# Tracked changes in LaTeX

Maybe people keep telling you Word is great but you are just too emotionally attached to LaTeX to consider using anything else. It just looks so beautiful. Besides, you would have to leave your beloved linux environment (maybe that’s just me), so you stick with what you know. You work for many weeks long and hard, finally producing a draft of a paper that gets the all clear from your supervisor to submit to journal X. Eventually you hear back and the reviewers have responded with some good ideas and a few pedantic points. Apparently this time the journal wants a tracked changes version to go with your revised manuscript.

Highlighting every change sounds like a lot of bother, and besides, you’d have to process the highlighted version to generate the clean version they want you to submit alongside it. There must be a better way, and one that doesn’t involve converting your document to Word.

Thankfully, the internet has an answer! Check out this little package changes which will do just what you need. As long as you annotate using \deleted{}, \replaced{} and \added{} along the way, you will have to change just one word of your tex source file in order to produce the highlighted and final versions. It even comes with a handy bash script to get rid of the resulting mess when you’re happy with the result, leaving you with a clean final tex source file.

The die-hard Word fans won’t be impressed, but you will be very satisfied that you have found a nice little solution that does just the job you want it to. It’s actually capable of much more, including comments by multiple authors, customisation of colours and styles, and an automatically generated summary of changes. I have heard good things about ShareLaTeX for collaboration, but this simple package will get you a long way if you are not keen to start paying money yet.

# Rational Design of Antibody Protease Inhibitors

On the theme of my research area and my last presentation I talked at group meeting about a another success story in structurally designing an antibody by replicating a general protein binding site using grafted fragments involved in the original complex. The paper by Liu et al is important to me for two major reasons . Firstly they used an unconventional antibody for protein design, namely a bovine antibody which is known to have an extended CDR H3. Secondly the fragment was not grafted at the anchor points of the CDR loop.

SFTI-1 is a cyclic peptide and a known trypsin inhibitor. It’s structure is stabilised by a disulphide bridge. The bovine antibody is known to have an extended H3 loop which is essentially a long beta strand stalk with a knob domain at the end. Liu et al removed the knob domain and a portion of the B strand and grafted the acyclic version of the SFTI-1 to it. As I said above this result is very important because it shows we can graft a fragment/loop at places different then the anchor points of the CDR. This opens up the possibility for more diverse fragments to be grafted because of new anchor points,  and also because the fragment will sit further away from the other CDRs and the framework allowing more conformational space. To see how the designed antibody compares to the original peptide they measured the Kd and found a 4 fold increase (9.57 vs 13.3). They hypothesise that this is probably due to the fact that the extended beta strand on the antibody keeps the acyclic SFTI-1 peptide in a more stable conformation.

The problem with the bovine antibody is that if inserted in a human subject it would probably elicit an immune response from the native immune system. To humanise this antibody they found the human framework which shares the greatest sequence identity to the bovine antibody and then grafted the fragment on it. The human antibody does not have an extended CDR H3 and to decide what is the best place of grafting they tried various combinations again showing again that the fragments do not need to grafted exactly at the anchor points. Some of the resulting  human antibodies showed even more impressive Kds.

The best designed human antibody had a 0.79nM Kd, another 10-fold gain . Liu et al hypothesised that this might be due to the fact that the cognate protein forms contacts with residues on the other CDRs even though there is no crystal structure to show this. In order to test this hypothesis they mutated surface residues on the H2 and L1 loop to Alanine which resulted in a 6.7 fold decrease in affinity. The only comment I would have to this is that the mutations to the other CDRs might have destabilized the other CDRs on the antibody which could be the reason for the decrease in affinity.

# Quantifying dispersion under varying instrument precision

Experimental errors are common at the moment of generating new data. Often this type of errors are simply due to the inability of the instrument to make precise measurements. In addition, different instruments can have different levels of precision, even-thought they are used to perform the same measurement. Take for example two balances and an object with a mass of 1kg. The first balance, when measuring this object different times might record values of 1.0083 and 1.0091, and the second balance might give values of 1.1074 and 0.9828. In this case the first balance has a higher precision as the difference between its measurements is smaller than the difference between the measurements of balance two.

In order to have some control over this error introduced by the level of precision of the different instruments, they are labelled with a measure of their precision $latex 1/\sigma_i^2$ or equivalently with their dispersion $latex \sigma_i^2$ .

Let’s assume that the type of information these instruments record is of the form $latex X_i=C + \sigma_i Z$,  where $latex Z \sim N(0,1)$ is an error term, $latex X_i$ its the value recorded by instrument $latex i$ and where $latex C$ is the fixed true quantity of interest the instrument  is trying to measure. But, what if $latex C$ is not a fixed quantity? or what if the underlying phenomenon that is being measured is also stochastic like the measurement $latex X_i$. For example if we are measuring the weight of cattle at different times, or the length of a bacterial cell, or concentration of a given drug in an organism, in addition to the error that arises from the instruments; there is also some noise introduced by dynamical changes of the object that is being measured. In this scenario, the phenomenon of interest, can be given  by a random variable $latex Y \sim N(\mu,S^2)$. Therefore the instruments would record quantities of the form $latex X_i=Y + \sigma_i Z$.

Under this case, estimating the value of $latex \mu$, the expected state of the phenomenon of interest is not a big challenge. Assume that there are $latex x_1,x_2,…,x_n$ values observed from realisations of the variables $latex X_i \sim N(\mu, \sigma_i^2 + S^2)$, which came from $latex n$ different instruments. Here $latex \sum x_i /n$ is still a good estimation of $latex \mu$ as $latex E(\sum X_i /n)=\mu$.  Now, a more challenging problem is to infer what is the underlying variability of the phenomenon of interest $latex Y$. Under our previous setup, the problem is reduced to estimating $latex S^2$ as we are assuming $latex Y \sim N(\mu,S^2)$ and that the instruments record values of the from $latex X_i=Y + \sigma_i Z$.

To estimate $latex S^2$ a standard maximum likelihood approach could be used, by considering the likelihood function:

$latex f(x_1,x_2,..,x_n)= \prod e^{-1/2 \times (x_i-\mu)^2 /(\sigma_i^2+S^2)} \times 1/\sqrt{2 \pi (\sigma_i^2+S^2) }$,

from which the maximum likelihood estimator of $latex S^2$ is given by the solution to

$latex \sum [(X_i- \mu)^2 – (\sigma_i^2 + S^2)] /(\sigma_i^2 + S^2)^2 = 0$.

Another more naive approach could use the following result

$latex E[\sum (X_i-\sum X_i/n)^2] = (1-1/n) \sum \sigma_i^2 + (n-1) S^2$

from which $latex \hat{S^2}= (\sum (X_i-\sum X_i/n)^2 – ( (1-1/n ) \sum(\sigma_i^2) ) ) / (n-1)$.

Here are three simulation scenarios where 200 $latex X_i$ values are taken from instruments of varying precision or variance $latex \sigma_i^2, i=1,2,…,200$ and where the variance of the phenomenon of interest $latex S^2=1500$. In the first scenario $latex \sigma_i^2$ are drawn from $latex [10,1500^2]$, in the second from $latex [10,1500^2 \times 3]$ and in the third from $latex [10,1500^2 \times 5]$. In each scenario the value of $latex S_2$ is estimated 1000 times taking each time another 200 realisations of $latex X_i$. The values estimated via the maximum likelihood approach are plotted in blue, and the values obtained by the alternative method are plotted in red. The true value of the $latex S^2$ is given by the red dashed line across all plots.

First simulation scenario where $latex \sigma_i^2, i=1,2,…,200$ in $latex [10,1500^2]$. The values of  $latex \sigma_i^2$ plotted in the histogram to the right. The 1000 estimations of $latex S$ are shown by the blue (maximum likelihood) and red (alternative) histograms.

First simulation scenario where $latex \sigma_i^2, i=1,2,…,200$ in $latex [10,1500^2 \times 3]$. The values of $latex \sigma_i^2$ plotted in the histogram to the right. The 1000 estimations of $latex S$ are shown by the blue (maximum likelihood) and red (alternative) histograms.

First simulation scenario where $latex \sigma_i^2, i=1,2,…,200$ in $latex [10,1500^2 \times 5]$. The values of $latex \sigma_i^2$ plotted in the histogram to the right. The 1000 estimations of $latex S$ are shown by the blue (maximum likelihood) and red (alternative) histograms.

For recent advances in methods that deal with this kind of problems, you can look at:

Delaigle, A. and Hall, P. (2016), Methodology for non-parametric deconvolution when the error distribution is unknown. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78: 231–252. doi: 10.1111/rssb.12109

# Improving the precision of local community detection

Communities are everywhere. You can find groups of connected people on Facebook who went to university together, communities of people on youtube that like the same band, sets of products on amazon that tend to be purchased together, or modules of proteins that act in the same biological process in the human body. With the amount of data we can collect on these networks (facebook friendship, protein interactions, etc), you would think we can easily recover these groups. You would think wrong…

There are a lot of methods out there to find communities in networks. As a paper that compares community detection methods by Lancichinetti and Fortunato put it: “A thorough analysis of all existing methods would require years of work (during which many new methods would appear), and we do not believe it is feasible”. Thus, clearly we have not found the answer.

So what do people generally do to find these communities? Well… very different things. To explain a few of the approaches, it helps to break them down into broad categories such as global and local methods.

## Global Methods

Global community detection methods generally aim to optimized an objective function over the entire network. For example, I could decide that I will score the network partition based on a function I will call “greatness”, G. I will define the greatness of a community by the sum of the number of nodes each node is connected with that are in the community (the sum of the node in-degree). Or simply:

Here, A_ij is the adjacency matrix that contains a 1 where nodes i and j interact, and 0 otherwise. The sigmas in the equation correspond to the community identifiers of nodes i and j respectively, and thus the delta function is only 1 when sigma_i and sigma_j are equal, meaning nodes i and j are in the same community, and 0 otherwise.

If you now come to the conclusion that this is a pretty stupid thing to define, then you’ve understood the concept of a global method (alternatively… get rid of that negativity!). A global community detection method would now look how to partition the network so that we can maximize G. In the case of greatness defined here, everything will be put into the same community, as there’s no penalty for putting two nodes together that shouldn’t belong together. Adding a further node to a community can never decrease the greatness score!

If you want a real overview over these and other methods, check out this 100-page review article.

## Local Methods

In contrast to global methods that attempt to gather information on the whole network, local community detection methods dive into the network to try to optimize the community partitioning. This means on the one hand they can be very fast at assigning nodes to communities, as they take into account less information. However, on the other hand using less information in this assignment means they can be less precise.

An example of a local method is link clustering, which I described in an earlier post. Link clustering assesses the similarity of two edges that share a node, rather than comparing the nodes themselves. The similarity of these edges is calculated based on how similar the surroundings of the nodes on the endpoints are. Using this score, all edges that are more similar than a cutoff value, S, can be clustered together to define the edge-communities at resolution S. This partitioning translates to nodes by taking the nodes that the edges in the edge-communities connect.

## Improving local community detection

Link clustering is a pretty nice and straightforward example of how you can successfully find communities. However, the authors admit it works better on dense than on sparse networks. This is an expected side-effect of taking into account less information to be faster as mentioned above. So you can improve it with a pretty simple idea: how much more information can I take into account without basically looking at the whole network?

Local community detection methods such as link clustering are generally extended by including more information. One approach is to include information other than network topology in the local clustering process (like this). This approach requires information to be present apart from the network. In my work, I like using orthogonal data sets to validate community detection, rather than integrate it into the process and be stuck without enough of an idea if it actually worked. So, instead I’m trying to include more of the local surroundings of the node to increase the sensitivity of the similarity measure between edges. At some point, this will blow up the complexity of the problem and turn it into a global method. So I’m now walking a line between precision and speed. A line that everyone always seems to be walking. On this line, you can’t really win… but maybe it is possible to optimize the trade-off to suit my purposes. And besides… longer algorithm run times give me more time for lunch breaks ;).

# Conformational Variation of Protein Loops

Something many structural biologists (including us here in OPIG!) are guilty of is treating proteins as static, rigid structures. In reality, proteins are dynamic molecules, transitioning constantly from  conformation to conformation. Proteins are therefore more accurately represented by an ensemble of structures instead of just one.

In my research, I focus on loop structures, the regions of a protein that connect elements of secondary structure (α-helices and β-sheets). There are many examples in the PDB of proteins with identical sequences, but whose loops have different structures. In many cases, a protein’s function depends on the ability of its loops to adopt different conformations. For example, triosephosphate isomerase, which is an important enzyme in the glycolysis pathway, changes conformation upon ligand binding, shielding the active site from solvent and stabilising the intermediate compound so that catalysis can occur efficiently. Conformational variability helps triosephosphate isomerase to be what is known as a ‘perfect enzyme’; catalysis is limited only by the diffusion rate of the substrate.

Structure of the triosephosphate isomerase enzyme. When the substrate binds, a loop changes from an ‘open’ conformation (pink, PDB entry 1TPD) to a ‘closed’ one (green, 1TRD), which prevents solvent access to the active site and stabilises the intermediate compound of the reaction.

An interesting example, especially for some of us here at OPIG, is the antibody SPE7. SPE7 is multispecific, meaning it is able to bind to multiple unrelated antigens. It achieves this through conformational diversity. Four binding site conformations have been found, two of which can be observed in its unbound state in equilibrium – one with a flat binding site, and another with a deep, narrow binding site [1].

SPE7; an antibody that exists as two different structures in equilibrium – one with a shallow binding site (left, blue, PDB code 1OAQ) and one with a deep, narrow cleft (right, green, PDB 1OCW). Complementary determining regions are coloured in each case.

So when you’re dealing with crystal structures, beware! X-ray structures are averages – each atom position is an average of its position across all unit cells. In addition, thanks to factors such as crystal packing, the conformation that we see may not be representative of the protein in solution. The examples above demonstrate that the sequence/structure relationship is not as clear cut as we are often lead to believe. It is important to consider dynamics and conformational diversity, which may be required for protein function. Always bear in mind that the static appearance of an X-ray structure is not the reality!

[1] James, L C, Roversi, P and Tawfik, D S. Antibody Multispecificity Mediated by Conformational Diversity. Science (2003), 299, 1362-1367.

# “Identifying Allosteric Hotspots with Dynamics”: Using molecular simulation to find critical sites for the functional motions in proteins

Allosteric (allo-(“other”) + steric (repulsion of atoms due to closeness or arrangement)) sites regulate protein function from a position other than the active site or binding site. Consider the latch on a pair of gardening scissors (Figure 1): depending on the position of the latch (allosteric site) the blades are prevented from cutting things at the other end (active site).

Figure 1 Allostery explained: A safety latch in gardening scissors.

Due to the non-trivial positions of allosteric sites in proteins their identification has been challenging. Selected well characterised systems such as GPCRs have known allosteric sites that are being used as targets in drug development. However, large scale identification of allosteric sites across the Protein Data Base (PDB) has not yet been feasible, partly because of the lack of tools.

To tackle this problem the Gerstein Lab developed a computational protocol based on various molecular simulations and network methods to find allosteric hotspots in proteins across the PDB. They introduce two different pipelines; one for identifying allosteric residues on the surface (surface-critical) and one for buried residues (interior-critical).

For the search of exterior-critical residues they us MC simulations to repeatedly probe the surface of the protein with short Monte Carlo (MC) with a short peptide. Based on hard spheres and simple energy calculations this method seems to be an efficient way of detecting possible binding pockets. Once the binding pockets have been found, the collective motions of the structure are simulated using an elastic mass-and-spring network (an anisotropic network model [ANM]). Binding pockets that undergo significant deformation during these simulations are considered to be surface-critical.

For interior-critical residues they start by weighting residue-residue contacts on the basis of collective movement. Communities within the weighted network are then identified and the residues with the highest betweenness interactions between communities are chosen as interior-critical residues. Thus, interior-critical residues have the highest information flow between two densely inter-connected groups of residues.

The protocol as been implemented in STRESS (STRucturally identified ESSential residues) and is freely available at stress.molmovdb.org.

Publication: http://www.ncbi.nlm.nih.gov/pubmed/27066750

# Comp Chem Kitchen

I recently started  “Comp Chem Kitchen” with Richard Cooper and Rob Paton in the Department of Chemistry here in Oxford. It’s a regular forum and seminar series  for molecular geeks and hackers, in the original, untarnished sense of the word: people using and developing computational methods to tackle problems in chemistry, biochemistry and drug discovery. Our hope is that we will share best practices, even code snippets and software tools, and avoid re-inventing wheels.

In addition to local researchers, we invite speakers from industry and non-profits from time to time, and occasionally organize software demos and tutorials.

We also provide refreshments including free beer. (We are grateful to Prof. Phil Biggin and the MRC Proximity to Discovery Fund for offering to support CCK.)

CCK-1

Our first meeting, CCK-1,  was held in the Abbot’s Kitchen on May 24, 2016, at 5 pm, and was a great success—standing room only, in fact! The Abbot’s Kitchen—originally a laboratory—is a beautiful stone building built in 1860 in the Victorian Gothic style, alongside the Natural History Museum, at a time when Chemistry was first recognized as a discipline.

We heard a fascinating talk from Jerome Wicker from the Department of Chemistry who spoke about “Machine learning for classification of solid form data extracted from CSD and ZINC”, and described a method that could successfully discriminate (~80%) whether a small molecule would crystallize or not. The software tools he discussed included RDKit, CSD, and scikit-learn. There were also two lightning talks, each 5 minutes long, one from OPIG member Hannah Patel,  from the Department of Statistics, on “Novelty Score: Prioritising compounds that potentially form novel protein-ligand interactions and novel scaffolds using an interaction centric approach”, who briefly described her Django-based web interface to her RDKit-based tool to analyse structures of ligands bound to proteins and help guide future medicinal chemistry to find novel compounds. We also had a talk from Dr Michael Charlton from InhibOx spoke about “Antibacterial Drug Discovery and Machine Learning”.

CCK-2

Our next Comp Chem Kitchen, CCK-2, will be held next Tuesday (June 14th, 2016), and you can register free for CCK-2 here.

We will have talks from:

Hope to see you there!  (Did I say free beer?)

# RDKit User Group Meeting 2016

If you’re interested in small molecules and cheminformatics, you might like to know that registration for the 2016 RDKit User Group Meeting is now open. The meeting will be held from the 26th-28th of October in Basel, Switzerland. From their announcement:

This year’s RDKit User Group Meeting will take place from 26-28 October at the Novartis Campus in Basel, Switzerland and is co-organized by people from Roche and Novartis.

Registration for the RDKit UGM is free:

https://www.eventbrite.com/e/5th-rdkit-user-group-meeting-tickets-22539677783

The previous years’ format seemed to work pretty well and the feedback was positive, so we will stick to the same format this year:

Days 1 and 2: Talks, lightning talks, roundtable(s), discussion, and talktorials. For those who haven’t attended before, talktorials are somewhere between a talk and a tutorial, they cover something interesting done with the RDKit and include the code used to do the work. During the presentation you’ll give an overview of what you did and also show the pieces of the code that are central to the work. The idea is to mix the science up with the tutorial aspects.

Day 3 will be a sprint: those who choose to stay will spend an intense day working in small groups to produce useful artifacts: new bits of code, KNIME nodes, KNIME workflows, tutorials, documentation, IPython notebooks, etc. We will once again try to structure this a bit by collecting a bunch of ideas for things to work on in advance.

There will also be, of course, social activities.

Looks like fun!

# Structure prediction through sequence physical properties

Today, protein structure is mainly predicted by aligning the unknown amino acid sequence against all sequences for which we already know the physical structure. Whilst sequences differing in length can be readily catered for by inserting or deleting (with or without affine gap penalties) the odd amino acid, there will frequently be cases where there are mutations. To compensate for this, the likes of the BLOSUM matrix is used to score the likelihood of one amino acid having mutated into another. For example, a hydrophobic residue is more likely to be swapped for something similar, than it is to be replaced with something strongly hydrophilic.

Whilst this is an entirely reasonable basis, there are many other physical property factors which which can be considered. In fact, the Amino Acid Index currently lists 544 physicochemical and biochemical properties of amino acids. A paper by Yi He et al. PNAS 2015;112:5029-5032 recently made use of a subset of these properties to predict structure. Their work shown below shows target T0797 (B) from CASP 11 compared with a purely physical structure predicted using their method (A) and the PSI BLAST candidate for the same sequence (C).

Even though their structure only had three residues in common with the target sequence, it is plainly more similar than the PSI BLAST attempt. The RMSD between structures A and B is also reported as being 0.73 Å, whilst PSI BLAST returns an RMSD of 2.09 Å.

# Visualising Biological Data, Pt. 2

Here’s a little quick round-up of some of the tools/algorithms that I’ve seen in VIZBI, which I believe can be useful for many. For more details, I strongly advise you check out the posters page (vizbi.org/Posters/2016). There were a few that I would’ve liked to re-visit, but the webapps weren’t available (e.g. MeshCloud from the Human Genome Center, Tokyo), so maybe I’ll come back with a part 3. Here are my top five:

1. Autodesk’s Protein Viewer* (shout-outs to @_merrywang on Twitter)
As a structural bioinformatician, I’m going to be really biased here, and say that Autodesk’s Molecule viewer was the best tool that was showcased in the conference. It combines not only the capacity to visualise millions of molecules from the PDB (or your own PDB files), it also allows annotation and sharing, effectively, “snapshots” of your workspace for collaboration (see this if you want to know what I mean). AND it’s free! It’s not the fastest viewer on the planet, nor the easiest thing to use, but it is effective.

2. Vectorbase
Not related to protein structures, but a really interesting visualisation that shows information on, for example, insecticide resistance. With mosquitoes being such a huge part of today’s news, this kind of information is vital for fighting and understanding the distribution of insects across the globe.

3. Phandango
This is a genome browser which, from a one-man effort, could be a game-changer. The UI needs a little bit of work I think, but otherwise, a really valuable tool for crunching lots of genomic data in a quick fashion.

4. i-PV Circos
This is a neat circular browser that helps users view protein sequences in a circularised format. With this visualisation format becoming more popular as the days go by, I think this has the potential to be a leader in the field. At the moment the website’s a bit dark and not the most user-friendly, but some of the core functionality (e.g. highlighting residues and association of domains) is a real plus!

5. Storyline visualisation
Possibly my favourite/eye-opener tool from the entire conference. Storyline visualisation helps users understand how things progress in realtime — this has been used for movie plot data (e.g. Star Wars character and plot progression) but the general concept can be useful for biological phenomena – for example, how do cells in diseased states progress over time? How does it compare to healthy states? Can we also monitor protein dynamics using a similar concept? I think the fact that it gives a very intuitive, big-picture overview of the micro-scale dynamics was the reason why I’ve been incredibly interested in Kwan-Liu Ma’s work, and I recommend checking out his website/publications list to grab insight on improving data visualisation (in particular, network visualisation when you want to avoid hairballs!)

The list isn’t ranked in any way, and do check these out! There were other tools I would’ve really liked to review (e.g. Minardo, made by David Ma @frostickle on Twitter), but I suppose I can go on and on. At the end of the day, visualisation tools like these are meant to be quick, and help us to not only EXPLORE our data, but to EXPLAIN it too. I think we’re incredibly fortunate to have some amazing minds out there who are willing to not only create these tools, but also make them available for all.