Aβ Binder Redesigns and the Round 2 Puzzle
In two previous blog posts (here and here), I wrote about the challenges of creating a protein to bind the Aβ polypeptide, a naturally-occurring polypeptide in the brain which self-assembles in Alzheimer's disease to produce toxic, aggregated forms that somehow kill brain cells. I've had a chance now to go through the designs that players created for the first Aβ Binder Redesign puzzle, and I've got to say that I'm very impressed. The goal was to redesign an existing, symmetric protein consisting of a pair of identical subunits in order to break the symmetry of the binding interface and optimally bind the Aβ molecule. The best designs that you produced achieved this beautifully – so congratulations!
We're still a few steps away from being able to test these designs in the wet lab, however. We first need to stabilize the structure a bit more with some more secondary structural elements, and then connect the two subunits to make one long chain. In the current puzzle, we would like to add a couple of additional β-strands to the β-sheet formed by the binder and the Aβ polypeptide. Ideally, these should pack well against the flanking helices, should hydrogen-bond perfectly to the adjacent β-strands, and should present core side-chains that pack without voids with the existing core (which you can redesign as needed). Here's a rough illustration of the backbone configuration that I'm hoping to see:
This puzzle is a prelude to a subsequent puzzle in which we will change the loop connectivity to make the binder into one long chain (instead of two separate chains) – which means that you need not worry much about the loops in this puzzle. We have chosen as starting points nine of our favourite designs from the first puzzle. Congratulations to the following players for their excellent designs:
|MurloW, wisky||Anthropic Dreams|
|andrewxc||Richard Dawkins Foundation|
The designs that we selected tended to be high-scoring designs, but this was not the only criterion that we considered. In particular, we paid close attention to good packing in the core and at the binder-Aβ interface (i.e. to the absence of deeply-buried voids). Small voids in the spaces between polar, surface-exposed residues are often less important, since water molecules can happily fill those.
This brings up a good point: the score assigned by the Foldit game engine is not a perfect representation of how “good” a structure is. This makes the game a little bit unfair for the players, but this isn't due to malice on our part: we'd like to make Foldit as fair a game as we possibly can! The truth is, improving the scoring is quite difficult, and it's an area of active research for the scientists who try to model protein folding computationally. We are constantly trying to develop better means of evaluating, mathematically, how likely it is that a particular computer model of a protein represents reality. The more reliable we make our scoring functions, the better we can predict in advance what's likely to fold up in the test tube, and the fairer Foldit becomes since the highest-scoring designs become the ones that are most likely to be good designs in reality. Because this is a very difficult problem that we have not fully solved, we still need to do a lot of manual checking to determine whether a design really is plausible or not, which means that, while players should use the score as a guide, there are structural features (like voids) that you should also pay close attention to.
(As an aside, scoring is one of two fundamental problems in the computational protein folding field. The other big area of research is devising algorithms to search for good structures given a protein sequence, or good protein sequences given a desired structure that we'd like to design. Just as computers were taught to play chess by emulating human players, we would like to solve this second problem by emulating the way in which the most skilled humans fold proteins by hand. This is the whole point of Foldit!)( Posted by v_mulligan 128 1800 | Tue, 11/19/2013 - 08:35 | 5 comments )
Fragment Filter and the Remix Tool
Foldit designs have come quite a long way, but there are several issues that we're still seeing in player-generated models.
One of the biggest problems is that some regions (primarily loops) in your models don't match the shape of regions in known natives. Since some shapes are more common than others, this means that these shapes are less statistically likely, and have a reduced probability to fold up in practice.
With that in mind, we've created...
The Fragment Filter!
So first - What is a fragment? A fragment is a shape for a small region of the protein. Imagine a loop of 8 residues. Loops can have drastically different shapes, and each of these shapes is a fragment of size 8.
The Fragment Filter tells you when part of your protein has an unrealistic fragment.
It does this by marching through the protein, looking at every fragment of size 4 (overlapping). For each of these fragments, it does a look-up into a database of known fragments that have been generated from a set of natives. If it cant find any fragments with shapes similar to the one you have in your protein, you'll get a score penalty, and it will highlight the fragment as a bad fragment.
The bad fragment is shown by the glowing red residues.
In this way, we ensure that the models you are generating are composed of statistically likely fragments, increasing the probability that they will fold up in the lab (and therefore the probability that we will test your design!).
From our analysis, top Foldit design solutions usually have 1-3 of these bad fragments.
So if your design has a bad fragment... how do you fix it?
The Remix Tool
Remix changes the shape of the selected region to match a 'good' fragment.
There are a couple of basic steps to using the Remix tool:
1. Select the region to remix.
Remix will work on selections of size 4-10. Even if there are only 4 bad residues, you can try larger sizes to try to get better results out of Remix. Just make sure the bad fragment is included in the Remixed region.
Here we've selected the bad fragment. In the Classic interface, Remix will work on secondary structure.
The Remix button will pop up when you select regions of size 4-10.
2. Remix the region!
When you hit Remix, the tool will do a lookup into a database and find a shape that has the same size and similar endpoints. Then it copies this shape onto your protein.
Remix wont always find something to copy (though it usually does), so you might have try selecting a different region, or modify the one you have until it finds something.
By using Remix, we've fixed the bad fragment, but this has created issues in the rest of the protein.
3. Resolve the resulting issues.
Similar to the Idealize tool, using Remix will create global changes even though you're just modifying a local area. This can cause issues elsewhere in the protein that you'll have to fix.
One way to fix them is to use low clashing importance and bands to stabilize the structure.
Another option (my favorite) is to insert a cut near one end of the region before Remixing, and then wiggle that region after Remixing until the cut can be closed. This prevents the global changes from ever happening, and usually works pretty well!
By using other game tools, we can bring the structure closer to its original shape, without the bad fragment.
Remember that you can fix the bad fragments any way that you want - Remix might just make it easier. It will be available on any puzzle that uses the Fragment Filter.
Like always, we'd love any feedback that you have on both the Fragment Filter and the Remix tool.( Posted by jflat06 128 1018 | Wed, 11/13/2013 - 22:02 | 5 comments )
Alzheimer's puzzle update from Baker Lab scientist Vikram Mulligan
In an earlier blog post, I wrote about the challenges of designing a protein therapeutic for Alzheimer's disease. Such a molecule must be able to bind with both high affinity and specificity to the neurotoxic Aβ polypeptide that causes the disease. Furthermore, the binding must prevent Aβ from self-associating to form neurotoxic protein oligomers or aggregates. In addition, a successful protein therapeutic must be stable in the body for a long time, evading clearance by the immune system and by proteases. It must also be able to cross the blood-brain barrier. Ideally, we would also like to add some functionality that would allow one binder molecule to clear or break down many Aβ molecules, though this will be an additional challenge. All of these are difficult problems to solve, but by breaking the overall goal into its component parts, we can begin to address these one at a time – and Foldit players can help!
The first puzzle that we posted involved the redesign of the core and interface residues of an existing, homodimeric Aβ binder in order to break the symmetry of the binding interface and improve both the affinity and the specificity for the asymmetric Aβ molecule. I am currently examining the designs that players produced in Puzzle 801b (many of which look quite impressive), and we will be revisiting these designs in the near future with a new puzzle. This puzzle will involve linking the secondary structural elements with new loop regions (which Foldit players will design) in order to convert the homodimer into a single large, monomeric protein. As we prepare the follow-up puzzle, we also want to get players started on a parallel strategy that we'll be exploring: complete de novo design of an Aβ binder.
In the new de novo design puzzle, we will be giving you the Aβ polypeptide in a rigid backbone conformation that has been observed in the complex with the affibody binder in NMR studies. We will also be giving you eight short, separate peptides (two of 24 residues, two of 20 residues, and four of 10 residues). We would like you to pretend that these short peptides are secondary structural elements of a single-chain protein, in which these peptides would be linked by loops – but we'll worry about the lengths, conformations, and sequences of the loops in a subsequent puzzle. For now, we'd like you to assemble these components around the Aβ polypeptide in such a way as to favor high-affinity, high-specificity binding. We are starting half of these peptides out as helices, and the other half in extended, strand-like conformations, but you should feel free to change the secondary structure as you see fit in order to achieve excellent binding. While the Foldit score should guide you, keep in mind that there are features that we are looking for that are not always captured well by the score alone. A successful design will have excellent shape-complementarity between all of the components – i.e. the surface of each element should have bumps and grooves that pack tightly against the surface of the next element. In addition, good designs should have buried hydrophobic residues making the core (phenylalanine, methionine, isoleucine, leucine, tyrosine, valine, and tryptophan). The surface should be non-hydrophobic. Most of the Aβ polypeptide's hydrophobic surface should be buried by the binder, both for purposes of specific recognition and to prevent Aβ from engaging in hydrophobic interactions with other, unbound Aβ molecules. Voids in the core of the binder or in the Aβ-binder interface are a bad thing – they're rarely seen in natural protein structures, and are energetically very unfavorable, though it is difficult for us to capture this properly in the mathematics of the Foldit score.
I'm also interested in your feedback about how this type of puzzle feels – in part because, if this turns out to be an effective means of designing a protein to bind a target, I'd like to try to develop automated algorithms to do the same, emulating the strategies used by the best Foldit players. Do you like having the freedom to move secondary structural elements independently, or is this too open-ended and unconstrained a puzzle? Does this open up new strategies that you would not have if you were dealing with a single long chain? Can you think of new Foldit manipulation tools that would make this type of puzzle easier for you? Would you like to see more design puzzles of this type, with follow-up loop design puzzles, or do you prefer the classic, “fold-a-long-chain” type of puzzles?
Please leave your comments below!( Posted by bkoep 128 1044 | Tue, 11/12/2013 - 00:09 | 10 comments )
The Baker Lab tests Foldit player-designed proteins
The Baker Lab has recently ordered materials to construct a batch of Foldit protein designs in the lab! We will express and purify the proteins from bacteria culture. For those proteins that we can successfully produce and isolate, we will conduct experiments to measure stability and solubility. If a designed protein is exceptionally well-behaved, we may be able to determine its native structure at atomic resolution.
Our process for selecting Foldit designs was as follows:
First, Baker Lab scientists visually inspected top-scoring and "Scientist-Shared" player designs for their favorite folding candidates. For the most part, these are well-ordered, contain significant secondary structure, have a well-packed and hydrophobic core, and also include enough complexity to demonstrate significant design.
These designs were then queued for ab initio folding predictions with the Rosetta@home distributed computing project. Computer power from volunteers around the world was used to generate hundreds of thousands of fold predictions for each design sequence. These results were then compiled into an "energy landscape" for each design sequence. For symmetric multi-chain designs, additional docking simulations were executed to sample alternative inter-molecular interactions. Only designs which show strong preference for the intended fold and docking scheme are included below.
In some cases, the prediction results suggested that a design could be significantly improved with minimal modification. Some of the designs to be tested are slightly modified versions of the player designs shown below. Please note that many solutions from past Foldit puzzles are still being processed.bkoep 128 1044 | Fri, 11/01/2013 - 06:28 | 2 comments )
Baker Lab scientist Vikram Mulligan describes how Foldit can help Alzheimer's disease research
Alzheimer's disease is an invariably fatal neurodegenerative disease that is now the seventh most common cause of death in Canada and the United States. Currently, one death in thirty-six is attributable to this disease. There is no cure for Alzheimer's, and existing treatments only provide temporary symptomatic relief, doing little to affect disease progression or longevity. Like most of the late-onset neurodegenerative dieases, Alzheimer's disease has protein misfolding and aggregation at its root. The Aβ polypeptide, a byproduct of protein cleavage that is produced continuously in the human brain and released into the extracellular space, self-associates to form toxic, soluble complexes (oligomers) and insoluble masses (aggregates). Although this self-assembly and aggregation process is poorly understood, one or more of the oligomeric or aggregated forms of Aβ produced along this pathway is toxic to neurons. Eventually, enough of this material accumulates to trigger massive neuronal loss in the brain, resulting in progressive dementia and, eventually, death.
Although small-molecule drugs are being developed to target downstream events in order to protect neurons, it would be better to target the root cause, if we could. Unfortunately, it is extremely difficult to prevent the self-association of large polypeptide chains into enormous aggregates using small molecules, since small molecules do not present enough surface area to compete with proteins that are prone to self-association. For this reason, we wish to develop protein therapeutics that can compete with self-assembly in order to target the Aβ aggregation cascade at the root of Alzheimer's disease.
Such a project will involve many challenges: getting the protein into the body and across the blood-brain barrier, having it evade the immune system, and ensuring that it persists long enough to have a desired effect will be just a few. A first step in this will involve the creation of a small protein able to bind tightly and specifically to monomeric Aβ, preventing Aβ self-association. Later stages will involve attempts to add functionality to the binder, in order to promote Aβ clearance or degradation.
As a starting point, we will be using an affibody molecule created in 2007 by a group at the Swedish Royal Institute of Technology (KTH) [1,2]. This molecule was created in a structure-agnostic manner, by screening a combinatorial library for Aβ binders, and was subsequently found to bind as a homodimer (i.e. two identical protein molecules bind a single Aβ molecule). Unfortunately, this means that a symmetric interface binds an asymmetric Aβ molecule, limiting the affinity and specificity of the binder for Aβ. We would like FoldIt players to redesign this two-chain molecule as one long polypeptide chain, breaking the symmetry of the interface and improving its affinity by separately optimizing both sides of the interface for the part of the Aβ molecule that each touches. The initial puzzle involves redesigning interface and core residues for optimal binding. Subsequent steps will involve designing new loops to connect the two subunits.
1. Grönwall C, Jonsson A, Lindström S, Gunneriusson E, Ståhl S, Herne N. (2007) Selection and characterization of Affibody ligands binding to Alzheimer amyloid beta peptides. J Biotechnol. 128(1):162-83.
2. Hoyer W, Grönwall C, Jonsson A, Ståhl S, Härd T. (2008) Stabilization of a beta-hairpin in monomeric Alzheimer's amyloid-beta peptide inhibits amyloid formation. Proc Natl Acad Sci U S A. 105(13):5099-104.