For Journal Club this week I decided to discuss the following paper by M. Rarey et al., which describes a method of using SMARTS patterns to discriminate between two sets of molecules. Link to paper here.

Given two sets of molecules can one generate a pattern that discriminates between two sets? This relates to a key question in drug design: can we predict whether molecules bind or not given a set of binders and a set of non-binders. The method is of particular interest because it makes use of data available, unlike conventional methods. However, for this technique to work, the correct molecular classification is required to discriminate between the two sets of molecules.

Originally molecules were classified using physiochemical properties for example, molecular weight or log P. However these classifications are too general and do not encompass enough molecular detail for accurate discrimination. An alternative is to using topological fingerprints. These encode a set the presence of a set of topological features using a series of bits. One of the limitations with this classification is that it is restricted by the predefined set of structures and features. This method makes use of chemical patterns which advantageously can can classify a chemical feature that cannot be sufficiently described by molecular substructure.

SMARTS (a molecular description language based on SMILES) allow description of structures with varying levels of specificity. For example one can specify atomic element, whether the atom is a subset of elements, whether it is aliphatic or aromatic, or whether it is in a ring. The method makes use of this description of molecules as the group have already developed some software to visualise SMARTS strings and modify them: the SMARTSeditor.

The method involves combining automatic pattern generation and visualisation to form SMARTSminer. Given two distinct molecule sets, the algorithm derives connected chemical patterns to differentiate both sets by using a sub-graph mining technique: solutions are extended by single elements iteratively.

The SMARTSminer was then used to test a series of test cases using the DUD (Database of Useful Decoys) data set. This seems strange when the data set has been shown to be inaccurate and perhaps there are more accurate test sets available, such as DUDe (Database of Useful Decoys enchanced). Let us look a couple of these case sets in more detail.

1. Discrimination between Active Molecules on Similar Targets

The first case set looks at discriminating between molecules that are active for COX-1 and COX-2. COX proteins are cyclooxygenase that are involved in inflammatory reaction. These proteins are targeted by inhibitors such as aspirin and ibuprofen for the relief of inflammatin and pain. Both COX-1 and COX-2 are similar targets with similar molecular weight and 65% sequence identity. Selective inhibition is only due to a difference in residue at position 523.

Separation of the sets of molecules was possible with a pattern identified that hit 21/25 of the molecules active for COX-1 and 15/348 of molecules of molecules active for COX-2. When the positive and negative set are reversed a pattern is identified that matched 313/348 of COX-2 actives but only 1 of the COX-1 ligands. The group state that perfect separation is not possible as there is an overlap of 2 molecules.

It is interesting that patterns were identified that could discriminate between the two sets. However, there is no discussion of how to use this information. Additionally the pattern determined has not been tested on any molecules outside of the training set – there are no blind tests. This seems strange as a blind test could emphasise the usefulness of this method if it was successful.

2. Discrimination between Active and Inactive Molecules

The second case investigates determining whether a pattern can be generated that discriminates between active and inactive targets. The test case used target SAHH (S-adenosyl-homocysteine hydrolase). A pattern was generated that matched all active molecules and only 1% of inactives. What is particularly exciting is that the pattern found contains part of the interaction network hydrogen bonding partners of the ligand, as shown in the figure below (the pattern identified is highlighted in green).

I find it very surprising that the group did not follow up with blind tests of molecules not used in the training set – especially as the pattern identified a key part of the binding mechanism.

To summarise a new method, SMARTSminer, calculates discriminative patterns between two sets of molecules using the SMARTS language. The authors state that the method has shown applicability in several use cases covering the application of actives vs decoys, kinase classifications, analysis of data sets and characterisation of reaction centers. However, I’m not sure I can agree with that statement. I believe further blind tests would be required to prove the applicability of the method once the pattern has been found. I also believe that an analysis of whether the pattern is over fitted to the training data is also required.

Hey Blopig Readers,

I had the privilege to go down to Heidelberg last week to go and see some stunning posters and artwork. I really recommend that you check some of the posters out. In particular, the “Green Fluorescent Protein” poster stuck out as my favourite. Also, if you’re a real Twitter geek, check out #Vizbi for some more tweets throughout the week.

So what did the conference entail? As a very blunt summary, it was really an eclectic collection of researchers around the globe who showcased their research with very neat visual media. While I was hoping for a conference that gave an overview of some of the principles that dictate how to visualise proteins, genes, etc., it wasn’t like that at all! Although I was initially a bit disappointed, it turned out to be better – one of the key themes that were re-iterated throughout the conference is that visualisations are dependent on the application!

From the week, these are the top 5 lessons I walked away with, and I hope you can integrate this into your own visualisation:

1. There is no pre-defined, accepted way of visualising data. Basically, every visualisation is tailored, has a specific purpose, so don’t try to fit your graph into something pretty that you’ve seen in another paper. We’re encouraged to get insight from others, but not necessarily replicate a graph.
2. KISS (Keep it simple, stupid!) Occam’s razor, KISS, whatever you want to call it – keep things simple. Making an overly complicated visualisation may backfire.
3. Remember your colours. Colour is probably one of the most powerful tools in our arsenal for making the most of a visualisation. Don’t ignore them, and make sure that they’re clean, separate, and interpretable — even to those who are colour-blind!
4. Visualisation is a means of exploration and explanation. Make lots, and lots of prototypes of data visuals. It will not only help you explore the underlying patterns in your data, but help you to develop the skills in explaining your data.
5. Don’t forget the people. Basically, a visualisation is really for a specific target audience, not for a machine. What you’re doing is to encourage connections, share knowledge, and create an experience so that people can learn your data.

I’ll come back in a few weeks’ time after reviewing some tools, stay tuned!

Why network comparison?

Many complex systems can be represented as networks, including friendships (e.g. Facebook), the World Wide Web trade relations and biological interactions. For a friendship network, for example, individuals are represented as nodes and an edge between two nodes represents a friendship. The study of networks has thus been a very active area of research in recent years, and, in particular, network comparison has become increasingly relevant. Network comparison, itself, has many wide-ranging applications, for example, comparing protein-protein interaction networks could lead to increased understanding of underlying biological processes. Network comparison can also be used to study the evolution of networks over time and for identifying sudden changes and shocks.

How do we compare networks?

There are numerous methods that can be used to compare networks, including alignment methods, fitting existing models,
global properties such as density of the network, and comparisons based on local structure. As a very simple example, one could base comparisons on a single summary statistic such as the number of triangles in each network. If there was a significant difference between these counts (relative to the number of nodes in each network) then we would conclude that the networks are different; for example, one may be a social network in which triangles are common – “friends of friends are friends”. However, this a very crude approach and is often not helpful to the problem of determining whether the two networks are similar. Real-world networks can be very large, are often deeply inhomogeneous and have multitude of properties, which makes the problem of network comparison very challenging.

A network comparison methodology: Netdis

Here, we describe a recently introduced network comparison methodology. At the heart of this methodology is a topology-based similarity measure between networks, Netdis [1]. The Netdis statistic assigns a value between 0 and 1 (close to 1 for very good matches between networks and close to 0 for similar networks) and, consequently, allows many networks to be compared simultaneously via their Netdis values.

The method

Let us now describe how the Netdis statistic is obtained and used for comparison of the networks G and H with n and m nodes respectively.

For a network G, pick a node i and obtain its two-step ego-network. That is, the network induced by collection of all nodes in G that are connected to i via a path containing at most two edges. By induced we mean that a edge is present in the two-step ego-network of i if and only if it is also present in the original network G. We then count the number of times that various subgraphs occur in the ego-network, which we denote by for subgraph w. For computational reasons, this is typically restricted to subgraphs on 5 or fewer nodes. This processes is repeated for all nodes in G, for fixed k=3,4,5.

1. Under an appropriately chosen null model, an expected value for the quantities is given, denoted by . We omit some of these details here, but the idea is to centre the quantities to remove background noise from an individual networks.
2. Under an appropriately chosen null model, an expected value for the quantities is given, denoted by .  We omit some of these details here, but the idea is to centre the quantities to remove background noise from an individual networks.
3. Calculate:  
4. To compare networks G and H, define:where A(k) is the set of all subgraphs on k nodes andis a normalising constant that ensures that the statistic  takes values between -1 and 1.  The corresponding Netdis statistic is:  which now takes values in the interval between 0 and 1.
5. The pairwise Netdis values from the equation above are then used to build a similarity matrix for all query networks. This can be done for any , but for computational reasons, this typically needs to be limited to . Note that for     k=3,4,5 we obtain three different distance matrices.
6. The performance of Netdis can be assessed by comparing the nearest neighbour    assignments of networks according to Netdis with a ‘ground truth’ or ‘reference’ clustering. A  network is said to have a correct nearest neighbour whenever its nearest neighbour according to Netdis is in the same cluster as the network itself.  The overall performance of   Netdis on a given data set can then be quantified using the nearest neighbour score (NN),   which for a given set of networks is defined to be the fraction of networks that are assigned correct nearest neighbours by Netdis.

Why Netdis?

The Netdis methodology has been shown to be effective at correctly clustering networks from a variety of data sets, including both model networks and real world networks, such Facebook. In particular, the methodology allowed for the correct phylogenetic tree for five species (human, yeast, fly, hpylori and ecoli) to be obtained from a Netdis comparison of their protein-protein interaction networks. Desirable properties of the Netdis methodology are the following:

\item The statistic is based on counts of small subgraphs (for example triangles) in local neighbourhoods of nodes. By taking into account a variety of subgraphs, we capture the topology more effectively than by just considering a single summary statistic (such as number of triangles). Also, by considering local neighbourhoods, rather than global summaries, we can often deal more effectively with inhomogeneous graphs.

• The Netdis statistic contains a centring by subtracting background expectations from a null model. This ensures that the statistic is not dominating by noise from individual networks.
• The statistic also contains a rescaling to ensure that counts of certain commonly represented subgraphs do not dominate the statistic. This also allows for effective comparison even when the networks we are comparing have a different number of nodes.
• The statistic is normalised to take values between 0 and 1 (close to 1 for very good matches between networks and close to 0 for similar networks). The statistic gives values between 0 and 1 and based on this number, we can simultaneously compare many networks; networks with Netdis value close to one can be clustered together. This offers the possibility of network phylogeny reconstruction.

A new variant of Netdis: subsampling

Despite the power of Netdis as an effective network comparison method, like many other network comparison methods, it can become computationally expensive for large networks. In such situations the following variant of Netdis may be preferable (see [2]). This variant works by only querying a small subsample of the nodes in each network. An analogous Netdis statistic is then computed based on subgraph counts in the two-step ego networks of the sampled nodes. From numerous simulation studies and experimentations, it has been shown that this statistic based on subsampling is almost as effective as Netdis provided that at least 5 percent of the nodes in each network are sampled, with the new statistic only really dropping off significantly when fewer than 1 percent of nodes are sampled. Remarkably, this procedure works well for inhomogeneous real-world networks, and not just for networks realised from classical homogeneous random graphs, in which case one would not be surprised that the procedure works.

Other network comparison methods

Finally, we note that Netdis is one of many network comparison methodologies present in the literature Other popular network comparison methodologies include GCD [3], GDDA [4], GHOST [5], MI-Graal [6] and NETAL [7].

[1]  Ali W., Rito, T., Reinert, G., Sun, F. and Deane, C. M. Alignment-free protein
interaction network comparison. Bioinformatics 30 (2014), pp. i430–i437.

[2] Ali, W., Wegner, A. E., Gaunt, R. E., Deane, C. M. and Reinert, G. Comparison of
large networks with sub-sampling strategies. Submitted, 2015.

[3] Yaveroglu, O. N., Malod-Dognin, N., Davis, D., Levnajic, Z., Janjic, V., Karapandza,
R., Stojmirovic, A. and Prˇzulj, N. Revealing the hidden language of complex networks. Scientific Reports 4 Article number: 4547, (2014)

[4] Przulj, N. Biological network comparison using graphlet degree distribution. Bioinformatics 23 (2007), pp. e177–e183.

[5] Patro, R. and Kingsford, C. Global network alignment using multiscale spectral
signatures. Bioinformatics 28 (2012), pp. 3105–3114.

[6] Kuchaiev, O. and Przulj, N. Integrative network alignment reveals large regions of
global network similarity in yeast and human. Bioinformatics 27 (2011), pp. 1390–
1396.

[7] Neyshabur, B., Khadem, A., Hashemifar, S. and Arab, S. S. NETAL: a new graph-
based method for global alignment of protein–protein interaction networks. Bioinformatics 27 (2013), pp. 1654–1662.

 

Multilayer networks are a generalisation of network that may incorporate different types of interactions [1]. This could be different time points in temporal data, measurements in different individuals or under different experimental conditions. Currently many measures and methods from monolayer networks are extended to be applicabile to multilayer networks. Those include measures of centrality [2], or methods that enable to find mesoscale structure in networks [3,4].

Examples of such mesoscale structure detection methods are stochastic block models and community detection. Both try to find groups of nodes that behave structurally similar in a network. In its most simplistic way you might think of two groups that are densely connected internally but only sparsely between the groups. For example two classes in a high school, there are many friendships in each class but only a small number between the classes. Often we are interested in how such patterns evolve with time. Here, the usage of multilayer community detection methods is fruitful.

Mucha et al. analysed the voting pattern in the U.S. Senate [4]. They find that the communities are oriented as the political party organisation. However, the restructuring of the political landscape over time is observable in the multilayered community structure. For example, the 37th Congress during the beginning of the American Civil War brought a major change in the voting patterns. Modern politics is dominated by a strong partition into Democrats and Republicans with third minor group that can be identified as the ‘Southern Democrats’ that had distinguishable voting patterns during the 1960.

Such multilayer community detection methods can be insightful for networks from other disciplines. For example they have been adopted to describe the reconfiguration in the human brain during learning [5]. Hopefully they will be able to give us insight in the structure and function of protein interaction.

[1] De Domenico, Manlio; Solé-Ribalta, Albert; Cozzo, Emanuele; Kivelä, Mikko; Moreno, Yamir; Porter, Mason A.; Gómez, Sergio; and Arenas, Alex [2013]. Mathematical Formulation of Multilayer Networks, Physical Review X, Vol. 3, No. 4: 041022.

[2] Taylor, Dane; Myers, Sean A.; Clauset, Aaron; Porter, Mason A.; and Mucha, Peter J. [2016]. Eigenvector-based Centrality Measures for Temporal Networks

[3]  Tiago P. Peixoto; Inferring the mesoscale structure of layered, edge-valued, and time-varying networks. Phys. Rev. E 92, 042807

[4] Mucha, Peter J.; Richardson, Thomas; Macon, Kevin; Porter, Mason A.; and Onnela, Jukka-Pekka [2010]. Community Structure in Time-Dependent, Multiscale, and Multiplex Networks, Science, Vol. 328, No. 5980: 876-878.

[5] Bassett, Danielle S.; Wymbs, Nicholas F.; Porter, Mason A.; Mucha, Peter J.; Carlson, Jean M.; and Grafton, Scott T. [2011]. Dynamic Reconfiguration of Human Brain Networks During Learning, Proceedings of the National Academy of Sciences of the United States of America, Vol. 118, No. 18: 7641-7646.

 

At the group meeting on the 3rd of February I presented the results of the paper “A General Method for Insertion of Functional Proteins within Proteins via Combinatorial Selection of Permissive Junctions” by Peng et. al. This is interesting to our group, and especially to me, because this is a novel way of designing an antibody, although I suspect that the scope of their research is much more general, their use of antibodies being a proof of concept.

Their premise is that the structure of a protein is essentially secondary structures and tertiary structure interconnected through junctions. As such it should be possible to interconnect regions from different proteins through junctions, and these regions should take up their native secondary and tertiary structures, thus preserving their functionality. The question is what is a suitable junction? This is important because these junctions should be flexible enough to allow the proper folding of the different regions, but also not too flexible as to have a negative impact on stability. There has been previous work done on trying to design suitable junctions, however the workflow presented in this paper is based on trying a vast number of junctions and then identifying which of them work.

As I said above their proof concept is antibodies. They used an antibody scaffold (the host), out of which they removed the H3 loop and then fused to it, using  junctions, two different proteins: Leptin and FSH (the guests). To identify the correct junctions they generated a library of antibodies with random three residues sequences on either side of the inserted protein plus a generic linker (GGGGS) that can be repeated up to three times.

They say that the theoretical size of the library is 10^9 (however I would say it is 9*20^6), and the actually achieved diversity of their library was of size 2.88*10^7 for Leptin and 1.09*10^7. Next step is to identify which junctions have allowed the guest protein to fold properly. For this they devised an autocrine-based selection method using engineered cells that have beta lactamase receptors which have either Leptin or FSH as agonists. A fluoroprobe in the cell responds to the presence of beta lactamase producing a blue color, instead of green and therefore this allows the cells with the active antibody-guest  designed protein (clone) to be identified using FRET-based fluorescence-activated cell sorting.

They managed to identify 6 clones that worked for Leptin and 3 that worked for FSH with the linkers being listed in the below table:

There does not seem to be a pattern emerging from those linker sequences, although one of them repeats itself. For my research it would have been interesting if a pattern did emerge, and then that could be used as a generic linker for future designers. However, this is still another prime example of how well an antibody scaffold can be used a starting point for protein engineering.

As a bonus they also tested in vivo how their designs work and they discovered that the antibody-leptin design (IgG-Leptin) has a longer lifetime. This is probably due to the fact that being a larger protein this is not filtered out by the kidneys.

At the meeting on February 10 I covered the article by Sormanni et al. describing a methodology for computationally designing antibodies against intrinsically disordered regions of proteins.

Antibodies are proteins that are a natural part of our immune system. For over 50 years lab-made antibodies have been used in a wide variety of therapeutic and diagnostic applications. Nowadays, we can design antibodies with high specificity and affinity for almost any target. Nevertheless, engineering antibodies against intrinsically disordered proteins remains costly and unreliable. Since as many as about 33.0% of all eukaryotic proteins could be intrinsically disordered, and the disordered proteins are often implicated in various ailments and diseases such methodology could prove invaluable.

The initial step in the protocol involves searching the PDB for protein sequences that interact in a beta strand with segments of the target sequence. Next, such peptides are joined together using a so-called “cascade method”. The cascade method starts with the longest found peptide and grows it to the length of the target sequence by joining it with other, partially overlapping peptides coming from beta strands of the same type (parallel, antiparallel). In the cascade method, all fragments used must form the same hydrogen bond pattern. The resulting complementary peptide is expected to “freeze” part of the discorded protein by forcing it to locally form a beta sheet. After the complementary peptide is designed, it is grafted on a single-domain antibody scaffold. This decision has been made as antibodies have a longer half-life and lower immunogenicity.

To test their method the authors initially assessed the robustness of their design protocol. First, they run the cascade method on three targets – a-synuclein, Aβ42 and IAPP. They found that more than 95% of the residue position in the three proteins could be targeted by their method. In addition, the mean number of available fragments per position was 570. They also estimated their coverage on a larger scale, using 1690 disordered protein sequences obtained from DisProt database and from measured NMR chemical shifts. About 90% of residue positions from DisProt and 85% positions from the chemical shift could be covered by at least one designed peptide. The positions that were hard to target usually contained Proline, in agreement with the known result that Prolines tend to disrupt secondary structure formation.

To test the quality of their designs the authors created complementary peptides for a-synuclein, Aβ42 and IAPP and grafted them on the CDR3 region of a human single domain antibody scaffold. All designs were highly stable and bound their targets with high specificity. Following the encouraging result the authors measured the affinity of one of their designs (one of the anti-a-synuclein antibodies). The K­_d was found to lie in the range 11-27 μM. Such affinity is too low for pharmaceutical purposes, but it is enough to prevent aggregation of the target protein.

As the last step in the project the authors attempted a two-peptide design, where a second peptide was grafted in the CDR2 region of the single-domain scaffold. Both peptides were designed to bind the same epitope. The two peptide design managed to reach the affinity required for pharmaceutical viability (affinity smaller than 185 nM with 95% confidence). Nevertheless, the two loop design became very unstable rendering it not viable for pharmaceutical purposes.

Overall, this study presents a very exciting step towards computationally designed antibodies targeting disordered epitopes and deepens out understanding of antibody functionality.

Some times real networks contain few nodes that are connected to a large portion of the nodes in the network. These nodes, often called ‘hubs’ (or global hubs), can change global properties of the network drastically, for example the length of the shortest path between two nodes can be significantly reduced by their presence.

The presence of hubs in real networks can be easily observed, for example, in flight networks airports such as Heathrow (UK) or Beijing capital IAP (China) have a very large number of incoming and outgoing flights in comparison to all other airports in the world. Now, if in addition to the network there is a partition of the nodes into different groups ‘local hubs’ can appear. For example, assume that the political division is a partition of the nodes (airports) into different countries. Then, some capital city airports can be local hubs as they have incoming and outgoing flights to most other airports in that same country. Note that a local hub might not be a global hub.

There are several ways to classify nodes based on different network properties. Take for example, hub nodes and non-hub nodes. One way to classify nodes as hub or non-hub uses the participation coefficient and the standardised within module degree (Gimera &  Amaral, 2005).

Consider a partition of the nodes into groups. Let be the degree of node and the number of links or edges to other nodes in the same group as node . Then, the participation coefficient of node  is:

.

Gimera &  Amaral (2005) proposed a classification of the nodes of the network based on their corresponding values of the previous statistics. In particular they proposed a heuristic classification of the nodes depicted by the following plot

Guimera and Amaral (2005), named regions R1-R4 as non-hub regions and R5-R7 as hub regions. Nodes belonging to: R1 are labelled as ultra-peripheral nodes, R2 as peripheral nodes, R3 as nun-hub connector nodes, R4 as non-hub kinless nodes, R5 as provincial nodes, R6 as connector hubs and R7 as kinless hubs. For more details on this categorisation please see Guimera and Amaral (2005).

The previous regions give an intuitive classification of network nodes according to their connectivity under a given partition of the nodes. In particular it gives an easy way to differentiate hub nodes of non-hub nodes. However the classification of the nodes into these seven regions (R1-R7) depends on the initial partition of the nodes.

First, happy new year to all our Blopig fans, and we all hope 2016 will be awesome!

A couple of months ago, I was covering this article by Shalom Rackovsky. The big question that jumps out of the paper is, has modelling reached its limits? Or, in other words, can bioinformatics techniques be used to model every protein? The author argues that protein structures have an inherent level of variability that cannot be fully captured by computational methods; thus, he raises some scepticism on what modelling can achieve. This isn’t entirely news; competitions such as CASP show that there’s still lots to work on in this field. This article takes a very interesting spin when Rackovsky uses a theoretical basis to justify his claim.

For a pair of proteins P and Q, Rackovsky defines their relationship depending on their sequence and structural identity. If P and Q share a high level of sequence identity but have little structural resemblance, P and Q are considered to be a conformational switch. Conversely, if P and Q share a low level of sequence identity but have high structural resemblance, they are considered to be remote homologues. 

From here on comes the complex maths. Rackovsky’s work here (and in papers prior, example) assume that there are periodicities in properties of proteins, and thus apply fourier transforms to compare protein sequences and structures.

In the case of comparing protein sequences, instead of treating sequences as a string of letters, protein sequences are characterised by an N x 10 matrix. N represents the number of amino acids in protein P (or Q), and each amino acid has 10 biophysical properties. The matrix then undergoes Fourier Transformation (FT), and the resulting sine and cosine coefficients for proteins P and Q are used to calculate the Euclidean distance between each other.

When comparing structures, proteins are initially truncated into length-L fragments, and the dihedral angle, bond length and bond angle for each fragment is collected into a matrix. The distribution of matrices allows us to project proteins onto a pre-parameterised principal components space. The Euclidean distance between the newly-projected proteins is then used to quantify protein structural similarity.

In both sequence and structure distances, the distances are normalised and centred around 0,0 by calculating the average distance between P and its M-nearest neighbours, and then adjusted by the global average. Effectively, if a protein has an average structural distance, it will tend toward 0,0.

The author uses a dataset of 12000 proteins from the CATH set to generate the following diagram; the Y-axis represents sequence similarity and the X-axis is the structural similarity. Since these axes are scaled to the mean, the closer you are to 0, it means you’re closer to the global average sequence or structure distance.

Rackovsky argues that, while the remote homologue and conformational switch seem like rare phenomena, it accounts for approximately ~50% of his dataset. Although he does account for the high density of proteins within 0,0, the paper does not clearly address the meaning of these new metrics. In other words, the author does not translate these values to something we’re more familiar with (e.g.RMSD, and sequence identity % for structural and sequence distance). Although the whole idea is that his methods are supposed to be an alignment-free method, it’s still difficult to draw relationships to what we already use as the gold standard in traditional protein structure prediction problems. Also, note that the structure distance spans between -0.1 and 0.1 units whereas sequence identity spans between -0.3 and 0.5. The differences in scale are also not covered – i.e., is a difference of 0.01 units an expected value for protein structure distance, and why are the jumps in protein structure distance so much smaller than jumps in sequence space?

The author makes more interesting observations in the dataset (e.g. α/β mixed proteins are more tolerant to mutations in comparison to α- or β-only proteins) but the observations are not discussed in depth. If α/β-mixed proteins are indeed more resilient to mutations, why is this the case? Conversely, if small mutations change α- or β-only proteins’ structures to make new folds, having any speculation on the underlying mechanism (e.g. maybe α-only proteins are only sensitive to radically different amino acid substitutions, such as ALA->ARG) will only help our prediction methods. Overall I had the impression that the author was a bit too pessimistic about what modelling can achieve. Though we definitely cannot model all proteins that are out there at present, I believe the surge of new sources of data (e.g. cryo-EM structures) will provide an alternative inference route for better prediction methods in the future.

Every time I talk about protein interaction networks, I put up a the nice little figure below. This figure suggests that experiments are done to detect which proteins interact with each other, and lines are drawn between them to create a network. Sounds pretty easy, right? It does, and it probably should, because otherwise you wouldn’t listen past the words “Yeast Two-Hybrid”. However, sometimes it’s important to hear about how it’s not at all easy, how there are compromises made, and how those compromises mean there are some inherent limitations. This post is one of those things it is important to listen to if you care about protein interaction networks (and if you don’t… they’re pretty cool, so don’t be a downer!).

So what’s wrong with that nice figure? Just to deal with the obvious: the colour scheme isn’t great… there’s a red protein and the interactions are also red… and then a yellow protein and red interactions aren’t that easy on the eyes. But you’re not here to judge my choice of colours (I hope… otherwise you’d be better off criticizing the graphs in this beast). You’re here to hear me rant about networks… much more fun ;). So here goes:

1. Not all interactions come from the same experiments, they investigate different “types” of binding.
2. Each experiment has experimental errors associated with it, the interactions are not all correct.
3. The network is not complete (estimation is that there are somewhere between 150k and 600k interactions depending on what paper you read).
4. People tend to investigate proteins that they know are associated with a disease, so more interactions are known for certain proteins and their neighbours resulting in an “inspection bias”.
5. That’s a stupid depiction of a network, it’s just a blue blob and you can’t see anything (hence termed “ridiculogram” by Mark Newman).
6. The arrow is wrong, you are not reporting interactions between proteins!

Points 1-5 should more or less make sense as they’re written. Point 6 however sounds a little cryptic. They’re called PROTEIN interaction networks, why would they not be interactions between proteins, you might ask.. and you’d be right in asking, because it’s really quite annoying. The problem is one of isoforms. The relation gene to protein is not 1-to-1. After a gene is transcribed into RNA, that piece of RNA is cut up and rearranged into mRNA, which is then translated into a protein. This rearranging process can occur in different ways so that a single gene may encode for different proteins, or as they are called, different protein isoforms. Testing for interactions between isoforms is difficult, so what tends to be done is that people test for interactions for THE ONE isoform of a protein to rule them all  (the “reference isoform”) and then report these interactions as interactions for the gene. Sneaky! What you end up seeing are interactions mostly tested between reference isoforms (or any that happened to be in the soup) and reported as interactions for the gene product.

So how much does it matter if we don’t know isoform interaction information? Are there even different interacting partners for different isoforms? Do they have different functions? Well… yes, they can have different interactions and different functions. As to whether they matter… according to Corominas et al that answer is also a resounding yes… or at least in Autism Spectrum Disorder (ASD) it is.

The paper is the result of a 5-year investigation which investigates isoform interactions and the effect of knowing them vs not knowing them on predicting candidate ASD genes of interest. And seeing as a bunch of people spent a lot of time on this stuff, it was definitely worth a read. Corominas et al found that in an ASD-related protein interaction network, there is a significant number of interactions that would not be found if only the reference isoform interactions were used. Furthermore, compared to a “high-quality” literature curated network, the generated isoform-resolved ASD network added a lot of interactions. They then went on to show that these additional interactions played an important role in the decision of which genes to prioritize as important “players in Autism”.

Should these results make us renounce the use of currently available non-isoform-resolved protein interaction networks lest we burn in the depths of bioinformatics hell? Well… probably not. While the paper is very interesting and shows the importance of isoforms, it does so in the context of Autism only. The paper itself states that ASD is a brain-related disease which is an environment known for many isoforms. In many cases, it will likely be the case that the “dominant isoform” is just that… dominant. Moreover, the results may sound a little stronger than they are. The literature curated network that was compared to, to say that this isoform-resolved network is really important, was quoted as being “high-quality”. It is likely that many of the isoform interactions would be included in lower quality networks, but they have simply not been as well-studied as dominant isoforms. Thus, their isoform-resolved network would just confirm lower quality interactions as high-quality ones. That being said, if you want to look at the specific mechanisms causing a phenotype, it is likely that isoform information will be necessary. It really depends on what you want to achieve.

Let’s say you’re playing Texas Hold’em poker and you have two pairs. You’d like to have that full house, but it’s elusive and you’re stuck with the hand you have when your opponent bids high. That’s the situation we were in with protein interaction networks: you know you’re missing something, but you don’t know how bad it is that you’re missing it. This paper addresses part of that problem. We now know that your opponent could have the flush, but possibly only if you’re in Vegas. If you only want to play in a local casino, you’ll likely be fine.

For my second short project I have developed Theox, molecular diversity software, to aid the selection of synthetically viable molecules from a subset of diverse molecules. The selection of molecules for synthesis is currently based on synthetic intuition. The developed software indicates whether the selection is an adequate representation of the initial dataset, or whether molecular diversity has been compromised. Theox plots the distribution of diversity indices for 10,000 randomly generated subsets of the same size as the chosen subset. The diversity index of the chosen subset can then be compared to the distributions, to determine whether the molecular diversity of the chosen subset is sufficient. The figure shows the distribution of the Tanimoto diversity indices with the diversity index of the subset of molecules shown in green.