa4 had to be excluded from further consensus calculations since its center was >2SD from the average predicted center for the segment. It did not predict any transmembrane helices in porin.
4. TMAP: This technique successfully predicted all four transmembrane alpha helices in all six LGICs without suggesting any additional possible regions, except for its predictions of TM4 in GABARa1 and TM4 in GlyRa1. These latter predictions had to be excluded from further consensus calculations since their centers were >2SD from the average predicted center for their respective segments. It did not predict any transmembrane helices in porin.
5. TMFinder: This technique inaccurately predicted small transmembrane helices in the very early parts of the Torpedo nAChRa1, the nAChRa4 and the GABARa1 receptors. It correctly predicted all four transmembrane regions in all six LGICs except for TM2 in GlyRa1. In addition, TM1 in 5HT3, TM1 in nAChRa4 and TM1 in nAChRa7 had to be excluded from further consensus calculations since their centers were >2 SD from the average predicted center for their respective segments. Unfortunately, it also predicted two transmembrane alpha helices in the porin control.
6. TMHMM: This technique successfully predicted all four transmembrane alpha helices in all six LGICs without suggesting any additional possible regions, except for missing TM2 in GlyRa1. It also did not predict any transmembrane helices in porin.
7. TMPred: This technique inaccurately predicted a small transmembrane helix in the very early part of GlyRa1. It correctly predicted all four transmembrane regions in all six LGICs except for TM2 in 5HT3 and GlyRa1. Unfortunately, it also predicted three transmembrane alpha helices in the porin control.
8. HMMTOP: This technique inaccurately predicted a small transmembrane helix in the very early part of GlyRa1. It correctly predicted all four transmembrane regions in all six LGICs. However, its prediction for TM3 in GABARa1 had to be excluded from further consensus calculations since its center was >2 SD from the average predicted center for TM3. It also did not predict any transmembrane helices in porin.
9. MEMSAT: This technique successfully predicted all four transmembrane alpha helices in all six LGICs without suggesting any additional possible regions. However, its prediction for TM4 in the Torpedo nAChRa1 had to be excluded from further consensus calculations since its center was >2 SD from the average predicted center for the segment. Unfortunately, it also predicted two low scoring transmembrane alpha helices in porin.
10. TopPred2: This technique successfully predicted all four transmembrane alpha helices in all six LGICs without suggesting any additional possible regions or missing any consensus regions. However, it did predict one transmembrane alpha helix in the porin control.
The validity of the prediction technique
Our prediction that the transmembrane secondary structure of the LGICs is alpha helical is based upon the statistical consensus of multiple transmembrane topology prediction algorithms. Each of these algorithms was derived for the prediction of transmembrane alpha helices. While we see that each technique has certain false-negative and false-positive rates, the value of such predictions lies in the consensus of results. For example, in this study only four of the algorithms predicted alpha helices in porin, a protein of known beta sheet structure. Moreover, among those four false predictions there was no consensus for any particular sequence. The value of finding a consensus among several algorithms was demonstrated by Nilsson et al., who showed that the accuracy of prediction varies strongly with the number of methods that agree (Nilsson et al., 2000). In their study, the topology of nearly half of all Escherichia coli inner membrane proteins can be predicted with ~90% correctness by a simple majority-vote approach using five of the same algorithms presented in our study. Further review of the accuracy of transmembrane topology prediction algorithms may be found in the excellent comparisons made by others (Moller et al., 2001; Tusnady and Simon, 2001b).
A variety of techniques have been used to predict the secondary structure of the transmembrane regions of LGICs (Gorne-Tschelnokow et al., 1994; Gready et al., 1997; Lynch et al., 1997; Ortells et al., 1997; Bertaccini and Trudell, 1999; Barrantes et al., 2000; Yamakura et al., 2001). Recently, Le Novere et al. performed an analysis of the secondary structure of an entire nicotinic acetylcholine receptor (Le Novere et al., 1999). They performed a consensus analysis, somewhat similar to that present in this study, by using the results of several globular protein secondary structure prediction algorithms (PHD, PREDATOR, DSC and NNSSP) combined with a multiple sequence analysis of nAChR, GlyR, 5HT3 and GABAR. In their analysis of the transmembrane segments, they applied several of the same transmembrane topology algorithms employed in this work (DAS, TMPred, Toppred2, SOSUI, PHDhtm), but only as a means of defining the extents of the transmembrane regions. Once these had been set, their subsequent predictions of the transmembrane secondary structure were based on the consensus of the aforementioned globular protein secondary structure prediction algorithms applied to these transmembrane regions. Their results suggest that the extracellular portion of the LGICs is composed primarily of beta sheets, that the transmembrane region is composed of a unique mixed alpha helix–beta sheet combination and that the cytoplasmic domain is composed of alpha helices and coiled protein. Their implementation of multiple secondary structure prediction algorithms to arrive at a consensus prediction is a valuable technique and provides a secondary structure prediction for the extracellular domain that is consistent with that of a recently crystallized homologous acetylcholine binding protein (Brejc et al., 2001). However, their conclusions about the TM domain can be criticized because they applied secondary structure prediction algorithms designed for globular protein to the prediction of transmembrane segments. They mention in their paper, `Each transmembrane segment of the receptor is predicted to fold in a mixed alpha/beta structure. This prediction should be taken with extreme caution, since, as noted above, the programs used were not designed to work on membrane proteins. Prediction methods based on analyses of globular proteins could incorrectly predict strands in helical transmembrane regions'. In the present study, the high degree of consensus is based on transmembrane secondary structure prediction algorithms that were specifically designed for the prediction of transmembrane alpha helices.
Corroboration with experiment
In general, transmembrane domains of proteins have been shown experimentally to be either helical bundles or beta-barrels (Wallin et al., 1997). Despite this, there has been considerable controversy over the predicted secondary structure of the LGICs. Our initial hypothesis was that the common motif of the LGIC superfamily is a pentamer of subunits, with each subunit consisting of four anti-parallel alpha helices (Bertaccini and Trudell, 1999). The general motif of an ion channel composed of five helices belonging to different subunits arranged around a central pore is strongly supported by a series of papers by Unwin and co-workers (Unwin, 1993; Miyazawa et al., 1999). The suggestion that the other three transmembrane segments in each subunit are also alpha helices is much more controversial (Unwin, 1993; Akabas and Gorne-Tschelnokow et al., 1994; Karlin, 1995; Ortells et al., 1997; Le Novere et al., 1999; Williams and Akabas, 1999; Leite et al., 2000).
Recently, Methot et al. conducted an elegant FTIR analysis of the Torpedo nAChR (Methot et al., 2001). They reconstituted Torpedo nAChR in a liposomal preparation and demonstrated a mixed content of alpha helix and beta sheet in the whole receptor. They then exposed the receptor to proteolytic cleavage both external and internal to the liposome to leave only the transmembrane component of the protein embedded in the liposomal membrane. Subsequent FTIR analysis clearly demonstrated a preponderance of alpha helical character. In fact, the only hint of beta sheet character found in this portion was similar to that noted in other proteins that have been shown to be entirely alpha helical in structure by X-ray crystallography.
TM2 is the pore-lining segment and is the segment most studied by experiment. It is believed to be alpha helical based upon several experimental studies: Cryoelectron microscopy was used to demonstrate that electron density in the transmembrane region is consistent with five alpha helices lining the central pore of the nicotinic acetylcholine receptor (Unwin, 1993; Miyazawa et al., 1999). These alpha helices were identified as TM2 in the nicotinic acetylcholine receptor in a series of studies in which specific amino acid residues were mutated to cysteine and labeled with water-soluble reagents (Akabas et al., 1994; Williams and Akabas, 1999).
The crystal structure of a bacterial mechanosensitive ion channel has been published (1msl) (Chang et al., 1998). It is believed to be a possible primordial progenitor for the majority of ion channels in higher organisms. This ion channel contains five alpha helices arranged around a central pore with a right hand supertwist and a funnel shape that is narrowest at the intracellular face (Wilson et al., 2000). This structure supports the alpha helical nature of a central pentamer of TM2s in LGICs. This structure is entirely consistent with our predictions and also those based on electron density in the cryoelectron micrographs of acetylcholine receptors by Unwin and co-workers (Unwin, 1993; Miyazawa et al., 1999). Although early studies predicted a `kink' at the highly conserved leucine residue lining the ion channel (Unwin, 1993), this `kink' was not observed in a more recent NMR study (Opella et al., 1999). Moreover, application of simulated annealing via restrained molecular dynamics (SA/MD) to a model of the nAChR ion channel showed that the `kink' may be achieved by cumulative small distortions of the backbone from canonical alpha-helical geometry, rather than a marked loss of alpha-helical geometry in the vicinity of the conserved leucine (Sansom et al., 1998).
Although there is considerable experimental confirmation that TM2 is an alpha helix, that segment is the least well predicted by the algorithms used in this study. This is most likely because several of the algorithms, whether directly or indirectly, contain hydrophobicity as a fundamental predictor of a transmembrane alpha helix. While TM2 spans the membrane, it is also a helix that lines the ion pore. As a result, it contains a well-defined stripe of hydrophilic residues on the side facing that pore. This rather hydrophilic composition is probably the reason for its lower prediction accuracy relative to the other transmembrane segments.
The other three transmembrane segments identified by hydropathy algorithms have long been assumed also to be alpha helices. Although some studies suggest that this may not be the case, their conclusions can be refuted. For example, exposure of cysteine mutants of TM1 and TM2 of the acetylcholine receptor to a water-soluble probe revealed that TM2 was labeled in a manner consistent with an alpha helix (Akabas et al., 1994), whereas TM1 was labeled in an irregular manner and only near the extracellular surface, indicating incomplete exposure to the aqueous pore region (Akabas and Karlin, 1995). In addition, this exposure changed during channel gating, suggesting that the tertiary structure of the channel changes during gating. In fact, most of the irregularly labeled residues (P211–I215) were either outside of or at the extracellular end of our predicted TM1 region (Y213–Y234). A similar result was obtained with the hydrophobic probe 3-trifluoromethyl-3-(m-[125I]iodophenyl)diazirine in the Torpedo nAChR (Blanton and Cohen, 1994). The probe reacted non-specifically with residues 222, 223, 227 and 228, a pattern of incorporation inconsistent with that expected from either a `face' of an alpha helix or a beta sheet. However, when a central pore is constructed in the shape of a funnel with the small opening at the cytoplasmic surface, the TM2 alpha helices form a tight cylinder at the intracellular surface but are splayed apart toward the extracellular surface. This splay of the helices allows exposure of other TM segments, notably TM1, to the hydrophilic reagents in the pore region (Yamakura et al., 2001). A similar conclusion was reached by Sansom et al. with a model based on distance restraints (Sansom et al., 1998).
Additional uncertainty about the secondary structure of TM1 was provided by proteolytic digestion and mass spectrometric studies of the glycine receptor (Leite et al., 2000). This study found cleavage sites within the putative TM1 and TM3 transmembrane helices. The authors interpreted these cleavage sites as evidence for a mixture of alpha helices and beta sheets within this region (Leite et al., 2000). An alternative explanation is that the very aggressive detergent (Triton X-100) used in isolation of receptors before proteolytic digestion resulted in denaturation of the receptor and exposure of inappropriate cleavage sites. There is similar concern about the FTIR spectrometric study that found only 50% alpha helical content after vigorous removal of the extracellular domain with proteinase K (Gorne-Tschelnokow et al., 1994). It is likely that proteolysis not only removed stabilizing interactions with the receptor domain (Brejc et al., 2001), but also removed inter-segment loops that orient the helices and prevent them from fraying. In contrast to the conclusion of the previous study, recent studies in the nicotinic acetylcholine receptor supported an all-alpha-helical structure in the transmembrane domain (Methot et al., 2001)
The TM3 segment is the subject of much research and controversy. There is strong evidence that the TM2–3 loop is involved in transmission of a signal from agonist binding at the agonist-binding site to the ion channel pore (Ryan et al., 1994; Lynch et al., 1997). If both TM2 and TM3 were alpha helices connected by a short loop, there would be analogy to the established mode of light activation of rhodopsin (Farrens et al., 1996; Palczewski et al., 2000). In addition, it was recently shown that activation by GABA increases the water accessibility of TM3 membrane-spanning residues in GABA receptors (Williams and Akabas, 1999). This result suggests that channel gating involves motion of the transmembrane segments relative to each other. However, initial cryoelectron microscopy by Unwin did not observe clear evidence of alpha helices in the region of TM3 (Unwin, 1993). In the absence of this evidence, he proposed other, more random, structures for this segment. Support for Unwin's suggestion was provided by a molecular modeling study in which it was found to be impossible to place five subunits of strictly anti-parallel alpha helices around a central pore without severe steric overlap. This deficiency can be readily circumvented by incorporating a super twist in packing of the alpha helices (Bertaccini and Trudell, 2001; Yamakura et al., 2001). Our conclusion that TM3 is also an alpha helix is supported by a recent NMR study of a synthetic peptide corresponding to the putative TM3 from Torpedo californica that concluded the segment has an alpha helical structure (Lugovskoy et al., 1998).
We have predicted that TM4 is an alpha helix and the available experimental data also corroborate this. The lipid–protein interface of the Torpedo nicotinic acetylcholine receptor was identified with the hydrophobic probe 3-trifluoromethyl-3-(m-[125I]iodophenyl)diazirine (Blanton and Cohen, 1994). The periodicity of the resulting labeling was consistent with both TM3 and TM4 being alpha helices. The cholesterol binding domain in the nicotinic acetylcholine receptor has been located within TM4 with [125I]azidocholesterol (Corbin et al., 1998). Early studies demonstrated that the TM4 segment could be substituted with homologous sequences without loss of receptor function (Tobimatsu et al., 1987). However, more recent studies have shown that substitution of Cys418 in TM4 by tryptophan altered ion channel function (Tamamizu et al., 1999). The latter studies suggest that although TM4 is thought to be distant from the channel pore, it is important in gating of the ion channel.
In conclusion, the consensus of 10 algorithms specifically designed to predict secondary structure in transmembrane regions provides strong evidence that the four TM segments of LGICs are alpha helices. The poor performance of these algorithms on TM2, the segment with the strongest experimental evidence for alpha helical structure, is readily explained by the partially hydrophobic face that is necessary for a pore-lining segment. It seems appropriate to use X-ray structures of anti-parallel four helix bundles as templates for modeling the tertiary structures of subunits in the superfamily of LGICs (Bertaccini and Trudell, 2001; Yamakura et al., 2001)
The beginning, ending and central residues in the sequences of Torpedo nicotinic AChR alpha-1, human AChR alpha-4 and human AChR alpha-7 were calculated from the TM segments predicted by the secondary structure prediction algorithms DAS, PHDhtm, SOSUI, TMAP, TMFfinder, TMHMM, TMPred, TopPred2, HHMTOP and MEMSAT. The multiple sequence alignment from CLUSTAL W shown in Figure 1 was used to assign residue numbers corresponding to Torpedo AChR alpha-1. Using these assigned numbers, a consensus average beginning, end, center, segment length and the respective SD were calculated, Values shown in bold type were excluded from the averages because their centers were >2SD from the mean central residue. Values are blank where an algorithm did not predict a TM segment in that region.
The beginning, ending and central residues in the sequences of human GlyR alpha-1, GABA alpha-1 and 5HT3R were calculated from the TM segments predicted by the secondary structure prediction algorithms DAS, PHDhtm, SOSUI, TMAP, TMFfinder, TMHMM, TMPred, TopPred2, HHMTOP and MEMSAT. The multiple sequence alignment from CLUSTAL W shown in Figure 1 was used to assign residue numbers corresponding to Torpedo AChR alpha-1. Using these assigned numbers, a consensus average beginning, end, center, segment length and the respective SD were calculated, Values shown in bold type were excluded from the averages because their centers were >2SD from the mean central residue. Values are blank where an algorithm did not predict a TM segment in that region.
This research was supported by NIGMS grant 2 PO1 GM47818 (J.R.T.), VA grant PAIRE BER0000ADO (E.B.) and Stanford University Department of Anesthesia grant 1HEA185 (E.B.).