chapter_tailanchor.tex

\chapter{Collation and analysis of tail-anchored protein transmembrane helices reveals subcellular variation in flanking charged residue distribution and the transmembrane helix core hydrophobicity}
\chaptermark{Collation and analysis of tail-anchored...}

\sloppy

\section{Abstract}

Tail\--anchored proteins are a functionally diverse group of post\--translationally inserted membrane proteins that have a cytosolic N\--terminal, a transmembrane $\alpha$ helix, and a short C\--terminal tail.
The transmembrane helix and flanking regions of the tail\--anchored protein have been shown to contain sufficient information for subcellular targeting.
Here, we built datasets based on sequence definitions of plausible tail\--anchored proteins using the  UniProt database.
We show that any statistical differences between the hydrophobicity of the transmembrane helix between tail\--anchored proteins belonging to mammalian, plant, or yeast organisms are, if at all present, small.
Yet, the hydrophobicity of transmembrane helices from different subcellular locations is, on average, different between the mitochondria and other membranes along the secretory pathway.
Notably, in the case of mitochondria, this appears to be a difference in the hydrophobic residue preference of alanine instead of leucine or isoleucine and does not seem to be an increase of intramembrane polar residues.
In the endoplasmic reticulum, Golgi, and plasma membrane there are positively charged residues inside the cytoplasm and negatively charged residues outside.
Yet we identify a charge skew reversal of the ``positive\--inside'' and ``negative\--outside'' rules in the mitochondrial tail\--anchored proteins to negative\--inside and positive\--outside.
It is unclear to what extent these differences are the result of membrane environment adaptations or biological features useful for the biogenesis and accurate localisation of the protein.
Furthermore, structural homology modelling of the spontaneously inserting transmembrane helices of PTP1b and cytochrome b5 revealed that tail\--anchored proteins may gain their ability to integrate into the membrane unaided due to a strip of conserved relatively polar residues and strong flanking charge that would allow effective membrane coupling and anchoring.

\section{Introduction}

\gls{ta} proteins are defined by their single carboxy-terminal~\gls{tmh} with a cytosolic facing amino-terminus and are a topologically distinct class of intracellular proteins.
The integration of \gls{ta} proteins into the membrane is post\--translational rather than co-translational; the ribosome is not in complex with the membrane bound insertion machinery.
\gls{ta} proteins are involved in a range of key cellular functions such as translocation (Sec61$\beta$ and Sec61$\gamma$) \cite{Osborne2005} and apoptosis (Bcl-2 apoptotic protein family) \cite{Hockenbery1990}.
Additionally, within the~\gls{ta} class of proteins are a set of vesicle fusion proteins called~\gls{snare} proteins \cite{Ungar2003}, which contain typically hydrophobic \gls{tmh}s \cite{Kalbfleisch2007}.
The idea that \gls{snare} proteins are modular and capable of spontaneous insertion has significant implications for biomedical application in liposome\--based drug delivery.
Furthermore, this could aid future research for testing complex biological molecular networks~\cite{Allen2013, Nordlund2014}.

The \gls{ta} protein's \gls{tmh} is both the anchor and the targeting factor for the \gls{er} \cite{Kutay1993}.
Furthermore, the hydrophobicity of the \gls{tmh} appears to be a determining factor in the delivery pathway that~\gls{ta} proteins use for insertion~\cite{Rabu2008, Rabu2009}, for which there is evidence demonstrating that there are several mechanisms of protein\--membrane integration~\cite{Rabu2009, Johnson2013}(Figure \ref{fig:biogenesis-overview}).
%Alternative mechanisms (Kutay et al., 1995; Nyathi et al., 2013; Chartron et al., 2012)

\begin{sidewaysfigure}
\centering
\includegraphics[width=1\textwidth]{TA_chapter/biogenesis-overview}
        \captionof{figure}[An overview of the biogenesis of tail-anchored proteins.]{\textbf{An overview of the biogenesis of tail-anchored proteins.}

        (i) The intensively studied Get (yeast) or TRC40 (mammalian) pathway can target either for membrane integration or degradation of the \gls{ta} protein.
        Text in brackets indicates mammalian proteins.
        (ii) In yeast, a novel mechanism was identified in which Snd1 binds to the folded \gls{ta} protein and delivers it to the membrane\--bound Snd2 and Snd3 complex.
        (iii) \gls{srp} of the co-translational insertion mechanism has been shown to be able to integrate \gls{ta} proteins. Sec61s role in this is disputed, and it is likely other proteins are involved in membrane partitioning instead.
        (iv) A recently identified insertase, the \gls{emc}, can integrate \gls{ta} proteins with relatively polar \gls{tmh}s into the \gls{er}.
        (v) A handful of \gls{ta} proteins with relatively polar \gls{tmh} regions have been observed spontaneously integrating into the membrane using Hsp40 and Hsc70 as chaperones.
        This system may be employed for mitochondrial localisation.
        (vi) Peroxisomal proteins with an abundance of charge in the tail region are chaperoned by Pex19 into association with the membrane\--bound Pex3.
}

\label{fig:biogenesis-overview}
\end{sidewaysfigure}

\gls{ta} proteins have several pathways for biogenesis in the~\gls{er} membrane.
\gls{ta} proteins were originally thought to be inserted into the membrane via different machinery than the co-translational machinery, but unexpectedly \gls{srp} was found to be a factor for post\--translational targeting confirmed by both cross-linking studies \cite{Abell2004} and an \textit{in vitro} pull-down experiment \cite{Leznicki2010}.
It was shown that \gls{srp} could deliver the \gls{ta} protein to the membrane\--bound \gls{sr} in association with a highly conserved Sec translocon.
Further cross-linking experiments confirmed Sec61 is also involved during \gls{ta} protein membrane insertion \cite{Abell2003}.
Previous studies had shown the Sec61 translocon is not necessary for \gls{ta} protein membrane integration by biochemical reconstitution experiments \cite{Kutay1993} and conditional mutants in yeast \cite{Steel2002, Yabal2003}.
So whilst it is hard to determine if Sec61 can be part of the post\--translational pathway, we can conclude that it is certainly not essential, indeed almost no observable impact on biogenesis is had when it is removed \cite{Kutay1993, Steel2002, Yabal2003}.
Nevertheless, this suggests the possibility of at least one insertion mechanism that is related to the co-translational method of insertion.
Most likely is that \gls{srp} binds to the \gls{ta} protein after it is released from the ribosome and chaperones it until it is close enough to the established post translational machinery \cite{Casson2017} (Figure \ref{fig:biogenesis-overview}iii).

A second redundant system is also known to be involved in \gls{ta} protein biogenesis and is referred to as the TRC40 (also known as Asna1) pathway in mammals (Figure \ref{fig:biogenesis-overview}i).
A conserved homologue was found in \textit{Saccharomyces cerevisiae}, Get3 \cite{Schuldiner2008}, named after the action it has in the GET pathway.
Unlike co-translational insertion, the post\--translational chaperone proteins do not couple with the ribosome, so the \gls{ta} protein must be exposed to the cytosolic environment for at least some time \cite{Guna2018}.
At some point after the \gls{ta} protein emerges from the ribosomal exit tunnel, the \gls{ta} protein \gls{tmh} associates with Sgt2 (SGTA in mammals).
An \textit{in vitro} assay revealed that Sgt2 associates with Get5  \cite{Wang2010} (UBL4A) as part of a dimerised Get4 (TRC35) and Get5 (Ubl4A) complex (two copies of each)\cite{Chang2010, Chang2012, Chartron2010, Chartron2012}.

In mammals, at this point SGTA either associates with preferential TRC40 which targets the \gls{ta} protein for \gls{er} membrane biogenesis or if there are excess \gls{ta} proteins SGTA also associates with Bag6 which targets the \gls{ta} protein for degradation \cite{Shao2017}.
This ``race'' between Bag6 and TRC40 ensures a level of quality control within the system.

Get3 associates first with the Sgt2\--Get4\--Get5 complex via an interaction with the N-terminal of Get4 \cite{Wang2010}.
A dimerised ADP-bound Get3 \cite{Mateja2009, Hu2009, Bozkurt2009, Suloway2009, Yamagata2010} associates with and shields the C\--terminal region of the \gls{ta} protein \cite{Stefanovic2007, Schuldiner2008, Favaloro2008}.
This shielding may be especially important since Get3 is involved in reducing the misfolding of any nascent \gls{ta} proteins, the \gls{tmh} of which would be unviable in the cytosol due to their hydrophobicity \cite{Jonikas2009}.
Fluorescence studies revealed that tagged Get3 appears at both the cytosol and the \gls{er} membrane so apparently shuttles the \gls{ta} protein between the transmembrane complex of Get1 and Get 2 (WRB and CAML in mammalian cells), that contains cytosolic domains that receive the Get3, Get4, Get5, Sgt2 complex \cite{Huh2003, Zalisko2017}.
Yet it is an interesting note that a single molecule fluorescence study revealed that the minimum machinery required for \gls{ta} protein insertion from this system is a Get1 and Get2 heterodimer \cite{Zalisko2017}.
The Get pathway exclusively delivers \gls{ta} proteins to the \gls{er} membrane, and indeed has been recently shown to be responsible for some of the mislocalisation of mitochondrial \gls{ta} proteins during overexpression or signal masking to the \gls{er} \cite{Vitali2018}.
The significance of this is that the Get machinery can recognise and tolerate integration of non\--\gls{er} proteins.
Yet there is also evidence that the deep groove of Get3 \cite{Mariappan2011, Stefer2011} predisposes it to only effectively integrating the more hydrophobic \gls{tmh}s of \gls{ta} proteins \cite{Wang2010, Rao2016}.
As an example, increasing the hydrophobicity of the \gls{tmh} in squalene synthase (an \gls{er} \gls{ta} \gls{tmp} involved in sterol synthesis) in a TRC40 inhibited system reduced the biogenesis of the protein, where the wild\--type was unaffected by TRC40 inhibition \cite{Guna2018a}.
Around a half of \gls{ta} proteins are estimated to not use the TRC40 pathway \cite{Guna2018a}.

Redundancy of the GET/TRC40 pathway and \gls{srp} pathway may be explained in part by a novel \gls{srp} and Get independent pathway.
This pathway utilises the Snd pathway and was discovered in yeast \cite{Aviram2016} (Figure \ref{fig:biogenesis-overview}ii).
Snd1 binds to the \gls{ta} protein after it exits the ribosome and delivers it to the Snd2 and Snd3 membrane\--bound complex which integrates the \gls{ta} protein into the membrane.
So far only the human homologue of Snd2 has been identified (hSnd2) with relatively low sequence identity \cite{Hassdenteufel2017}.
However, it is suspected that functional mammalian homologues exist for Snd1 and Snd3 also, albeit with low sequence similarity.

Even more recently identified as a \gls{ta} protein insertase was the \gls{emc} \cite{Guna2018a} (Figure \ref{fig:biogenesis-overview}iv).
In the interaction study, squalene synthase did not effectively crosslink with any TRC pathway machinery.
Calmodulin (a eukaryotic highly conserved calcium\--sensing globular protein with a flexible symmetrical structure that performs a variety of functions) sufficiently prevented aggregation and acted as a chaperone in \gls{er} microsomes in a chaperone free \textit{E. coli} translation system with purified translation factors.
In this system, calmodulin acted similarly to the mammalian chaperone protein SGTA, and although in native cytosol calmodulin was preferred, SGTA could also be used by the protein.
By contrast VAMP2 (a \gls{ta} protein involved in vesicle binding), which is known to interact with TRC40 as a chaperone \cite{Shao2017}, was unable to insert in this system.
Abolition by knock\--outs of the \gls{emc} components greatly reduced the insertion potential of squalene synthase, along with five mutants thereof, and six other TRC40 independent proteins in \gls{er} membranes from semi-permeabilised cultured cells, but not that of VAMP2.
Insertion of Sec61$\beta$ (a post\--translationally inserted protein ) was partially dependent on both systems, perhaps indicating the midway point between the two \cite{Guna2018a} (Figure \ref{fig:biogenesis-overview}).

Some \gls{ta} proteins, for example BCL\--2 involved in apoptosis, are targeted to the \gls{mom} instead of the \gls{er}, however the mitochondrial mechanism of membrane integration works independently from the \gls{tom} \cite{Setoguchi2006, Kemper2008}.
In the absence of the mitochondrial inner and outer translocons and their associated machinery, insertion machinery and GET/TRC machinery, heat shock proteins Hsp40 and Hsc70 chaperones along with ATP are also sufficient for enough biogenesis of \gls{ta} proteins for viable cell growth \cite{Rabu2008, Rabu2009, Ngosuwan2003, Colombo2009, Kemper2008, Meineke2008, Setoguchi2006, Kemper2008} (Figure \ref{fig:biogenesis-overview}v).
Synaptobrevin, one of the first identified~ \gls{snare} proteins, is capable of spontaneous insertion if the naturally occurring \gls{ta} domain is replaced by the~\gls{tm} that is known to be able to spontaneously insert into the \gls{er}.
One can, therefore, conclude that spontaneous insertion is a property of the \gls{tmh}, not of the rest of the protein \cite{Nordlund2014}.
Molecular dynamics simulations showed that direct insertion \gls{tmh}s thermodynamically mimics the energies of \gls{tmh}s integrated by the translocon \cite{Ulmschneider2014} so in theory, no integration machinery is strictly necessary if the \gls{tmh} can ``correctly'' interact with the membrane interface.
Furthermore, it was revealed that scrambling the \gls{tmh} sequence, but maintaining hydrophobicity, reduced the insertion potential of spontaneously inserting \gls{tmh}s \cite{Brambillasca2006}.
This phenomenon cannot, therefore, be explained entirely by the marginal hydrophobicity of the \gls{tmh}.

The few peroxisomal \gls{ta} proteins first associate with Pex19 which forms a complex with the membrane\--anchored Pex3 protein from which the \gls{ta} protein is integrated into the membrane \cite{Chen2014, Yagita2013, Costello2017}(Figure \ref{fig:biogenesis-overview}vi).

Given a ``choice'', it is speculated that hydrophobicity determines the integration pathway \cite{Costa2018, Rabu2008, Rabu2009}.
Altering the hydrophobicity, at least in the case of the spontaneously inserting PTP1b, also determines the localisation of the \gls{ta} protein to either the mitochondrial membrane or the \gls{er} membrane, or rather a more hydrophobic \gls{ta} protein \gls{tmh} is less likely to localise to the mitochondrial membrane \cite{Fueller2015}.
Broader analysis has shown that hydrophobicity \cite{White1999} stratified by TM tendency score \cite{Zhao2006} can distinguish between the \gls{er} and mitochondrial localised \gls{ta} proteins \cite{Guna2018}.
However, the diversity and known biosynthetic redundancy of these proteins may mean that no single factor applied en masse is able to distinguish the \gls{tmh} recognition factors and investigation into this area is becoming increasingly complex \cite{Guna2018}.

By regenerating a list of likely \gls{tmh}s \cite{Kalbfleisch2007} and using a manually curated list of \gls{ta} proteins \cite{TheUniProtConsortium2014}, this investigation aims to find relationships between biochemical factors and a disposition to a certain insertion mechanism and terminal localisations.
Here, we also present evidence for a conserved polar strip along the spontaneously inserting \gls{ta} protein \gls{tmh}s, which may be the key to the initial interaction of these \gls{tmh}s with the membrane interface.

\section{Methods}

\subsection{Building a list of tail-anchors}
Steps carried out by Kalbfleisch \textit{et al.} (Traffic 8: 1687\-1694) to generate a list of all \gls{ta} proteins in the human proteome ~\cite{Kalbfleisch2007} were recreated using up to date tools and applied to other model representative species.
Whilst their study focused on the human proteome, here we take into account the entire TrEMBL and SwissProt database and then stratify the datasets by the organism at the end of the pipeline.
An overview of the generation of these datasets is shown in Figure \ref{fig:dataset-overview}.

\begin{figure}[!ht]
\centering
\includegraphics[width=1\textwidth]{TA_chapter/dataset-overview}
        \captionof{figure}[The sources, methods, and filters applied to the sequences in the tail\--anchored protein datasets.]{\textbf{The sources, methods, and filters applied to the sequences in the tail\--anchored protein datasets.}
        From top to bottom are the sources of the sequences and the filters and methods applied to each of the datasets of sequences. The database symbol is used to denote when the dataset was used to capture results and is available as supplementary material. For the dataset size and more information, see the methods section text.
}

\label{fig:dataset-overview}
\end{figure}

\subsubsection{SwissProt tail anchored dataset according to filters}
There were 557012 protein records downloaded from SwissProt via UniProt~\cite{TheUniProtConsortium2014} (downloaded 24--04--2018).
106149~\gls{tmh}s (\url{TRANSMEM} annotation) were found between 76953 records (\url{annotation:(type:transmem) AND reviewed:no}).
This keyword is contained in a record according to either experimental evidence~\cite{TheUniProtConsortium2014} or a conservative meta-analysis of~\gls{tmh} prediction using TMHMM~\cite{Krogh2001}, Memsat~\cite{Jones2007}, Phobius~\cite{Kall2004,Kall2007} and the hydrophobic moment plot method of Eisenberg and co-workers~\cite{Eisenberg1984}.
11141 of those records had only a single~\gls{tmh}.
11110 of those~\gls{tmh}s were within the length thresholds of 16 to 30 residues (None of those had the annotation for splice isoforms according to \url{NON_TER} annotation).
5548 of those had had no~\gls{sp} annotation (\url{SIGNAL}).
4332 of those had annotation (based on \url{TOPO_DOM} annotation) that the N terminal was cytoplasmic.
615 of those had the~\gls{tmh} within 25 residues of the C\--terminal, the same threshold used by Kalbfleisch \textit{et al.,}~\cite{Kalbfleisch2007}.
Running CD-Hit 4.5.3 on the WebMGA web-server~\cite{Huang2010, Wu2011} at 90\% identical sequence at 90\% coverage thresholds resulted in 443 representative proteins.
This threshold was chosen as a compromise between avoiding over-representation of a certain protein and maintaining a viable sample size.

From this representative list, 46 were Archaeal, 66 were bacterial, and 320 were eukaryotic and 11 came from dsDNA viruses.
When counting proteomes with greater than 20 records, 49 belonged to the \textit{A. thaliana} proteome, 48 to \textit{Mus musculus} (Mouse), 46 to the \textit{Homo sapiens} proteome, 24 to \textit{S.cerevisiae}. %19 from RAT!

65 were annotated under the mitochondrion location (query \url{locations:(location:"Mitochondrion [SL-0173]")}), 157 in the \gls{pm} (query \url{locations:(location:"Cell membrane [SL-0039]")}, 82 in the Golgi (query \url{locations:(location:"Golgi apparatus [SL-0132]")}), and 98 in the \gls{er} (query \url{locations:(location:"Endoplasmic reticulum [SL-0095]"}).
Only 16 records were found for the peroxisome (query \url{locations:(location:"Endoplasmic reticulum [SL-0204]"}) which is not high enough a sample size for accurate statistical analysis.


\subsubsection{TrEMBL tail anchored dataset according to filters}
111425234 records were stored in the TrEMBL database at the time of download (downloaded 25--04--2018).
22107826 of those contained \url{TRANSMEM} annotation (\url{annotation:(type:transmem) AND reviewed:no}).
18053 of these were single\--pass proteins.
All of these were within the length restrictions of between 16 and 35 residues for the \gls{tmh} region.
17973 of those did not contain a signal sequence when looking for \url{SIGNAL} annotation.
5157 of those contained a cytoplasmically located N terminal according to \url{TOPO_DOM} annotation.
155 records had a~\gls{tmh} within 25 residues of the C\--terminal residue.
When considering which species these records come from, no more than 1 record belonged to any given species.
To avoid representing a well annotated SwissProt record that includes species annotation by a poorly annotated TrEMBL record without species annotation, these TrEMBL records were omitted from the sequence redundancy protocol and further analysis.

\subsubsection{UniProt curated list}
A query for \url{locations:(location:"Single\--pass type IV membrane protein [SL-9908]")} was used in UniProt which returned 2633 UniProtKB IDs; 463 SwissProt results and 2170 TrEMBL results.
Type IV anchors are sometimes split into two topological groups; A (A cytosolic facing N terminal domain) and B (The N terminal is targeted to the lumen), however, the UniProt nomenclature is strictly N terminal being cytosolic.
This manually created list contained some \gls{ta} proteins that did not exactly fit the generally accepted definition of a \gls{ta} protein and were excluded from further analysis.
These could be examples of misannotation in the databases or exceptional \gls{ta} proteins that behave as post\--translationally inserted \gls{ta} proteins despite not matching the exact criteria.
101 exceeded the \gls{ta} length restrictions of 25 residues between the \gls{tmh} and the terminal residue.
8 contained annotation for \url{SIGNAL}, indicating an \gls{sp}, inconsistent with the \gls{ta} protein definition.
20 were multipass proteins.
A full list of which records exceeded these limits and by how much is included in the supplementary files.
Running these records through CD-HIT at 90\% redundancy yielded 956 clusters; 269 SwissProt records and 687 TrEMBL records~\cite{Huang2010, Wu2011}.
No further filters were applied to this list.
Proteomes represented by more than 20 records include \textit{A. thaliana} (53 records), Human (30), \textit{Mus musculus} (30), and \textit{S. cerevisiae} (27). % and 20 to Rat.

426 were annotated under the mitochondrion location (query \url{locations:(location:"Mitochondrion [SL-0173]")}) 47 from SwissProt and 379 automatically assigned in TrEMBL.
397 in the \gls{er} (query \url{locations:(location:"Endoplasmic reticulum [SL-0095]")}), 88 from SwissProt and 308 automatically annotated in TrEMBL.
1 TrEMBL record (UniProt ID A0A1E5RT24) in the \gls{er} set contained an ``X'' residue in the C\--terminal flank and was omitted from the analyses.
Two subcellular location datasets had no automatically ascribed records and only contained manually annotated SwissProt records; 31 in the \gls{pm} (query \url{locations:(location:"Cell membrane [SL-0039]")}, and 83 in the Golgi (query \url{locations:(location:"Golgi apparatus [SL-0132]")}).
There were only 8 \gls{ta} proteins located in the peroxisome (\url{locations:(location:"Golgi apparatus [SL-0204]")}), making them an unsuitable dataset for statistical analysis.

\subsubsection{Remapping the previous dataset}
189 of the 411 proteins from the previous Kalbfleisch \textit{et al.,} 2007 study~\cite{Kalbfleisch2007} were successfully mapped to 222 UniProtKB IDs using the UniProt mapping tools with the RefSeq Protein to UniProtKB option~\cite{TheUniProtConsortium2014}.

\subsubsection{Tail anchor protein chaperone interactors}
As discussed in the introduction text,  there is evidence surrounding the \gls{tmh} biochemical factors that determine which chaperone will interact with a given \gls{ta} protein.
To gain a more quantitative understanding of the relationship between the \gls{ta} protein \gls{tmh} and the potential chaperones, it would be ideal to have a large \gls{ta} dataset stratified by known chaperone interactions.

Known interactor lists from BioGrid from the chaperones were checked against the SwissProt automatically filtered and the UniProt curated \gls{ta} protein datasets.
There were 91 interaction pairs for Hsp40 (Biogrid ID 119699) which mapped to 206 UniProt records.
Hsc70 (Biogrid ID 109544) had 534 interaction pairs 61 of which were mapped to 91 UniProt records.
Hsp40 and Hsc70 both returned 0 record hits after filtering the UniProt records IDs through the UniProt manually curated list and the automatically generated SwissProt list both before redundancy removal.

Snd1 (Biogrid ID 32240) had 237 interaction pairs which mapped to 239 records.
Snd1 returned 15 hits when filtering it through the \gls{ta} datasets.

Sgt2 (BioGrid ID 34410) had 260 BioGrid interactor ids which mapped to 264 UniProt records.
SGTA (BioGrid ID 112347) had 155 interactor ids from BioGrid, 153 of which were mapped to 274 records.
Sgt2 and SGTA returned 14 and 5 hits respectively.

50 BioGrid ids from TRC40 (Biogrid ID 106931) interaction pairs were mapped to 90 UniProt records.
Get3 (BioGrid ID 31962) had 456 Biogrid interactor ids, which mapped to 465 UniProt records.
After filtering those records through the \gls{ta} anchor datasets TRC40 and Get3 returned 7 and 22 hits respectively.

In yeast, Pex19 (BioGrid ID 31994) contained 466 interactor pairs which mapped to 384 UniProt IDs.
The \textit{Homo sapiens} Pex19 (BioGrid ID 111782) contained 230 interaction pairs which successfully mapped to 218 UniProt records.
When the \gls{ta} list filters were applied, 2 \textit{Homo sapiens} and 7 \textit{Saccharomyces cerevisiae} records were found.

For SRP54, ideally the plant version of the protein was required, for which there is more precedent for post\--translational protein interaction and biogenesis into the chloroplasts \cite{Abell2004}, however, of the 3 plant \gls{srp}s available on Biogrid, between them only 10 interactor pairs were available, none of which were common with our \gls{ta} lists.
It is also worth noting that plants have two forms of SRP; one for the chloroplastic and one cytoplasmic which would have to be compared separately.
The human SRP54 (Biogrid ID 112607) had 37 interactors which mapped to 85 UniProt IDs, but again, none of which were in our \gls{ta} lists.
On the other hand, the yeast SRP54 (BioGrid ID 36258) had 270 interactors which mapped to 273 UniProt records.
4 of those were found in our \gls{ta} lists.

\subsection{Calculating hydrophobicity}
Windowed hydrophobicity was calculated using a window length of 5 residues, and half windows were permitted.
Average hydrophobicity takes the total of the raw amino acid hydrophobicity values and divides them by the number of amino acids in the slice.
Unless explicitly stated, values reported in the results are based on the Kyte \& Doolittle scale~\cite{Kyte1982} which is based on the water\---vapour transfer free energy and the interior-exterior distribution of individual amino acids.
%Hydrophobicity values were also validated by the White and Wimley scale~\cite{White1999}, the Hessa scale~\cite{Hessa2005}, and the Eisenberg scale~\cite{Eisenberg1984}.

\subsection{Calculating sequence information entropy}
Information entropy is essentially an estimate of the entropy of a string of characters.
In the context of biology, it can be thought of as an estimation of the non\--randomness of a sequence.
Sequence complexity can be used to analyse DNA sequences~\cite{Pinho2013, Oliver1993, Troyanskaya2002} and is a component of the TMSOC z-score which can predict function beyond anchoring of a \gls{tmh}; an increase in complexity is associated with increased likelihood of function~\cite{Wong2011, Wong2012, Baker2017}.
Here we focus on the analysis of the complexity of a string of characters in protein sequences.

Broadly speaking, the entropy of a string of characters in information theory can be defined as in equation~\ref{simpleentropy2}, and we treat the protein sequence \gls{tmh} as a string with or without its flanking regions.

\begin{equation} \label{simpleentropy2}
H \left( S \right)=-\sum _{ i=1 }^{ n }{ { P }\left({ s }_{ i }\right)\log _{ 2 }{ P } \left({ s }_{ i } \right)}
\end{equation}

Where $H$ is the entropy of a sequence $S$ with values $ \left\{ {s}_{1},\ldots, {s}_{n} \right\} $, and $P \left( s_i \right)$ is the probability ($P$) of a character $i$ through each position in $S$.
This allows us to quantify the average relative information density held within a string of information~\cite{Shannon1948}.

\subsection{Statistics}

The null hypothesis of homogeneity of two distributions was examined with the Kolmogorov Smirnov, the Kruskal-Wallis, and the 2-sampled Student's T-test statistical tests.
These tests were all ran through the Python SciPy stat v0.17 package~\cite{VanderWalt2011}.
To note, the~\gls{ks} test scrutinises for significant maximal absolute differences between distribution curves; the~\gls{kw} test is after skews between distributions and the student T-test statistical test checks the average difference between distributions.

Since the $p$\-‑value is a product of a fraction of test statistics obtained from a permutated set of the samples, it exponentially increases as $N$ increases; the $p$\--value is a strong function of $N$.
We rely on the Bahadur slope ($B$) as a measure of distance between two distributions~\cite{Bahadur1967, Bahadur1971, Sunyaev1998, Baker2017}. A larger Bahadur slope shows a greater difference between the two distributions.

\begin{equation} \label{eq:bahadur2}
B=\frac{|\ln(p~value)|}{N}
\end{equation}

In the heatmaps (Figure \ref{fig:uniprot-heatmap}, Figure \ref{fig:swissprot-heatmap}), the relative percentage normalisation was used rather than a fraction of the absolute value.
This aims to answer the question of ``if we have a certain amino acid, which position is it likely to be in?'' and are able to sensitively identify clusters of skewed preference \cite{Baker2017}.

\begin{equation} \label{eq:independent_normalisation2}
        q_{i,r}=\frac{{100}\cdot{a_{i,r}}}{a_i}
\end{equation}

$a_i$ is the total abundance of residues of a specific amino acid type ($i$) of an aligned set of~\gls{tmh}-containing segments.
Peaks in $q_{i,r}$ as a function of $r$ (the position index) reveal the preferred positions of residues of type $i$.
\gls{tmh}s are oriented according to ``inside'' to ``outside'' from left to right respectively where ``inside'' refers to the cytoplasm.

\subsection{Modelling cytochrome b5 and PTP1b}
In order to assess the 3D arrangement of residues, the HHpred web server was used to query homologues of and model templates for Cytochrome b5 (UniProt accession code P00167) and PTP1b (UniProt accession code P18031)~\cite{Soding2005}.
Homologues were queried using three iterations of HHblitscd against the sequence database version uniprot20\_2016\_02 to generate the query Hidden Markov Model.
For cytochrome b5, a multiple alignment was generated from PDB accession codes and residues positions were approximated for the \gls{tmh} regions based on those of PDB accession codes 5NAO, 5DOQ, 5NAM, and 2MMU covering the \gls{tmh}.
Modeller was run from within the HHPRED server to generate the homology model \cite{Eswar2007, Webb2016}.
Similarly for PTP1b, the \gls{tms} was modelled using a homology model derived from a single sequence alignment of 5NAO based on the TMH$\pm$6 (the length of the C\--terminal tail) residues of PTP1b.

Both the \gls{tmh} regions for that of PTP1b and cytochrome 1b were verified by a consensus of sequence \gls{tmh} predictions (Scampi seq \cite{Bernsel2008}, Phobius \cite{Kall2004}, TMHMM \cite{Krogh2001}, MEMSAT3 \cite{Jones2007}, TMpred \cite{Hofmann1993}, HMMTOP \cite{Tusnady2001},
DAS--TMfilter \cite{Cserzo2004}, MINNOU \cite{Cao2006}, OCTOPUS \cite{Viklund2008}, PRODIV \cite{Viklund2004}, PRO--S \cite{Viklund2004}, S--TMHMM \cite{Viklund2004}, and proteus \cite{Montgomerie2008}).
However, it should be noted that not all these predictions unanimously agreed.
For PTP1b, several methods identified more than 1 \gls{tmh} (HMMTOP, TMPred) whilst Memsat identified a short \gls{tmh} in a completely different region for the \gls{tmh} (35A to 45R).
Besides that, only S--TMHMM and PRO agreed on the exact start and stop positions, and it so happens that these are also the majority consensus positions \cite{Kurowski2003} (409F to 429F).



%Verify3D \cite{Luthy1992}, and PROCHECK \cite{Laskowski1993}.

%>P1;UKNP
%sequence:UKNP:1    :A:137  :A::::
%MAEQSDEAVKYYTLEEIQKHNHSK-STWLILHHKVYDLTKFLEEHPGGEEVLREQAGGDATE--NFEDVGHSTDAREMSKTFIIGELHPDDRPKLNKPPETLITTIDSSSSWWTNWVIPAISAVAVALMYRLYMAED*
%>P1;2M33
%structure:2M33:1   :A:104 :A::Oryctolagus cuniculus::
%MAAQSDKDVKYYTLEEIKKHNHSK-STWLILHHKVYDLTKFLEEHPGGEEVLREQAGGDATE--NFEDVGHSTDARELSKTFIIGELHPDDRSKLSKPMETLITTVD------------------------------*
%>P1;2KEO
%structure:2KEO:21  :A:112 :A::Homo sapiens::
%------EKVTLVRIADLENHNNDG-GFWTVIDGKVYDIKDFQTQSLTENSILAQFAGEDPVV--ALEAALQFEDTRESMHAFCVGQYLEPDQEGVTIPDLG------------------------------------*
%>P1;3X34
%structure:3X34:8   :A:92  :A::Sus scrofa:0.76:
%-------AVKYYTLEEIQKHNNSK-STWLILHHKVYDLTKFLEEHPGGEEVLREQAGGDATE--NFEDVGHSTDARELSKTFIIGELHPDDRSKI------------------------------------------*
%>P1;1MJ4
%structure:1MJ4:3   :A:82  :A::Homo sapiens:1.2:
%-------STHIYTKEEVSSHTSPETGIWVTLGSEVFDVTEFVDLHPGGPSKLMLAAGGPLEPFWALYAVHNQSHVRELLAQYKIGEL--------------------------------------------------*
%>P1;2IBJ
%structure:2IBJ:3   :A:88  :A::Musca domestica:1.55:
%-----SEDVKYFTRAEVAKNNTKD-KNWFIIHNNVYDVTAFLNEHPGGEEVLIEQAGKDATE--HFEDVGHSSDAREMMKQYKVGELVAEERSN-------------------------------------------*
%>P1;5NAO
%structure:5NAO:16  :A:34  :A::Homo sapiens::
%-------------------------------------------------------------------------------------------------------------------VLSVLVVSVVAVLVYKFYF---*
%>P1;5DOQ
%structure:5DOQ:6   :C:28  :C::Geobacillus thermodenitrificans (strain NG80-2):3.05:
%------------------------------------------------------------------------------------------------------------------IMYAPMVVVALSVVAAFWVGLKD*
%>P1;5NAM
%structure:5NAM:16  :A:34  :A::Homo sapiens::
%-------------------------------------------------------------------------------------------------------------------VLSVLVVSVVAVLVYKFYF---*
%>P1;2MMU
%structure:2MMU:30  :A:53  :A::Mycobacterium tuberculosis::
%-------------------------------------------------------------------------------------------------------------SVWFVSLFIGLMLIGLIWLMVFQL----*

APBS as a PyMol plugin was used to map the electrostatic surface of the model \cite{Baker2001}.
Consurf \cite{Ashkenazy2010} was used to map the conservation scores based on 5 iterations of PSI-BLAST \cite{Altschul1997} with an E-value cut-off of 0.0001.
Hydrophobicity was mapped according to the Eisenberg aggregated hydrophicity scale \cite{Eisenberg1984} using a script accessed at \url{https://pymolwiki.org/index.php/Color_h}.

\subsection{Availability of materials}

The scripts and datasets associated with this study can be accessed at \url{https://github.com/JamesABaker/TA-protein-seq}.

\section{Results and discussion}

\subsection{A comparison of up-to-date tail-anchored protein datasets}
Here, we use two sources for \gls{ta} protein datasets.
One dataset is based on a previous method~\cite{Kalbfleisch2007} to obtain \gls{ta} datasets and consists of 9296 \gls{tmh} residues (13279 including up to $\pm$5 flanking residues) from 443 SwissProt entries with 90\% redundancy removal.
Another dataset contains the UniProt curated set of Type IV membrane proteins again with 90\% redundancy removal.
This dataset contains 20528 \gls{tmh} residues (27950 including up to $\pm$5 flanking residues) from 956 UniProt protein records.

% Venn diagram
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{TA_chapter/database-overlap}
        \captionof{figure}[A Venn diagram showing tail-anchored protein UniProt ids present in each of the datasets as well as those present in multiple datasets.]{\textbf{A Venn diagram showing tail-anchored protein UniProt ids present in each of the datasets as well as those present in multiple datasets.}
The number of ids present in redundant versions of
i) the supplementary materials table of a previous study by Kalbfleisch \textit{et al.,} predicting the complete set of human tail-anchored proteins was mapped to UniProtKB and is denoted by ``S1''~\cite{Kalbfleisch2007} is in red,
ii) and in green is the SwissProt dataset filtered according to typical~\gls{ta} features limited to the \textit{Homo sapiens} proteome~\cite{TheUniProtConsortium2014}, and
iii) in blue is the UniProt curated list of~\gls{ta} proteins~\cite{TheUniProtConsortium2014}.
Note that to avoid losing IDs to redundancy reduction this diagram was generated without the use of CD-HIT~\cite{Huang2010, Wu2011}, which is applied in the later statistical analysis.}

\label{fig:tadatasetoverlap}
\end{figure}

In order to get an understanding of the consistency of the datasets, before removing redundant proteins, we compared these two datasets to a remapped set of proteins from a previous 2007 method by Kalbfleisch \textit{et al.,} which aimed to predict \gls{ta} proteins from the NCBI database using predictive tools~\cite{Kalbfleisch2007} (see Figure \ref{fig:dataset-overview} for an overview of these methods).
The S1 dataset was built with an aim to gather \gls{ta} proteins in the human genome from the NCBI.
The greatest source of uncertainty here is that the original S1 list includes 411 records, however, only 222 of these were successfully mapped to the UniProt dataset.
This figure is closer to the 202 proteins from the original S1 list that excluded proteins that were either hypothetical or splice isoforms.
That being said, this mapping step prevents us from directly comparing the entire original S1 dataset.
We compared the up-to-date datasets to S1 to see how many records are shared, how many are now obsolete, and how many are unique.

Figure~\ref{fig:tadatasetoverlap} shows that S1 has 175 record ids of 222 records (78.8\%) which do not share overlap the up-to-date manually curated UniProt dataset~\cite{TheUniProtConsortium2014}.
Of the 170 unique records of that S1 dataset, 4 were manually annotated as not belonging to the \textit{Homo sapiens} proteome, 20 have the C\--terminal as annotated being cytoplasmic, only 125 had \texttt{TRANSMEM} annotation indicating a bona fide \gls{tmh}.
If we apply equivalent filters, only 42 have annotation verifying that they are \gls{ta} proteins.

Equivalent criteria to the original Kalbfleisch \textit{et al.,} 2007 \cite{Kalbfleisch2007} study was applied to the entire SwissProt database and then restricted to the human proteome dataset.
24 of these 47 records (51.1\%) are in the curated UniProt~\gls{ta}  dataset.
21 of the 49 (44.7\%) records from SwissProt filtered \textit{Homo sapiens} dataset can be found in the original S1 list.

The same method applied to an up-to-date dataset overlaps more with a manually curated dataset.
There is also a large degree of what we now believe to be mistakes that occurred in the older prediction tools and datasets, even when using similar methods.
As a trend, this shows that up-to-date datasets improve the reliability of this automated predicted method.
These automated criteria still do not fully align with the manually curated list.
Of 2633 records in the manually curated list, only 2241 have the \texttt{TRANSMEM} annotation.
Furthermore, it is not only the transmembrane annotation itself but also the type of transmembrane protein.
Small integral membrane protein 1 is a blood group antigen (UniProt ID B2RUZ4) that is just one example of a protein we know to be a post\--translationally inserted \gls{ta} protein, and yet in  UniProt it is annotated as a type II, not a type IV, transmembrane protein.
As a result of which it appears in the SwissProt automatically filtered list and not the manually curated list.
In an ideal database, where there are instances of discrepancy, a note on post\--translational or co-translational biogenesis would address this issue.
Ultimately, this points to the idea that datasets are a moving target as they are constantly updated with more accurate information using evermore reliable tools and methods.



\subsection{It is difficult to observe any hydrophobic variation of tail\--anchored protein transmembrane helices from different species}

In single\--pass proteins of eukaryotic species, there are typically various adaptations of the \gls{tmh} to adhere to the membrane constraints of the specific membrane.
For single\--pass proteins, previous studies have observed differences in terms of \gls{tmh} hydrophobicity between \textit{Saccharomyces cerevisiae} and \textit{Homo sapiens} \gls{tmp}s~\cite{Sharpe2010}, or in cress, yeast, bacteria, and \textit{Homo sapien} datasets~\cite{Baker2017}.
We would expect to see a similar trend between the \gls{tmh}s of \gls{ta} proteins from different species.
However, when assuming a zero-difference hypothesis, in these \gls{tmh} \gls{ta} protein datasets we cannot observe any species-level differences between the datasets at this sample size for \gls{tmh} hydrophobicity.

\begin{figure}[!ht]
\centering
\includegraphics[width=1\textwidth]{TA_chapter/species-hydrophobicity}
\captionof{figure}[Average values of species datasets from UniProt manually curated set and SwissProt automatically filtered dataset.]
{\textbf{Average values of species datasets from UniProt manually curated set and SwissProt automatically filtered dataset.}

The average hydrophobicity values from the Kyte \& Doolittle scale~\cite{Kyte1982}.for both the \gls{tmh} and the \gls{tmh}$\pm$5 residues.
Values are shown for both the UniProt manually curated set and the SwissProt filtered set. In the UniProt manually curated set we compare the mammalian set of \gls{ta} proteins (\textit{Homo sapiens} N=30 and \textit{Mus musculus} N=30) to \textit{A. thaliana} (N=57) representing plants and \textit{S. cerevisiae} (N=27) representing yeasts. For the SwissProt filtered set we compare the mammalian set of \gls{ta} proteins (\textit{Homo sapiens} N=46 and \textit{Mus musculus} N=48) to \textit{A. thaliana} (N=49) representing plants  and  \textit{S. cerevisiae} (N=24) representing yeasts.
Error bars are shown at $\pm 1 \sigma$ from the mean of the respective dataset.
}

\label{fig:average_species_hydrophobicity_ta}
\end{figure}

When comparing the average Kyte \& Doolittle~\cite{Kyte1982} hydrophobicity values for the~\gls{tmh}s from \textit{Homo sapiens} and \textit{Mus musculus}, \textit{A. thaliana}, and  \textit{S. cerevisiae}, we can see little difference between the mean values.
All of the mean values lie between 2.3-2.6 when we only consider the \gls{tmh} and at 1.3-1.6 when also considering flanking residues in close proximity to the~\gls{tmh} ($\pm$5 residues) (Figure~\ref{fig:average_species_hydrophobicity_ta}).
Indeed, we see no strong observable statistical differences in hydrophobicity ($P>3.35E-1$ in the SwissProt automatically filtered list Table \ref{table:speciestableswissprotstats}, and $P>2.40E-1$ in the UniProt curated list Table \ref{table:speciestableuniprotstats}).
There are also no consistent trends among the absolute Bahadur slopes; no datasets are greatly different from any other.

\begin{table}[htbp]
\centering
\captionof{table}[Hydrophobicity statistical comparisons between mouse and human, yeast, and plants in the SwissProt Filtered Dataset.]
{\textbf{Hydrophobicity statistical comparisons between mouse and human, yeast, and plants in the SwissProt Filtered Dataset.}
Here, we compare a mammalian set of \gls{ta} proteins (\textit{Homo sapiens} N=46 and \textit{Mus musculus} N=48) to \textit{A. thaliana} (N=49) representing plants  and  \textit{S. cerevisiae} (N=24) representing yeasts.
The hydrophobicity was predicted as the mean average of the values of the sequences of the \gls{tmh}, as well another group including up to $\pm$5 flanking residues, since predicting the boundary of \gls{tmh}s is difficult, according to the Kyte \& Doolittle hydrophobicity scale~\cite{Kyte1982}.
The Test column refers to the statistical score obtained from the test; H statistic for the Kruskal Wallis, the KS statistic for the Kolmogorov Smirnov test, and the t-statistic for the T-test.
$p$ is the p\--value of that statistical score.
$B$ refers to the Bahadur slope, an interpretation of the p\--value that accounts for the sample size powering the test~\cite{Bahadur1967, Bahadur1971}.}
\tiny
    % Table generated by Excel2LaTeX from sheet 'SwissProt filtered species'

    \begin{tabular}{clrrrrrrrrr}
          &       & \multicolumn{3}{c}{Mammal and Plant} & \multicolumn{3}{c}{Mammal and Yeast} & \multicolumn{3}{c}{Plant and Yeast} \\
          &       & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} \\
    \multirow{3}[0]{*}{TMH } &  KW & 0.93  & 3.35E-1 & 7.64E-3 & 0.10  & 7.56E-1 & 2.37E-3 & 0.84  & 3.60E-1 & 1.40E-2 \\
          &  KS & 0.13  & 6.36E-1 & 3.17E-3 & 0.12  & 9.24E-1 & 6.69E-4 & 0.19  & 5.28E-1 & 8.76E-3 \\
          &  T-test & -0.86 & 3.90E-1 & 6.58E-3 & 0.21  & 8.31E-1 & 1.57E-3 & 0.79  & 4.33E-1 & 1.15E-2 \\
    \multirow{3}[0]{*}{TMH and flanks } &  KW & 0.04  & 8.52E-1 & 1.12E-3 & 0.12  & 7.28E-1 & 2.69E-3 & 0.04  & 8.33E-1 & 2.51E-3 \\
          &  KS & 0.11  & 7.72E-1 & 1.81E-3 & 0.13  & 8.79E-1 & 1.09E-3 & 0.11  & 9.80E-1 & 2.81E-4 \\
          &  T-test & -0.22 & 8.23E-1 & 1.37E-3 & -0.38 & 7.04E-1 & 2.97E-3 & -0.19 & 8.50E-1 & 2.22E-3 \\
    \end{tabular}%
                \label{table:speciestableswissprotstats}

\end{table}%

\begin{table}[htbp]
\centering
\captionof{table}[Hydrophobicity statistical comparisons between mouse and human, yeast, and plants in the UniProt Curated Dataset.]
{\textbf{Hydrophobicity statistical comparisons between mouse and human, yeast, and plants in the UniProt Curated Dataset.}
Here, we compare a mammalian set of \gls{ta} proteins (\textit{Homo sapiens} N=30 and \textit{Mus musculus} N=30) to \textit{A. thaliana} (N=53) representing plants  and  \textit{S. cerevisiae} (N=27) representing yeasts.
The hydrophobicity was predicted as the mean average of the values of the sequences of the \gls{tmh}, as well another group including up to $\pm$5 flanking residues, since predicting the boundary of \gls{tmh}s is difficult, according to the Kyte \& Doolittle hydrophobicity scale~\cite{Kyte1982}.
The Test column refers to the statistical score obtained from the test; H statistic for the Kruskal Wallis, the KS statistic for the Kolmogorov Smirnov test, and the t-statistic for the T-test.
$p$ is the p\--value of that statistical score.
$B$ refers to the Bahadur slope, an interpretation of the p\--value that accounts for the sample size powering the test~\cite{Bahadur1967, Bahadur1971}.}
    \tiny
    % Table generated by Excel2LaTeX from sheet 'SwissProt filtered species'

    \begin{tabular}{clrrrrrrrrr}
                &       & \multicolumn{3}{c}{Mammal and Plant} & \multicolumn{3}{c}{Mammal and Yeast} & \multicolumn{3}{c}{Plant and Yeast} \\
                &       & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} \\
    \multirow{3}[0]{*}{TMH} &  KW & 0.71  & 4.01E-01 & 8.09E-03 & 0.03  & 8.72E-01 & 1.57E-03 & 0.57  & 4.48E-01 & 1.00E-02 \\
                &  KS & 0.13  & 6.93E-01 & 3.24E-03 & 0.13  & 9.11E-01 & 1.08E-03 & 0.20  & 4.16E-01 & 1.10E-02 \\
                &  T-test & -0.93 & 3.55E-01 & 9.15E-03 & -0.11 & 9.13E-01 & 1.04E-03 & 0.64  & 5.22E-01 & 8.12E-03 \\
    \multirow{3}[0]{*}{TMH and flanks} &  KW & 1.37  & 2.42E-01 & 1.26E-02 & 0.38  & 5.36E-01 & 7.17E-03 & 0.08  & 7.80E-01 & 3.11E-03 \\
                &  KS & 0.19  & 2.40E-01 & 1.26E-02 & 0.14  & 8.13E-01 & 2.38E-03 & 0.09  & 9.97E-01 & 3.21E-05 \\
                &  T-test & -1.17 & 2.45E-01 & 1.24E-02 & -0.79 & 4.35E-01 & 9.58E-03 & 0.20  & 8.43E-01 & 2.14E-03 \\
    \end{tabular}%
                    \label{table:speciestableuniprotstats}
    \end{table}%

Here, we are dealing with datasets at least an order of magnitude smaller than those broad studies \cite{Sharpe2010, Baker2017} which could explain the absence of the effect.
However, this only goes to show that if there is a biochemically distinct effect in \gls{ta} proteins in terms of hydrophobicity between species, it is indeed weak.


\subsection{There are biochemical differences between tail\--anchored transmembrane helices from different organelles}

Although the species datasets appeared to have no significant differences between them in terms of hydrophobicity, we also investigated the subcellular membranes.
We see clear differences in the biochemistry of the \gls{tmh} (Figure \ref{fig:average_organelle_factors_ta}).

\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{TA_chapter/organelle-averages}
\captionof{figure}[Average sequence\--based biochemical values of organelle datasets from UniProt manually curated set and SwissProt automatically filtered dataset.]
{\textbf{Average sequence\--based biochemical values of organelle datasets from UniProt manually curated set and SwissProt automatically filtered dataset.}

A) The average hydrophobicity values from the Kyte \& Doolittle scale~\cite{Kyte1982}, B) the average information entropy~\cite{Shannon1948} (see methods) for both the \gls{tmh} and the \gls{tmh}$\pm$5 residues.
Values are shown for both the UniProt manually curated set and the SwissProt filtered set.
In the UniProt manually curated set we compare \gls{ta} proteins from the~\gls{er} (N=400) to the Golgi (N=82), the~\gls{pm} (N=37), and the mitochondria (N=401).
For the SwissProt filtered set we compare \gls{ta} proteins from the~\gls{er} (N=98) to the Golgi (N=82), the~\gls{pm} (N=157), and the mitochondria (N=65).
Error bars are shown at $\pm 1 \sigma$ from the mean of the respective dataset.
}

\label{fig:average_organelle_factors_ta}
\end{figure}


In the UniProt manually curated dataset, the Kyte \& Doolittle hydrophobicity scores range from 1.7 in mitochondria to 2.7 in the \gls{pm} (Figure \ref{fig:average_organelle_factors_ta}A).


\begin{table}[htbp]
\centering
\captionof{table}[Statistical comparisons between TMH sequences from organelles in the UniProt Curated Dataset.]
{\textbf{Statistical comparisons between TMH sequences from organelles in the UniProt Curated Dataset.}
Here, we compare an organelle subset from the UniProt curated dataset of \gls{ta} proteins.
We compare \gls{er} (N=397) to Golgi (N=83), \gls{pm} (N=31), and the mitochondria (N=426).
The hydrophobicity was predicted as the mean average of the values of the sequences of the \gls{tmh}, as well another group including up to $\pm$5 flanking residues, since predicting the boundary of \gls{tmh}s is difficult, according to the Kyte \& Doolittle hydrophobicity scale~\cite{Kyte1982}.
The linguistic information entropy was calculated according to the methods section~\cite{Shannon1948}.
The Test column refers to the statistical score obtained from the test; H statistic for the Kruskal Wallis (KW), the KS statistic for the Kolmogorov Smirnov test (KS), and the t-statistic for the student's T-test (T-test).
$p$ is the p\--value of that statistical score.
$B$ refers to the Bahadur slope, an interpretation of the p\--value that accounts for the sample size powering the test~\cite{Bahadur1967, Bahadur1971}.}
    \tiny

    \begin{tabular}{clrrrrrrrrr}
                &       & \multicolumn{3}{c}{ER and Golgi} & \multicolumn{3}{c}{ER and PM} & \multicolumn{3}{c}{ER and mito} \\
                &       & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{P} & \multicolumn{1}{l}{B} \\
    \multirow{3}[0]{*}{Hydrophobicity of TMH } &  KW & 21.83 & 2.98E-06 & 2.66E-02 & 28.53 & 9.21E-08 & 3.80E-02 & 377.02 & 5.54E-84 & 2.34E-01 \\
                &  KS & 0.34  & 1.61E-07 & 3.27E-02 & 0.57  & 5.32E-09 & 4.47E-02 & 0.67  & 4.22E-82 & 2.28E-01 \\
                &  T-test & -6.45 & 2.72E-10 & 4.61E-02 & -8.86 & 2.30E-17 & 8.99E-02 & 23.53 & 6.58E-94 & 2.61E-01 \\
    \multirow{3}[0]{*}{... and flanks} &  KW & 0.21  & 6.48E-01 & 9.07E-04 & 17.53 & 2.83E-05 & 2.46E-02 & 490.46 & 1.13E-108 & 3.03E-01 \\
                &  KS & 0.19  & 1.10E-02 & 9.44E-03 & 0.50  & 4.69E-07 & 3.42E-02 & 0.82  & 5.58E-123 & 3.43E-01 \\
                &  T-test & 0.32  & 7.48E-01 & 6.07E-04 & -4.85 & 1.75E-06 & 3.11E-02 & 34.60 & 2.19E-162 & 4.53E-01 \\
    \multirow{3}[0]{*}{Sequence Entropy of TMH } &  KW & 4.66  & 3.09E-02 & 7.28E-03 & 27.54 & 1.54E-07 & 3.68E-02 & 24.03 & 9.48E-07 & 1.69E-02 \\
                &  KS & 0.24  & 4.78E-04 & 1.60E-02 & 0.46  & 4.20E-06 & 2.91E-02 & 0.18  & 2.10E-06 & 1.59E-02 \\
                &  T-test & 3.22  & 1.37E-03 & 1.38E-02 & 6.42  & 3.71E-10 & 5.10E-02 & -4.55 & 6.28E-06 & 1.46E-02 \\
    \multirow{3}[0]{*}{... and flanks} &  KW & 0.52  & 4.70E-01 & 1.58E-03 & 19.50 & 1.01E-05 & 2.70E-02 & 40.11 & 2.40E-10 & 2.70E-02 \\
                &  KS & 0.13  & 2.06E-01 & 3.31E-03 & 0.41  & 7.97E-05 & 2.22E-02 & 0.23  & 5.53E-10 & 2.60E-02 \\
                &  T-test & 1.08  & 2.82E-01 & 2.65E-03 & 4.47  & 1.00E-05 & 2.70E-02 & -5.84 & 7.51E-09 & 2.28E-02 \\
    \end{tabular}%
                    \label{table:organellesuniprotstats}
    \end{table}%

In the UniProt curated list, there are clear hydrophobic differences between all the organelle \gls{tmh} datasets excluding flanks ($P<2.98E-6$) which as a trend becomes less clear when considering the \gls{tmh}$\pm$5 flanking residues except for mitochondria which increases in significance when considering the flanks also (Table~\ref{table:organellesuniprotstats}).
The \gls{er} and mitochondrial tests are very significant ($P<4.22E-82$).
Consistently the Bahadur slope is at least an order of magnitude greater in the \gls{er} and mitochondrially located protein comparison than for the other considerations, so these differences cannot be accounted for by the larger sample size.
This gap in hydrophobicity appears to be due to a trend of the \gls{er}, \gls{pm}, and Golgi using isoleucine, valine, and leucine as their most common \gls{tmh} residues, whereas in the case mitochondrially located \gls{ta} proteins, the most common residue type is alanine in the UniProt manually curated dataset (16.3\% of total residues) followed by valine (12\% total residues)(Figure \ref{fig:uniprot-heatmap}).
Similarly, alanine is the second most common residue in mitochondrial located \gls{ta} proteins from the SwissProt automatically generated dataset at 11.9\% of the total residues after leucine which is 13.4\% of the total residues(Figure \ref{fig:swissprot-heatmap}).

\begin{figure}
\centering
\includegraphics[width=1\textwidth]{TA_chapter/UniProt-heatmaps}
\captionof{figure}[The normalised skews of each amino acids from tail\--anchored proteins grouped by localisation from the UniProt manually curated dataset.]
{\textbf{The normalised skews of each amino acids from tail\--anchored proteins grouped by localisation from the UniProt manually curated dataset.}

The residue position aligned with the centre of the TMH is on the horizontal axis, and the residue type is on the vertical axis.
Amino acid types are listed in order of decreasing hydrophobicity according to the Kyte and Doolittle scale \cite{Kyte1982}.
Flank lengths were restricted to $\pm$5 residues.
The edge residues from proteins with flank lengths and \gls{tmh} lengths that exceeded the plotted 31 residues were still included in the normalisation calculations despite not being plotted.
The colour scale represents the relative percentage of a particular amino acid and is shown with dark blue as 0, white as the 50th percentile value of the entire heatmap, and dark red as the highest percentage on the heat map.
The panels are constructed from TA proteins derived from the SwissProt automatic method with redundancy removal applied detailed in the methods section.
The datasets were further separated by subcellular locations: (a) the \gls{er}, (b) the Golgi, (c) the cell membrane, (d) the mitochondria.
These datasets are more thoroughly outlined in the methods section.
}

\label{fig:uniprot-heatmap}
\end{figure}

Analysis from 16 \gls{ta} proteins with known subcellular locations showed that both the C\--terminal tail charge and hydrophobicity are determinants of the terminal destination to the \gls{er}, mitochondria, and the peroxisome intracellular subcellular locations \cite{Costello2017}.
They found that less hydrophobicity and more charge in the ``tail'' determined the \gls{ta} protein for the mitochondria rather than the \gls{er}.
This corroborates what we see in terms of hydrophobicity (Figure \ref{fig:average_organelle_factors_ta}A).
When we consider charge difference between organelles on larger datasets, we see trends that reinforce this idea, however, rather than net charge, we see differences in charge distribution along the \gls{tmh} and the neighbouring flanks.
In the SwissProt automatically filtered dataset, in the \gls{er} 9.4\% of the residues are positively charged, and 2.5\% are negatively charged.
Most of the positively charged residues cluster following the ``positive\--inside'' rule between positions -15 and -10 for R and K, but so do the negatively charged residues D and E, effectively reducing this local charge by 2.5\%.
In mitochondria, we find that the proportion of charge is similar (10.4\% R and K, 3.3\% D and E) however the negatively charged residues cluster on the N flank (-15 to -8) and the positively charged residues cluster more strongly on the outside flank (positions 9 to 15) (Figure \ref{fig:swissprot-heatmap}).

In the UniProt manually curated \gls{er} set, 6.6\% of residues are positively charged and 3.3\% of residues are negatively charged.
K clusters strongly on the inside flank as expected, yet R clusters strongly between positions 7 to 15 and rather weakly at the inside flank (positions -15 to -10) (Figure \ref{fig:uniprot-heatmap}).
Similarly to the SwissProt sets, D prefers the inside flank but is tolerated in the outside flank.
The more abundant E residues behave very unusually and cluster at positions 5-10.
Generally, charged residues are suppressed in the \gls{tmh} core \cite{Sharpe2010, Baeza-Delgado2013}, especially in anchoring \gls{tmh}s \cite{Baker2017}.
It is unclear why this is observed, yet, altogether the 313 glutamic acid residues and 397 arginine residues that appear unusually deep in the \gls{tmh} core may be to an extent neutralising one another in the compacted \gls{tmh} arrangement, but are ultimately not that abundant compared to the total number of residues in this organelle dataset (11351 total residues).

In mitochondria, 1413 positively charged residues (11\% of the total residues in the mitochondrial dataset) were preferentially located at the outside flank and somewhat into the core (positions 6 to 15) than the expected ``inside'' flank (positions -15 to -5).
The 221 negatively charged residues (1.7\%) unusually cluster at the inside.
This results in a strong net positive\--outside charge signal since there are more positively charged residues on the outer flank un\--countered by the negatively charged residues, which are skewed with a preference for the inside flank.

\begin{figure}
\centering
\includegraphics[width=1\textwidth]{TA_chapter/SwissProt-heatmaps}
\captionof{figure}[The normalised skews of each amino acids from tail\--anchored proteins grouped by localisation from the SwissProt automatically filtered dataset.]
{\textbf{The normalised skews of each amino acids from tail\--anchored proteins grouped by localisation from the SwissProt automatically filtered dataset}

Similarly to figure \ref{fig:uniprot-heatmap}, the residue position aligned with the centre of the TMH is on the horizontal axis, and the residue type is on the vertical axis.
Amino acid types are listed in order of decreasing hydrophobicity according to the Kyte and Doolittle scale \cite{Kyte1982}.
Flank lengths were restricted to $\pm$5 residues.
The edge residues from proteins with flank lengths and \gls{tmh} lengths that exceeded the plotted 31 residues were still included in the normalisation calculations despite not being plotted.
The colour scale represents the relative percentage of a particular amino acid and is shown with dark blue as 0, white as the 50th percentile value of the entire heatmap, and dark red as the highest percentage on the heat map.
The panels are constructed from TA proteins derived from the SwissProt automatic method with redundancy removal applied detailed in the methods section.
The datasets were further separated by subcellular locations: (a) the \gls{er}, (b) the Golgi, (c) the cell membrane, (d) the mitochondria.
These datasets are more thoroughly outlined in the methods section.
}

\label{fig:swissprot-heatmap}
\end{figure}

Information entropy has been known to identify cryptic function in \gls{tmh}s when considered along with hydrophobicity \cite{Wong2011, Wong2012}.
In terms of information entropy, there is a marked decrease in entropy in the \gls{pm} subset (mean entropy = 3.15 in the \gls{tmh}, 2.67 including $\pm$5 flanking residues) from the UniProt curated dataset compared to the other organelle datasets (entropy $>$ 3.29 and $>$ 2.85 including the flanks).
However, this stark difference in \gls{tmh}s from \gls{pm}\--bound \gls{ta} proteins and the other organelle datasets cannot be observed in the SwissProt set (Figure~\ref{fig:average_organelle_factors_ta}).

No clear significant differences can be observed for the information entropy ($P>6.33E-2$).
This is unsurprising given that the hydrophobic nature of the \gls{tmh}s demands that certain residues must be over-represented, which lowers the information entropy.
In this case, we have a highly hydrophobic set, the \gls{pm} UniProt set, which likely contains a higher proportion of the most hydrophobic residues.
As a trend, the information entropy mirrors the hydrophobicity albeit with less range between dataset means (2.67-3.15 in the \gls{tmh} for information entropy, 1.72-2.74 for hydrophobicity)(Figure~\ref{fig:average_organelle_factors_ta}).


    \begin{table}[htbp]
    \centering
    \captionof{table}[Statistical comparisons between transmembrane helix sequences from organelles in the SwissProt Filtered Dataset.]
    {\textbf{Statistical comparisons between transmembrane helix sequences from organelles in the SwissProt Filtered Dataset.}
    Here, we compare organelle subsets from the SwissProt automatically filtered dataset of \gls{ta} proteins.
    We compare \gls{er} (N=98) to Golgi (N=82), \gls{pm} (N=157), and the mitochondria referred to as ``mito'' (N=65).
    The hydrophobicity was predicted as the mean average of the values of the sequences of the \gls{tmh}, as well another group including up to $\pm$5 flanking residues, since predicting the boundary of \gls{tmh}s is difficult, according to the Kyte \& Doolittle hydrophobicity scale~\cite{Kyte1982}.
    The linguistic information entropy was calculated according to the methods section~\cite{Shannon1948}.
    The Test column refers to the statistical score obtained from the test; H statistic for the Kruskal Wallis (KW), the KS statistic for the Kolmogorov Smirnov test (KS), and the t\--statistic for the student's T\--test (T-test).
    $p$ is the p\--value of that statistical score.
    $B$ refers to the Bahadur slope, an interpretation of the p\--value that accounts for the sample size powering the test~\cite{Bahadur1967, Bahadur1971}.}
        \tiny
        % Table generated by Excel2LaTeX from sheet 'SwissProt filtered species'

        %\begin{tabular}{clrrrrrrrrr}
         \begin{tabular}{ccccccccccc}
                                &       & \multicolumn{3}{c}{ER and Golgi} & \multicolumn{3}{c}{ER and PM} & \multicolumn{1}{l}{ER and mito} &       &  \\
                                &       & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} & \multicolumn{1}{l}{Test} & \multicolumn{1}{l}{$p$} & \multicolumn{1}{l}{$B$} \\

        \midrule
        \multirow{3}[0]{*}{TMH Hydrophobicity} &  KW & 11.96 & 5.43E-4 & 4.18E-2 & 0.02  & 8.77E-1 & 5.14E-4 & 8.46  & 3.64E-3 & 3.45E-2 \\
                                &  KS & 0.27  & 1.98E-3 & 3.46E-2 & 0.08  & 8.48E-1 & 6.44E-4 & 0.27  & 4.62E-3 & 3.30E-2 \\
                                &  T-test & -3.47 & 6.50E-4 & 4.08E-2 & -0.17 & 8.67E-1 & 5.60E-4 & 3.45  & 7.24E-4 & 4.44E-2 \\
        \midrule
        \multirow{3}[0]{*}{... including flanks} &  KW & 5.92  & 1.50E-2 & 2.33E-2 & 9.14  & 2.50E-3 & 2.35E-2 & 26.42 & 2.75E-7 & 9.27E-2 \\
                                &  KS & 0.21  & 2.85E-2 & 1.98E-2 & 0.26  & 4.88E-4 & 2.99E-2 & 0.43  & 4.93E-7 & 8.91E-2 \\
                                &  T-test & -2.52 & 1.25E-2 & 2.43E-2 & -3.09 & 2.23E-3 & 2.40E-2 & 4.95  & 1.87E-6 & 8.09E-2 \\
      \midrule

        \multirow{3}[0]{*}{TMH entropy} &  KW & 2.96  & 8.56E-2 & 1.37E-2 & 0.66  & 4.17E-1 & 3.43E-3 & 0.69  & 4.05E-1 & 5.54E-3 \\
                                &  KS & 0.13  & 4.32E-1 & 4.66E-3 & 0.10  & 5.27E-1 & 2.51E-3 & 0.18 & 1.40E-1 & 1.20E-2 \\
                                &  T-test & 1.58  & 1.15E-1 & 1.20E-2 & 0.79  & 4.32E-1 & 3.29E-3 & 1.03 & 3.06E-1 & 7.26E-3 \\
        \midrule
        \multirow{3}[0]{*}{... including flanks} &  KW & 2.62  & 1.06E-1 & 1.25E-2 & 2.87  & 9.04E-2 & 9.42E-3 & 0.05 & 8.31E-1 & 1.14E-3 \\
                                &  KS & 0.15  & 2.48E-1 & 7.75E-3 & 0.17  & 6.56E-2 & 1.07E-2 & 0.21 & 6.33E-2 & 1.69E-2 \\
                                &  T-test & 1.84  & 6.75E-2 & 1.50E-2 & 1.66  & 9.84E-2 & 9.09E-3 & 0.42 & 6.72E-1 & 2.44E-3 \\
        \end{tabular}%
                        \label{table:organellesswissstats}
        \end{table}%

Similarly, in the SwissProt filtered dataset, the mean \gls{tmh} hydrophobicity for mitochondria is the lowest at 1.9.
It appears to be the Golgi apparatus that is the peak at 2.4.
In the SwissProt dataset, when we compare each subset of only the \gls{tmh} to the \gls{er} subset, we find significance between the \gls{er} and the Golgi ($p<1.98E-3$), and the \gls{er} and the mitochondria ($p<4.62E-3$), however, the \gls{er} and \gls{pm} are more similar considering the Bahadur values are $<6.44E-4$, two orders of magnitude smaller than the other sets (Bahadur values $>3.3E-2$) (Table \ref{table:organellesswissstats}).
When we take into account the flanks, the \gls{er} and \gls{pm} dataset can be distinguished ($P<2.50E-3$), however, as a trend the other two comparisons, \gls{er} and Golgi become less significant, and \gls{er} and mitochondria become more significant.

The information entropy of the \gls{tmh} string was also examined.
No significance was observed in any consideration of the information entropy, but similarly to the UniProt subset, as a trend, the entropy mirrors the hydrophobicity (Figure \ref{fig:average_organelle_factors_ta}).

Lipid asymmetry is caused by sphingomyelin and glycosphingolipids on the non\--cytosolic leaflet and phosphatidylserine and phosphatidylethanolamine in the cytosolic leaflet in the Golgi and \gls{pm}, however, this asymmetry is not present in the \gls{er}~\cite{Daleke2007, Devaux2004}.
Another consideration is that sphingomyelin is not present in the \gls{er} but is present in the Golgi~\cite{Futerman2005} and \gls{pm}~\cite{Li2007, Tafesse2007}.
Furthermore, the \gls{pm} contains densely packed sphingolipids and sterols~\cite{Paolo2006}.
Mitochondria have bacterial lipids in their membrane and uniquely contain cardiolipin \cite{Choi2005}, which is also confirmed to be present in the \gls{mom} \cite{Gebert2009}.
The compositional differences of the membrane environments inevitably places different constraints on the \gls{tmh}s, which may drive the observed variations in their biochemistry.

%Need mitochondria reference, and numbers for the UniER etc.
The hydrophobicity of \gls{tmh}s in \gls{ta} proteins is lower in proteins targeted to the mitochondria compared to those targeted to the \gls{er} \cite{Borgese2007}.
Here we observe that average biochemical features are evidently of significance.
Furthermore, we see that the typically positive\--inside negative outside tandem in positively and negatively charged residues is reversed in the mitochondria to positive\--outside negative\--inside.
This variation in \gls{tmh} hydrophobicity and the charged residue skew reversal may be yet another nuance of the system which goes some way to explaining how signals are maintained in local environments even when the average values are ambiguous.
We also identify that alanine is a key reason behind the hydrophobic difference between subcellular organelles, with alanine being highly selected, if not favoured over other hydrophobic residues like leucine, in mitochondrial \gls{ta} proteins compared to the \gls{er}, the Golgi, and the \gls{pm}.
This could be an adaptation to the mitochondrial membrane, which contains a higher level of cardiolipins than other membranes \cite{VanMeer2008, Gebert2009}.
Regarding the charged residue distribution, there should also be a consideration of membrane potential.
Although membrane potentials are in flux, typically, the \gls{pm} has a potential of ~70mV with the negativity being on the cytoplasmic side.
It has been shown that the \gls{er} is again between 75-95mV with negativity on the lumenal side \cite{Qin2011, Worley1994}.
There is no detectable potential across the Golgi \cite{Schapiro2000}, and the mitochondrial inner membrane has a potential of 150-180 with negativity on the matrix side \cite{Perry2011}.
However, whilst those numbers go some way to satisfy the flanking charge reversal we see, they do not apply to the \gls{mom} in which the \gls{ta} proteins are localised; $\beta$ barrel porins on the \gls{mom} allow passive ion and water transfer across the membrane effectively diminishing the membrane potential.
Furthermore, these observed membrane potential values are by their nature not fixed values.
Not only would this be most striking in the nerve and muscle cells of course, but also would contrast between resting and stimulated secretory cells.

It has been known for some time that both the charge and the hydrophobic length of the \gls{ta} protein \gls{tmh} region are indicative of subcellular targeting of the protein \cite{Borgese2007}.
\gls{er} localised \gls{ta} proteins typically have a more hydrophobic \gls{tmh} than those targeted to the \gls{mom} \cite{Wang2010}.
The C\--terminal tail charge is particularly important for determining subcellular localisation, and can even override the hydrophobic signal if strong enough \cite{Costello2017}.
Here we see that while hydrophobicity is statistically different between the subcellular membranes, there is overlap.
Whilst there are differences in total average charge at the C\--terminal flank, this is not an absolute rule.
Perhaps these \gls{ta} protein anchor \gls{tmh}s are mechanistically similar to those of the signal anchor proteins, a group of singlepass \gls{tmp}s containing an uncleaved \gls{sp} that remains in the membrane.
The single hydrophobic segment in mitochondrial signal\--anchored proteins serves as both a mitochondrial targeting signal and a membrane anchor.
Signal\--anchored proteins, along with some \gls{ta} proteins, have been shown to be able to spontaneously insert into the membrane independently from the translocon~\cite{Elisa2012, Lan2000, Colombo2009}.

Nevertheless, it is tempting to conclude that these biochemical differences between differently localised \gls{ta} proteins are an adaptation to the membrane composition and environment.
But this must be tempered by noting that the spontaneously inserting cytochrome b5 localises to the mitochondrial membrane in the absence of cytosol, and to the \gls{er} in the presence of cytosol \cite{Costa2018}; there are also biological factors determining localisation.

In summary, the mitochondria located \gls{ta} protein \gls{tmh}s typically have a preference for alanine over leucine unlike their secretory counterparts and have a negative\--inside positive\--outside tendency counter to the overwhelming majority of \gls{tmp}s.
It is unclear if these features are a biophysical adaptation or part of a biological sorting process.

\subsection{More annotation is required to identify chaperone interaction factors of the transmembrane helix.}

\gls{ta} proteins known to interact with certain chaperones were acquired by filtering the interactor partner IDs for chaperones from BioGrid through the redundant versions of these UniProt manually curated lists and SwissProt automatically generated lists.

Hsp40, Hsc70, SRP54 (both plant and human) returned 0 hits, indicating a lack of annotation regarding \gls{ta} proteins with these chaperons probably due to the relatively polar, and non-trivially predictable, \gls{tmh}s of \gls{ta} proteins that these chaperones interact with.

Snd1 has 15 records that were in our \gls{ta} lists.
The average Kyte \& Doolittle hydrophobicity of these records was 2.60 in the \gls{tmh} itself and 1.58 including $\pm$ 5 flanking residues.
Sgt2, with 14 records, had a \gls{tmh} hydrophobicity of 2.47 and 1.51 including the flanks.
5 records were captured for SGTA with a \gls{tmh} hydrophobicity of 2.27 and 1.19 including the flanks.
TRC40 had the highest \gls{tmh} hydrophicity of 2.77 and 1.82 including flanks.
However, TRC40 also only had 7 records.
Get3 had 22 \gls{ta} interactor records with an average \gls{tmh} hydrophobicity of 2.36 and 1.48 including the flanks.
The 2 records for human Pex19 had an average hydrophobicity of 1.33 for the \gls{tmh} and 0.70 including the flanking residues.
The yeast Pex19 had a \gls{tmh} average hydrophobicity of 2.48 and 1.41 including the flanking residues.
The 4 yeast SRP54 interactors had an average \gls{tmh} of 2.43 and 1.98 including the flanking residues.

At the time of the investigation, these sample sizes are not statistically viable for analysis.
Whilst it appears TRC40 interactors have notably hydrophobic \gls{tmh}s, TRC40s yeast homologue Get3 has interactors with much more polar \gls{tmh}s, yet SGTA was lower than Sgt2 on average.
So although these average values differ and overlap between various chaperone systems, we tried to identify clearer patterns from the \gls{tmh} hydrophobic profiles (Figure \ref{fig:interaction-profile}).
Similarly, whilst a clear dip in hydrophobicity at position +5 in Get3 from between 2-3 across the rest of the \gls{tmh} core to 0.32, there is no such spike for TRC40 meaning this is probably not of any functional importance, but rather an artefact of overrepresented proteins in the Get3 dataset.
Snd1 also lies among these values, reinforcing Snd1 as a biological redundancy system \cite{Rabu2009, Johnson2013, Schuldiner2008}.

As expected, the human Pex19 interactors are as a trend among the most polar throughout the \gls{tmh} core, however, when we consider the yeast Pex19, this trend is less clear (Figure \ref{fig:interaction-profile}). In figure \ref{fig:interaction-profile},~at least at a handful of locations (-3, 3, 4 and 5) the SRP54 interactors have the most hydrophobic \gls{tmh} cores.

TRC40 has the highest \gls{tmh} mean average hydrophobicities at a given point at 3.9 at position -1.
However, the yeast homologue Get3 does not appear to have any preference for especially hydrophobic \gls{tmh}s (Figure \ref{fig:interaction-profile}).

In order to remove redundant proteins and investigate this further, more records with greater levels of accurate annotation need to be available to both BioGrid and UniProt.
We also observe a great deal of overlap between the profiles, indicating that as a trend this is more complex than hydrophobicity alone and that net/average \gls{tmh} polarity is not the absolute determinant of chaperone association.

However, this method demonstrates a potential way that this chaperone\--interaction problem can be investigated to verify that indeed hydrophicity plays a deterministic role in chaperone selection.

\begin{figure}
\centering
\includegraphics[width=1\textwidth]{TA_chapter/interaction-profile}
\captionof{figure}[The profile of transmembrane helix and flanks hydrophobicity from tail\--anchored protein groups stratified by chaperone interactors.]
{\textbf{The profile of transmembrane helix and flanks hydrophobicity from tail\--anchored protein groups stratified by chaperone interactors.}
On the horizontal axis is the position relative to the central TMH residue defined by UniProt.
On the vertical axis is the Kyte \& Doolittle hydrophobicity windowed across 5 residues allowing for half windows.
The chaperone interactors are colour coded according to the key.
}

\label{fig:interaction-profile}
\end{figure}

\subsection{Spontaneous insertion may be achieved by polar strips in the transmembrane helix of tail-anchored proteins}

The \gls{tmh}s of cytochrome b5 and PTP1b are among the least hydrophobic of the \gls{ta} proteins \cite{Rabu2008, Rabu2009}.
Indeed the \gls{tmh} is so polar that it is not trivial to predict and is not found in either dataset prepared herein.
Structural modelling and analysis thereof reveal features that may explain the ``missing hydrophobicity''~\cite{Hessa2005, Hedin2010, Hessa2007, Ojemalm2012} of these particular \gls{tmh}s.

A positive patch can be seen on the cytoplasmic side of either protein \gls{tmh}, and a negative patch on the lumenal tip at the C\--terminus.
There are large ``positive\--inside'' patches \cite{VonHeijne1989, Andersson1992, Sharpe2010, Baeza-Delgado2013, Pogozheva2013, Baker2017} and a strong ``negative\--outside'' charge \cite{Baker2017}~(Figure \ref{fig:cytb5-biochemistry}C) at the respective flanks of the \gls{tmh}.
Such a distribution of charge is typical of a \gls{tmh} anchor.
Once in the membrane, this may allow the \gls{tmh} to be an effective anchor despite such poor hydrophobicity since it satisfies electrostatic coupling to the membrane potential.

Furthermore, there is the question of overcoming the unfavourable interaction most \gls{tmh}s would hypothetically face when coming into contact with the highly polar membrane interface without the presence of membrane integration machinery.
We observe a highly conserved strip of relatively polar / non-hydrophobic residues on one side of the \gls{tmh} core (in cytochrome b5 these are N112, P116, A120, A124 Y127, and R128).
Similarly, a polar face exists for the PTP1b \gls{tmh} (R430, N434, Y426, T422, and T419).
These polar faces would not be as repulsed by the interfacial environment as either a more hydrophobic \gls{tmh} or an equally hydrophobic \gls{tmh} with a different sequence and structure order (Figure \ref{fig:cytb5-biochemistry} and Figure \ref{fig:ptp1b-biochemistry}).
Scrambling the cytochrome b5 \gls{tmh} sequence whilst maintaining the same hydrophobicity (DSNSS W W T N W V I P A I S A L I V A L M YR to DSNSS W W A S A I I A T M I P L L V N V W YR) reduces the insertion potential \cite{Brambillasca2006}.
Therefore we can conclude that there is more to this phenomenon than hydrophobicity alone.
It becomes apparent that the 3D arrangement of these relatively polar \gls{tmh} residues is conserved and is probably the key to spontaneous insertion of \gls{tmh}s.

\begin{figure}[!ht]
\centering
\includegraphics[width=1\textwidth]{TA_chapter/cytb5-biochemistry}
        \captionof{figure}[Structural biochemical analysis of a homology model of cytochrome b5.]{\textbf{Structural biochemical analysis of a homology model of cytochrome b5.}
        (A) The secondary structure of the protein coloured from the N terminus in blue to the C terminus in red coloured through the rainbow according to the residue number.
        (B) The hydrophobicity of the \gls{tmh} from white representing relatively polar residues to red showing relatively hydrophobic residues \cite{Eisenberg1984}.
        (C) The electrostatic surface with a threshold of $\pm5$ KT/e calculated by APBS in PyMol \cite{Baker2001}.
        Red patches are negatively charged whilst blue is positively charged.
        (D) The consurf scores on a scale of 1-9 (all residues had sufficient data) \cite{Ashkenazy2010}. Purple represents the most conserved whilst blue is the least.
        Note the correlation between the highly and modestly conserved \gls{tmh} residues and the relatively polar residues.
        Another observable feature is the very strong ``positive inside'' \cite{VonHeijne1989, Andersson1992, Sharpe2010, Baeza-Delgado2013, Pogozheva2013} and ``negative outside'' features which are associated with anchoring \cite{Baker2017}.
}

\label{fig:cytb5-biochemistry}
\end{figure}


\begin{figure}[!ht]
\centering
\includegraphics[width=1\textwidth]{TA_chapter/ptp1b-biochemistry}
        \captionof{figure}[Structural biochemical analysis of a homology model of PTP1b.]{\textbf{Structural biochemical analysis of a homology model of PTP1b.}
        (A) The secondary structure of the protein coloured from the N terminus in blue to the C terminus in red coloured through the rainbow according to the residue number.
        (B) The hydrophobicity of the \gls{tmh} from white representing relatively polar residues to red showing relatively hydrophobic residues \cite{Eisenberg1984}.
        (C) The electrostatic surface with a threshold of $\pm5$ KT/e calculated by APBS in PyMol \cite{Baker2001}.
        Red patches are negatively charged whilst blue is positively charged.
        Note the one hydrophobic face of the \gls{tmh} and the opposing relatively polar face.
        Another observable feature is the ``positive inside'' \cite{VonHeijne1989, Andersson1992, Sharpe2010, Baeza-Delgado2013, Pogozheva2013} and ``negative outside'' features which are associated with anchoring \cite{Baker2017}.
}

\label{fig:ptp1b-biochemistry}
\end{figure}

We speculate that this polar face allows the unassisted approach to the membrane's polar phospholipid head groups, and once in sufficiently close proximity, the hydrophobic side of the helix is entropically driven by the water environment into the membrane core (Figure \ref{fig:spont-insertion-patch}).
This close proximity is less likely to be achieved if there is no side of the \gls{tmh} to favourably interact with the lipid head groups, even though average hydrophobicity could be similar.
Once integrated, the charged \gls{tmh} flanking regions help sustain integration in lieu of a more hydrophobic \gls{tmh} core.

\begin{figure}[!ht]
\centering
\includegraphics[width=1\textwidth]{TA_chapter/spont-insertion-patch}
        \captionof{figure}[A cartoon of a potential method the cytochrome b5 and PTP1b transmembrane helix could integrate spontaneously into the membrane.]{\textbf{A cartoon of a potential method the cytochrome b5 and PTP1b transmembrane helix could integrate spontaneously into the membrane.}
        A) The marginally hydrophobic \gls{tmh} in teal cannot approach the membrane interface.
        B) Although the average hydrophobicity is the same as the teal \gls{tmh}, by having a more hydrophobic side (blue), and a more polar side (green), the \gls{ta} protein now has a \gls{tmh} surface that may interact more favourably with the interfacial region.
        Once interacting with the membrane sufficiently close to the interfacial region, the hydrophobic face is still being entropically driven by water molecules from the cytosol, which would lead to partitioning into the membrane.
        Once integrated, the strong ``positive\--inside'' ``negative\--outside'' charges on the \gls{tmh} flanks compensate for the lack of hydrophobicity in the core of the \gls{tmh}.

}

\label{fig:spont-insertion-patch}
\end{figure}

\section{Summary}

Here, we have observed a clear biochemical distinction between \gls{ta} proteins with different terminal subcellular destinations.
Previously it was known that both hydrophobicity and charge are involved in targeting to distinct subcellular locations \cite{Costello2017}.
In this study, we find that the location of the charge along and around the \gls{tmh} of \gls{ta} proteins differs among subcellular compartments.
Crucially, there is a shift in the charged residue inside\--outside tandem in the \gls{tmh} flanking residues in different organelles.
In the secretory pathway, the ``positive\--inside'' ``negative\--outside'' skew adheres to the membrane\--potential for electrostatic coupling, but in the \gls{mom} where there is no membrane potential, positively charged residues are skewed to be located outside the cytoplasm similarly to $\beta$ barrels \cite{Pogozheva2013} which are also located on the \gls{mom}, and the negatively charged residues are preferentially found inside the cytoplasm.
Furthermore, the missing hydrophobicity of mitochondrial \gls{ta} proteins can be in part attributed to the high abundance of alanine rather than leucine or isoleucine in the \gls{tmh}; it is not solely the tolerance of polar residues in the \gls{tmh}.
We expected to see evolutionary adaptations of \gls{tmh} hydrophobicity to species-specific membranes, even within eukaryotes~\cite{Baker2017, Sharpe2010}.
In this study using both a manually curated dataset from UniProt and an automatically filtered list using SwissProt annotation, we do not observe any strong differences.
Since we could not scrutinise a difference in the species, the strong hydrophobic differences between organelle \gls{tmh}s are indicative of a stronger adaptation pressure than between species as a whole.
These differences are likely to be partially adaptations to the organelle location membrane type and also possible cryptic biological factors that play a role in their targeting via chaperone\--binding affinity.
We could not find any clear trends or perform statistical work on the chaperone interactor datasets due to the small sample sizes, however, as the databases are enriched, this same method will be able to answer the questions about chaperone affinity with more accuracy in the future.

Furthermore, the spontaneously inserting \gls{ta} proteins PTP1b and cytochrome b5 appear to share a polar face that emerges in structural models and a strong ``positive\--inside'' ``negative\--outside'' electrostatic surface.
The polar face may be responsible for the promotion insertion potential in the absence of insertion proteins since when the sequence is scrambled, the insertion potential is reduced \cite{Brambillasca2006}.
The positively and negatively charged residues are distributed like an ideal anchoring \gls{tmh} \cite{Baker2017}, which could allow the marginally hydrophobic \gls{tmh} to perform as a suitable membrane anchoring feature.