Recent Advances in Computational Protein Design

Part 1: Protein Design by Computation

00:00:00.00 Hi, my name is Stephen Mayo. I'm a professor at Caltech,
00:00:07.17 and today I'm going to give a talk
00:00:09.08 entitled Recent Advances in Computation Protein Design.
00:00:13.01 The focus is going to be on designing
00:00:15.20 and testing proteins combinatorial libraries.
00:00:18.05 The first question, or the first thing you might ask
00:00:22.17 is why are we interested in computational protein design.
00:00:25.06 Well, fundamentally, we're interested in understanding
00:00:29.18 what the underlying physical principals
00:00:31.24 are that govern folding, stability, and function of proteins.
00:00:35.22 Typically, or historically, to address these sort of relationships,
00:00:40.06 we've used perturbation-based approaches.
00:00:44.05 For example, if you're interested in understanding how an enzyme works, a protein,
00:00:47.13 you might make mutations, perturbations in the active site of the enzyme
00:00:52.27 and then assess experimentally what those perturbations
00:00:56.18 did to the function of the protein.
00:00:58.24 What we're after ultimately is
00:01:01.01 using a design-based paradigm to study these relationships:
00:01:04.01 folding, stability, and function. And in a design-based paradigm,
00:01:09.00 what we're trying to do is to
00:01:10.06 first construct a mathematical model,
00:01:13.18 a computational model, of all the things
00:01:15.23 that we believe are true of proteins:
00:01:17.11 how they fold, how they achieve stability,
00:01:19.15 how they function.
00:01:20.26 And then within that computational construct,
00:01:23.21 framework, we want to then run a design calculation,
00:01:27.18 that is design new or novel proteins,
00:01:30.06 and ask the question experimentally,
00:01:32.04 are those proteins viable.
00:01:33.26 Do they fold? Are they stable?
00:01:35.20 Do they have function?
00:01:36.16 And if they have all these properties,
00:01:38.23 then we can conclude that our computational model,
00:01:41.09 our mathematical model, actually accurately captures the fundamental essence
00:01:45.03 of what a protein is.
00:01:46.05 On the other hand, if the experiments don't work,
00:01:49.26 then hopefully we can learn from those failures
00:01:52.10 to modify our models in a way that allows
00:01:56.12 us to go forward in terms of building new knowledge
00:02:00.00 into the model as we do designs and experimental validation.
00:02:05.20 As a chemist, ultimately, I'm interested in designing new systems,
00:02:11.17 engineering new systems.
00:02:13.14 And in particular, we're interested in things
00:02:16.03 like engineering nanobodies, biosensors,
00:02:18.23 industrial-grade enzymes, research reagents, etc.
00:02:22.18 All the way to using these sort of methods to design human therapeutics.
00:02:27.15 Okay, so in terms of an outline,
00:02:30.10 I'll give a fairly long introduction to computational protein design.
00:02:36.19 And then I'll talk about the specific topic of the presentation,
00:02:41.04 which is the development and testing of the methods
00:02:44.04 for designing protein libraries.
00:02:45.15 And then at the end, I'll give an example of
00:02:49.05 how these methods can be used to
00:02:52.26 design a research reagent, in this case a blue fluorescent protein
00:02:57.03 with enhanced optical properties.
00:02:59.27 Okay so before I start then in terms of protein design
00:03:05.15 in particular, let me just give you a brief overview of biology.
00:03:09.00 So everything you need to know about biology is actually on this slide.
00:03:12.00 Importantly, we start with DNA, and as you know,
00:03:17.20 DNA gets transcribed into RNA,
00:03:21.02 and then RNA gets translated into protein.
00:03:23.26 We're going to focus on the protein end of the spectrum.
00:03:27.21 A lot of the hype recently has been in genomics,
00:03:30.29 that is studies related to DNA,
00:03:33.23 but the real action ultimately is at the protein level,
00:03:36.28 because proteins are the things that actually do stuff in cells.
00:03:41.08 They are the molecular machines, they're involved in signal transduction,
00:03:44.23 in metabolism, in replication, in essentially
00:03:49.08 all aspects of cellular function.
00:03:51.06 So in general, proteins are linear polymers of amino acids.
00:03:56.28 The structure of the protein, and often time
00:04:01.19 the dynamical behavior of the protein, dictate the function of the molecule.
00:04:06.15 This important because we use this idea of
00:04:09.12 the relationship between structure and function
00:04:11.06 in our design effort.
00:04:13.10 As I said earlier, proteins are involved in essentially
00:04:17.29 all the cellular processes.
00:04:19.14 Across the bottom we have examples
00:04:21.19 of three different proteins, just sort of randomly selected.
00:04:26.04 On the left side is a protein, a small bacterial protein,
00:04:30.09 called protein G. And it is shown in full atom representation.
00:04:36.05 It is about 60 amino acids and
00:04:38.15 you can see it's a fairly complex structure even
00:04:40.09 though it's a relatively small protein.
00:04:41.22 In the middle is a protein roughly the size of 300 amino acids
00:04:45.25 shown in a ribbon diagram.
00:04:48.14 And this protein is an enzyme and
00:04:50.17 you can see if you look closely near the middle
00:04:52.00 of the molecule there will be a small molecule substrate.
00:04:56.00 This is around, as I said, 300 amino acids.
00:05:00.20 On the right side of the slide is a membrane-bound protein.
00:05:06.13 It is all helical and it resides in the membrane,
00:05:09.24 and the membrane will you know,
00:05:11.21 form a layer like this, and this thing would stick in the middle.
00:05:14.11 And you can also see that in the middle of that protein
00:05:16.14 there is a chromaphore that absorbs light
00:05:19.17 and is part of the function of the molecule.
00:05:23.06 Okay, so when I talk about protein design,
00:05:29.09 I often get questions that are related to protein folding.
00:05:32.26 So let me take a minute to distinguish between protein folding.
00:05:37.19 And there's protein folding prediction, and protein design,
00:05:41.28 which is often referred to as inverse folding.
00:05:44.15 And so in protein folding prediction,
00:05:47.08 we have an amino acid sequence
00:05:50.03 and the question is can we make a prediction about the structure
00:05:55.19 that the protein will fold to in solution.
00:05:58.28 Now we can think about this in terms of the diagram
00:06:01.29 on the bottom, we have sequence space.
00:06:05.20 Sequence space is large, and that blue dot
00:06:09.13 in that sequence space diagram is a representation of the amino acid
00:06:14.17 sequence of interest.
00:06:16.05 The mapping that we're interested in is going from sequence space
00:06:19.05 to structure space. And this is a difficult mapping.
00:06:27.12 One reason is that you're either correct or incorrect
00:06:32.01 in the sense of there's a single correct structure that this particular sequence will fold into.
00:06:36.22 On the other hand, protein design is essentially exactly the opposite problem.
00:06:42.14 We start with some structure that we're interested in,
00:06:46.26 and the question that we're asking is can we predict,
00:06:50.02 in our case compute, an amino acid sequence which will fold to that structure
00:06:55.03 if we put it into solution in the laboratory.
00:06:58.18 And so the mapping then is exactly the opposite.
00:07:00.01 We're going from a point in structure space to
00:07:04.14 a set of sequences in sequence space.
00:07:08.23 Now you'll notice that in the inverse direction,
00:07:12.05 there will be many solutions in sequence space
00:07:15.08 that satisfy our target structure. And this is degeneracy
00:07:19.06 , that is this one to many mapping, aids us in the protein design process.
00:07:25.00 So protein design, fundamentally, is a much more straight forward endeavor
00:07:31.01 than protein fold prediction because the idea
00:07:34.16 that there are many solutions which actually lead to the correct answer,
00:07:38.22 that is the correct structure in solution.
00:07:41.04 Library design, protein library design, is related.
00:07:45.18 So in this case, rather than trying to identify some single optimal sequence
00:07:50.18 for our target structure, we're going to again use the same mapping procedure,
00:07:54.26 but we're going to now specify an entire list of sequences,
00:08:00.06 all of which are predicted to fold to the correct structure.
00:08:03.28 And then in the laboratory, we can actually take this list of sequences,
00:08:08.10 this sequence alignment, and manifest it as a nucleic acid library,
00:08:13.07 as a protein library, and then screen or run a selection on that library
00:08:17.20 to find those examples which actually satisfy our design objective.
00:08:23.03 Folding to the correct structure, and ultimately having some desired functional characteristic.
00:08:29.01 Okay so why do this computationally?
00:08:32.28 That can be shown fairly easily on this slide.
00:08:36.12 If we have a single protein, comprised of p residues,
00:08:41.02 and as you know there are 20 naturally occurring amino acid types.
00:08:45.11 So for a single protein, there are on an order of 20 to the p different sequence combinations.
00:08:52.02 So there are lots of ways that you can actually arrange amino acid sequences
00:08:55.17 for some given length of protein.
00:08:58.02 So let's think about this then.
00:09:01.01 If we want to design a protein that is comprised of only 18 amino acids,
00:09:05.27 18 residues, the total number of combinations is on the order of 10^23 different solutions,
00:09:13.16 different arrangements.
00:09:15.13 So if we think about that from an experimental perspective,
00:09:17.18 if we wanted to synthesize, chemically synthesize,
00:09:21.02 one example of each of the 10^23 different sequence combinations,
00:09:25.13 we would need the mass of baseball, which is shown on the top.
00:09:28.19 And that's actually almost tractable. You can almost imagine
00:09:32.21 sending a graduate student into the laboratory to do peptide synthesis
00:09:35.21 and synthesize enough material that's roughly the size of a baseball.
00:09:39.27 I wouldn't want to do it but maybe someone would.
00:09:42.24 But this gets out of hand pretty quickly.
00:09:45.24 And so by the time you're at 37 amino acids, which is still a trivial-sized protein,
00:09:49.25 you're at 10^48 combinations, and the equivalent mass then would be the mass of the earth.
00:09:56.11 And then by 59 amino acids, which is still very small.
00:10:00.02 You know, most functional proteins are probably on the order of 100s of amino acids long.
00:10:04.23 At 59 amino acids, you're at 10^77 different sequence combinations,
00:10:09.14 and you're at the mass of the universe.
00:10:12.01 So clearly experimentally, there aren’t methods
00:10:15.04 that you can use in the laboratory that in a single shot,
00:10:18.12 will allow you to explore the full sequence diversity
00:10:21.22 available to even the small proteins.
00:10:23.20 Computationally, however, this is actually a very simple problem,
00:10:27.11 in terms of the numbers, at least.
00:10:29.04 So we can easily think about methods,
00:10:31.29 combinatorial optimization methods, which will handle numbers
00:10:34.27 like the astronomical sizes that we see here.
00:10:39.18 And then the real challenge is can we come up with the physics,
00:10:45.00 quote unquote, that describes a protein well enough that when we do the combinatorial optimization
00:10:48.27 on the computer, we end up with sequences that will in fact be viable in the laboratory.
00:10:53.11 Okay so over the years we focused on both developing methodologies
00:10:58.20 for computational protein design and on applying these methods
00:11:03.15 in the laboratory to various types of design problems.
00:11:07.09 Most computational protein design methodologies have these set of features in common.
00:11:13.20 First of all we have to have some representation of a protein backbone
00:11:17.23 that goes into the calculation.
00:11:20.29 Typically the backbone structures are derived from the protein structure database.
00:11:27.05 So for most of what we do and most of what others do,
00:11:29.27 the designs are based on the backbone of proteins that are known to exists,
00:11:35.09 whose structures have been solved either by crystallography or by NMR.
00:11:38.04 Importantly, when we think about amino acid side chains, amino acid side chains
00:11:43.13 have conformational flexibility, so a particular conformation state
00:11:47.19 of an amino acid side chain is call a rotamer, and we have to be able to describe
00:11:53.04 this conformational flexibility in terms of the calculation.
00:11:56.24 So we do that by building rotamer libraries that are again
00:12:01.02 derived from the known protein structure database.
00:12:05.22 So the rotamer library capture at some level the
00:12:10.00 conformational flexibility of the amino acids.
00:12:12.09 Now importantly, in terms of capturing
00:12:14.28 the underlying physics of the problem, chemical physics of the problem,
00:12:18.10 we have to specify atom based force fields.
00:12:23.25 And it's here that we actually are building our model.
00:12:27.24 And I'll show you more about that in a minute.
00:12:30.16 So these aren't force fields like on Star Trek
00:12:33.26 , these are simple mathematical equations that describe
00:12:37.22 the interactions between sets of atoms.
00:12:39.11 And how we write those equations down and the parameters
00:12:43.11 that go into those equations are ultimately the important aspect
00:12:47.09 of our ability to do this design.
00:12:49.14 Along with how we're treating side chain flexibility through the rotamer library
00:12:54.03 and how we are treating or not treating issues of how the protein backbone
00:12:57.22 may relax as the design progresses.
00:13:00.00 With those three elements then, we can move to more
00:13:05.28 computer science oriented issues, and that's where the combinatorial optimization algorithms
00:13:10.05 come into play.
00:13:11.14 So we can easily construct a combinatorial problem for design,
00:13:15.03 and employ either standard or improved or novel methods
00:13:20.11 for doing those sort of optimization calculations.
00:13:24.03 And so that's really the combination of straight-forward computer science and applied math,
00:13:29.04 but applied to the problem of protein design.
00:13:33.15 And then ultimately there's this issue of negative design.
00:13:36.11 For many protein design problems, it's not
00:13:41.28 sufficient to focus on achieving what you want, the positive design results:
00:13:49.02 I want to build a protein that does x.
00:13:51.27 You also have to, in many cases, consider
00:13:55.29 what you don't want, that is you must be able to design against the
00:14:00.01 bad outcomes. So if you're thinking about building a bridge,
00:14:03.10 for example, clearly you want to do a positive design aspect,
00:14:06.16 which is to build a bridge so that cars can drive past it,
00:14:09.11 but you also have to consider building a bridge to guard against
00:14:15.10 negative consequences, like what happens if there's a high wind,
00:14:18.06 you don't want that bridge to fall down.
00:14:19.27 So protein design is similar in that regard.
00:14:21.24 There are many types of protein design problems where an element
00:14:26.14 of negative design is actually critical.
00:14:30.07 So for example, designing against aggregation is an important feature.
00:14:34.25 Designing against your amino acid sequence folding to some alternative
00:14:40.16 structure that you don't want.
00:14:42.20 So there are ways, there are conceptual ways
00:14:45.05 and pragmatic ways to actually incorporate negative design into your calculation.
00:14:50.14 Okay so with that collection of methodology,
00:14:54.03 we've been able to write software that actually implements this and to use that software then to think about
00:15:03.19 design problems that are related to understanding sequence structure, stability and
00:15:10.21 functional relationships, the evolution of protein structure and function,
00:15:14.02 we can do things that are related to hypothesis-driven inquiry, and often
00:15:20.27 we're interested in using these methods to enable a new protein based biotechnology.
00:15:26.06 Okay so here are just some brief examples of things that we've done over the years.
00:15:31.23 So back in 1997, we showed for the first time that you can actually
00:15:36.07 do all this stuff and it actually works.
00:15:39.08 So we selected a small protein fragment, stripped off all the side chains,
00:15:43.24 ran our design calculations and came up with a new sequence,
00:15:46.27 and then in the laboratory showed that that new protein sequence
00:15:50.06 actually folded and had the same structure as the original design target.
00:15:55.03 And so you can see that in red is the design target,
00:16:03.22 and in blue is the structure of the actual design protein.
00:16:07.10 On the bottom left, you can see that to the extent that the atom-based
00:16:16.27 force fields are actually accurate, and the extent of the
00:16:20.28 combinatorial optimization algorithms are giving us amino acid sequences
00:16:24.00 that are optimal for their target fold,
00:16:26.27 that is that are thermodynamically optimal.
00:16:29.04 We should be able to design sequences or sequence variants
00:16:33.02 of naturally occurring proteins that are hyperthermal-stabilized for their folds.
00:16:37.23 And so on the bottom left is such an example.
00:16:40.12 We took a small bacterial protein and showed that
00:16:45.02 through computation that we can design a set of mutations that turn that protein
00:16:49.15 from being a normal mesophilic protein with a normal melting temperature
00:16:53.00 to a protein that is now hyperthermal-stable with a melting temperature
00:16:56.01 in excess of 100 degrees Celsius.
00:16:57.27 Also in terms of function, we can see that it's easy
00:17:04.23 to imagine ideas of transition state stabilization being
00:17:09.17 folded into these methods and then doing optimizations with substrates
00:17:15.20 and transition states in the calculation,
00:17:18.06 and hence designing an enzyme.
00:17:19.26 So on the lower right, we have an example where we took a non-catalytically active protein
00:17:26.19 and then designed a very primitive enzyme by building
00:17:32.20 a very primitive active site, and this molecule we refer to as a protozyme,
00:17:37.26 it's not a real enzyme, because the kinetic parameters aren't as good as a real enzyme,
00:17:41.22 but mechanistically shows all the standard features of real enzymes:
00:17:46.06 saturation kinetics, it can be inhibited, we can
00:17:50.05 even see by mass spectrometry analysis the intermediate in the reaction.
00:17:56.00 Okay, so in general when we think about doing a design,
00:18:02.12 we refer to this process flow chart.
00:18:06.14 First thing we obviously want to do is specify the design objective.
00:18:09.26 Once we have a design objective in mind, we can then employ
00:18:15.11 our computational design software, which in our case is called ORBIT,
00:18:20.05 which stands for Optimization of Rotamers by Iterative Techniques.
00:18:22.20 The output of that calculation is an amino acid sequence,
00:18:27.01 or a list of amino acid sequences,
00:18:28.21 and then we can actually take that list of amino acid sequences
00:18:31.25 and then use standard molecular biology techniques, gene assembly,
00:18:36.06 and expression in our favorite host, we typically use E. coli
00:18:40.20 but have used other hosts when appropriate.
00:18:43.11 Once that protein is expressed, we can isolate that protein
00:18:46.02 and purify it using standard techniques, and then evaluate the properties
00:18:49.21 of that protein using standard biophysical approaches.
00:18:51.21 It's at that second to last step that we learn whether
00:18:57.05 or not we've been successful in our design.
00:19:01.03 If not, then we can learn hopefully quantitatively where the failings are
00:19:05.21 and then iterate on the design process to hopefully get closer
00:19:11.06 to our design objective as we go multiple rounds into the process.
00:19:15.26 Okay, so let me just take you through a design calculation,
00:19:19.12 a very simple design calculation to give you an idea of what we're actually computing.
00:19:22.13 So this is a case where we have on the lower left,
00:19:26.22 a model of a protein.
00:19:30.09 Maybe it's a protein with two alpha helices.
00:19:33.01 And we're going to select two positions out of this protein
00:19:36.14 for which we are going to do a design.
00:19:38.05 Position P1 and position P2.
00:19:40.10 The first thing we have to do is specify the list of amino acids
00:19:44.26 that we're going to allow at each of these positions
00:19:47.11 in the design calculation.
00:19:48.20 So at position P1, we've selected alanine, valine,
00:19:52.14 and serine as the allowed amino acids.
00:19:55.10 But you'll notice that for example for the valine column,
00:19:57.21 for the valine group, there are three different types of valines
00:20:01.16 with the subscripts 1, 2, and 3.
00:20:04.16 And these are the 3 different rotameric states of valine.
00:20:06.28 So as you'll recognize chemically, valine has a single relevant rotatable bond
00:20:14.20 in the side chain, and it can exist in 3 energetically stable states,
00:20:19.11 3 different rotamer states.
00:20:21.17 And then serine, for this simple example, we're also using 3 different rotamers.
00:20:26.12 For position P2, we're selecting the same amino acids
00:20:30.15 and allowed rotamers to keep the calculations simple.
00:20:33.13 Now for a normal design calculation, there might be 10s to 100s of design positions,
00:20:39.02 and 100s to 1000s of rotamers per design position.
00:20:43.16 So what I'm hoping to convey here is that this is a simple example,
00:20:48.05 but the real example is much larger, real examples are much larger
00:20:53.20 and much more computationally challenging than what I'll show right now.
00:20:59.03 Ok and so there are 2 types of energies that have to be computed.
00:21:03.09 The first is what we often refer to as the one body energy,
00:21:06.25 and that is the energy of interaction between an amino acid side chain,
00:21:11.16 a rotamer, and the rest of the protein backbone.
00:21:14.04 And then as you'll see momentarily, there will be a
00:21:18.21 second type of energy, called the two body
00:21:20.21 interaction, that will be interactions between pairs of rotamers, pairs of side chains.
00:21:25.17 Okay, so the first thing we do then is we have to build
00:21:31.09 the rotamer onto the backbone. So we grab the first rotamer,
00:21:37.04 alanine, which has a single rotamer, and we put it on the protein backbone,
00:21:40.23 shown as the yellow ball.
00:21:42.10 And then we have to compute its interaction energy,
00:21:44.26 the interaction energy between the atoms of the alanine side chain
00:21:49.09 and the rest of the protein backbone.
00:21:50.29 And the way we compute the energy is by evaluating the potential functions
00:21:57.12 that are shown, the types of potential functions that are shown on the slide.
00:22:02.09 So the first potential function is a Vander Waals potential.
00:22:07.13 It scores the interactions between atoms by basically measuring their distance.
00:22:13.07 If atoms get too close together, the energy skyrockets
00:22:16.21 as shown on the graph.
00:22:18.19 The x-axis there is distance between atoms,
00:22:21.09 as they get closer together the energy goes positive and large,
00:22:24.15 which is bad, you always want your energies to be negative,
00:22:27.16 you want them to be as small as possible.
00:22:30.22 And however, if atoms are at some appropriate distance, than you get an
00:22:39.27 attractive energy, which is shown as a minimum in this plot.
00:22:43.11 Okay so that's the Vander Waals potential function, it basically scores steric interactions.
00:22:48.15 There are other types of potential functions.
00:22:51.10 Here is a simple electrostatic potential, based on Coulomb’s law,
00:22:54.12 and in the diagram, you see that there is an amino acid
00:22:59.10 with the blue tip is a lysine, which carries a positive charge,
00:23:02.25 and the amino acid with the forked red tips is either
00:23:09.21 an aspartic acid or a glutamic acid, carrying a negative charge,
00:23:12.18 and so there will be an electrostatic attraction between those two amino acids.
00:23:17.11 So we can score that, in this cause, using a simple Coulombic potential.
00:23:20.12 There's also hydrogen bonding potential, so certain polar amino acids will form strong interactions
00:23:28.16 that are primarily electrostatic in nature but also have a strong directional component.
00:23:33.25 We can capture that with that potential function.
00:23:36.04 And then in addition, we often use a solvation potential where
00:23:41.06 the solvation terms are divided into various groups.
00:23:47.14 We want to benefit the burial of hydrophobic amino acids,
00:23:53.24 that is benefit hydrophobic surface area burial.
00:23:56.06 We want to penalize, this is actual a negative design term, we want to penalize
00:24:02.10 the exposure of hydrophobic surface area. If we have too much
00:24:05.02 exposed hydrophobic surface area in our designs, then the proteins
00:24:08.14 will most likely aggregate, that is an alternative state that we don't like.
00:24:13.26 We also want to penalize the burial of polar surface area.
00:24:18.06 Charged amino acids and polar amino acids don't like to be
00:24:22.10 stuffed in a hydrophobic core of a protein, they rather be exposed to solvent
00:24:25.04 or interacting with other polar amino acids.
00:24:27.07 And that is sort of captured at least schematically in that last equation.
00:24:33.11 Okay then for all the rotamers, at each position,
00:24:39.27 one at a time, we evaluate these potential functions and then
00:24:44.11 we store the energies in this vector, this one body energy array.
00:24:49.10 Okay so then the rest of this we'll then go through and one at a time
00:24:54.06 you'll see the different amino acids with their individual rotamers getting
00:24:57.11 loaded onto the protein structure, and then the
00:24:59.19 calculation being done, and you'll see now that it switches to position P2, and then
00:25:05.08 its energies are also being stored in this array, this one dimensional array.
00:25:09.18 Okay so now we've actually computed all the one body energies for the system.
00:25:14.08 The next step is to compute the two body energies, and that's then illustrated here.
00:25:20.24 So again, for a normal calculation, we'd have many more positions,
00:25:26.09 so this matrix would actually be very large. Here we have a single pair
00:25:29.23 of positions, because there are only 2 positions in the calculation.
00:25:33.04 And for the first position, you'll notice I
00:25:36.25 have all the amino acids and the rotamers down the side, and for
00:25:42.14 the second position, all the amino acids and its relevant
00:25:47.15 rotamers are across the top.
00:25:49.16 And so I'm actually going to do the same sort of calculations,
00:25:52.10 but now the protein backbone is not important.
00:25:55.08 So the protein backbone sort of fades away and is not a part of
00:25:57.25 the calculation. So at position P1, I select the first choice alanine,
00:26:03.09 number 1, and at position P2 I select the second choice alanine, number 1 as well.
00:26:07.14 And I compute the interaction energy.
00:26:09.24 Now I'm just looking at the interaction energy between the atoms in the first rotamer
00:26:13.01 and the second rotamer. And I store that value
00:26:15.15 in the matrix. I'm using the same potential function,
00:26:19.07 the same set of potential functions that I used for the one body interactions.
00:26:23.17 And then I can then step through this and compute all the pairwise interactions
00:26:30.08 in the system. And eventually I will fill up the entire matrix.
00:26:33.24 So now I've computed all the energies. I've got a list of one body energies
00:26:38.02 and a list of two body energies, and now is where I can actually do a calculation,
00:26:43.01 do an optimization calculation. So for the optimization calculation,
00:26:46.25 the protein structure goes away, we no longer have to worry about the protein
00:26:50.04 backbone. In fact, in the simplest cases, we don't even care that this is a protein anymore.
00:26:56.22 All we actually have are a bunch of numbers that we're trying to optimize.
00:26:59.26 We're trying to march through this matrix,
00:27:02.15 and find a combination of numbers that gives us the lowest energy.
00:27:06.15 And so what I've done for this example is I've retained the
00:27:09.29 two body matrix intact, and I have taken the one body interaction vector
00:27:14.24 and I've broken it into two pieces. So down the side is the
00:27:20.10 piece for just the first position, and across the top, are the energy
00:27:28.10 values for the rotamer in the second position. So in this particular optimization,
00:27:34.01 I'm going to employ a technique called Monte Carlo simulations,
00:27:36.18 which is based on effectively random selections of potential solutions
00:27:43.06 and scoring those choices and then making decisions about whether
00:27:48.03 we want to keep or reject various changes.
00:27:50.09 And ultimately we'll be able to sample the matrix in a way that
00:27:54.12 allows us to find solutions which are low in energy. So the first thing
00:27:57.17 which we have to do is initialize a search.
00:27:59.26 So I have two positions that I'm worried about, position P1 and P2.
00:28:04.16 So I randomly select a rotamer for position P1, the second serine in this case.
00:28:10.15 I randomly select a rotamer for position P2, the first valine.
00:28:15.08 I grab the two one body energies and the single two body energy
00:28:23.06 and I just sum them. And the sum of those three numbers
00:28:25.12 is 21.8. So that's our initial energy.
00:28:28.10 Then I make random changes to the system, keeping track of what I'm doing,
00:28:32.23 and eventually via this Monte Carlo process, I'll end up finding the best solution.
00:28:38.11 Which in this case is the total energy of -9 kcal/mol and it turns out to be the third
00:28:45.12 serine rotamer on position P1 and the second serine rotamer on position P2.
00:28:51.00 So maybe they're on the helix making a hydrogen bond.
00:28:53.12 So once I've actually identified what the solution is, I can map that back onto the structure.
00:28:59.28 So now I've got a model, a sort of atomic detailed model of the result of the calculation.
00:29:09.11 And I can use that for other purposes. I can look at it,
00:29:12.05 I can use my intuition and make evaluations about whether or not this is a reasonable solution.
00:29:17.05 I can take that sequence then and pass it on to someone in the laboratory
00:29:22.09 who might actually make that protein. Okay, now the Monte Carlo simulation is actually very easy,
00:29:32.04 it's perhaps the simplest way of doing these optimizations. For many real world design
00:29:39.19 calculations, Monte Carlo is not necessarily sufficient, and there are more
00:29:43.21 sophisticated optimization methods that can be used to do protein design,
00:29:49.05 some of which are listed on the lower right.
00:29:52.09 The method that we prefer is a method called FASTER, which is fast, actually.
00:30:00.18 And very effective at finding low energy solutions for these sort of design problems.
00:30:06.18 Okay so here's other recent results that have been reported from different labs.
00:30:12.28 On the upper left is a case from the Hellinga lab at Duke,
00:30:18.04 where they've been able to use these sort of methods to design biosensors
00:30:22.12 for small molecules. In the middle is another example from that lab
00:30:26.11 where they were able to design enzymatic functionality into a protein.
00:30:33.06 On the upper right is an example from the Baker lab at the University of Washington,
00:30:38.18 where they were able to actually design a novel fold, which was an impressive result.
00:30:45.19 On the lower part of the slide is a protein therapeutics result from a company called Xencor,
00:30:51.20 which I was involved in co-founding. In this particular case,
00:30:56.14 we were able to design a variant of a protein
00:31:02.28 that might be useful for anti-inflammatory indications.

Part 2: Designing Protein Libraries and an Example of Making a Better Fluorescent Protein

00:00:04.09 So let me now move on to the main topic of the talk,
00:00:09.15 which is designing protein libraries.
00:00:13.08 So why design a protein library?
00:00:15.10 Well, of course what we're interested in in terms of design in general
00:00:18.09 is an ability to either generate protein variants that have enhanced properties
00:00:25.11 or brand new novel properties.
00:00:28.17 The issue, however, is that it's difficult to design for function.
00:00:35.12 The lack of dramatic advance in enzyme design (computational enzyme design)
00:00:41.00 is one indication of that.
00:00:43.10 It's really hard to capture details of chemical functionality
00:00:46.11 in these sorts of design calculations for many different reasons.
00:00:50.09 And so, perhaps, if we design a library of proteins,
00:00:55.01 we might be able to employ an experimental selection or screen
00:00:59.19 to find those variants that are actually useful for us in some particular way.
00:01:09.25 So we want to compensate for the inability of the calculation to give us exactly what we want
00:01:19.05 by designing a library.
00:01:20.21 And the idea in general is that
00:01:23.05 perhaps the calculation will put us in the right part of sequence space.
00:01:27.13 And if we screen that identified portion of sequence space,
00:01:30.20 we might find the needle in the haystack that we're looking for in terms of functionality.
00:01:36.02 Okay, so there are really two challenges when you think about designing a protein library.
00:01:42.05 One challenge is that you want to have a library that has,
00:01:46.07 for a given size of library -- 1,000 members, 10,000 members, whatever it is --
00:01:52.17 you want the members of the library to have a large diversity of function.
00:01:55.24 That is, for example, if you're trying to take an existing enzyme
00:01:59.25 and you want to change the active site around in order to
00:02:03.12 find variants that have new enzymatic functionality,
00:02:06.23 maybe these molecules might act on different substrates compared to the wild type substrate,
00:02:13.14 you want the members of the library to show a broad range of functionality
00:02:17.07 because if they all have the exact functionality of the wild type protein,
00:02:20.23 then of course you haven't done anything to the molecule.
00:02:23.15 Everything is the same, and it's not really useful.
00:02:26.13 So the way you typically go for high diversity of function is
00:02:32.17 use a high mutation rate.
00:02:35.14 That will give you lots of variants and lots of variants
00:02:37.18 tends to imply lots of diversity of function.
00:02:41.18 The other property of your library that you want to have
00:02:45.05 is that you want to have a preservation of function.
00:02:48.03 So of course if you're designing a library of enzyme variants,
00:02:52.11 you want all the variants to actually still have enzymatic capability.
00:02:55.29 And the way you typically actually achieve that is by using a low mutation rate.
00:02:58.29 So the problem is that these things are opposed to each other.
00:03:02.29 If you want to have high diversity in your library by,
00:03:05.21 for example, using a high mutation rate,
00:03:08.08 you invariably get low preservation of function
00:03:11.16 because high mutation rate typically means that you're going to eradicate function,
00:03:16.21 maybe compromise the stability of the protein, maybe the protein won't even fold anymore.
00:03:20.29 On the other hand, if you want to have high preservation of function,
00:03:24.29 because you don't want to be screening unfolded proteins,
00:03:28.25 you might want to use a low mutation rate,
00:03:31.10 but that then will drive down your diversity of function.
00:03:35.09 Traditional methods for making combinatorial libraries of proteins sort of end up at one extreme
00:03:44.21 of either diversity of function. They go in opposition.
00:03:48.09 What we want to do computationally, however, is
00:03:52.19 come up with a method that allows both high diversity of function (high mutation rate that is)
00:03:57.21 and high preservation of function.
00:03:59.03 And the way we might be able to achieve that is by letting the design calculation select what
00:04:06.22 large number of amino acids are allowable at the design positions,
00:04:11.02 which will give us high diversity of function because we'll using a high mutation rate,
00:04:15.21 and if the calculation chooses a set of amino acids that are going to be stabilizing
00:04:19.03 for the protein at those target/design positions,
00:04:22.14 then we should also be able to have high preservation of function.
00:04:25.14 Now, the problem is that I just told you that these sort of calculations
00:04:29.06 are poor when it comes to designing for functionality.
00:04:35.06 So how is it that I can think about designing a library of variants, in a calculation,
00:04:44.03 if I can't actually compute the function of interest:
00:04:46.06 catalysis or fluorescence of whatever it may be.
00:04:49.29 Well, the idea is that we want to employ this thing we call the Fold before Function hypothesis.
00:04:56.20 The idea is that in order to have a functional molecule, the protein has to be folded.
00:05:01.14 For most of the functions that we're thinking of, we require a stably folded protein.
00:05:07.28 And so our design calculation, in terms of designing a library,
00:05:11.25 can focus not on the actual chemical functionality,
00:05:16.00 it can focus on things we know we can do well,
00:05:19.20 which are more along the lines of fold stability.
00:05:23.04 So we run the calculation for fold stability,
00:05:27.13 generate variants that are predicted to be stably folded,
00:05:32.00 and then screen that library for functional properties.
00:05:35.17 So this if Fold before Function: we design for stability,
00:05:39.02 and then screen experimentally for the target function.
00:05:44.29 Okay, so let me briefly describe the algorithm for actually making the library.
00:05:49.18 So earlier in the talk I indicated that we have to start with a structure
00:05:54.23 and we use a pairwise-decomposable scoring function.
00:05:57.25 What that really means, as I indicated previously,
00:06:00.21 is that we have two types of energies: a one body energy and a two body energy.
00:06:04.19 A pairwise-decomposable scoring function just means that we can't have energies
00:06:09.15 that require more than the knowledge of two positions simultaneously.
00:06:14.00 So from that we get energies, these one and two body energies.
00:06:17.07 We then employ an optimization protocol, like Monte Carlo,
00:06:20.19 and from that we get then the sequence of amino acids that we want, the composition.
00:06:26.03 That's the general scheme for our normal calculation that I described earlier:
00:06:33.17 we start with the structure, we compute one and two body energies based on the rotamers
00:06:40.08 (these are also referred to as singles energies and pair energies) and then we do the optimization.
00:06:45.17 We end up with a single sequence shown on the far right.
00:06:49.27 To design a library we actually introduce the stuff shown in red.
00:06:54.13 So the first things we have to do is convert our one and two body rotamer-based energies
00:07:01.10 into one and two body-based amino acid energies.
00:07:05.29 And we do that in a way that allows us to control the composition of the amino acids.
00:07:18.25 That is, it allows us to do things that are interesting.
00:07:22.12 So, for example, if we want to ensure that we always
00:07:25.05 maintain the wild type amino acid at a design position,
00:07:28.17 we can add a composition benefit and score it such that in all the calculations
00:07:34.18 the wild type amino acid always shows up on the list.
00:07:39.09 And we can also add a diversity benefit that is a tunable parameter
00:07:43.25 that will effectively allow us to control the size of the library
00:07:51.17 that we want to evaluate experimentally.
00:07:54.22 So once we have these amino acid singles and doubles energies,
00:07:59.12 these one and two body amino acid energies,
00:08:01.25 we then have to convert them into what we call set energies.
00:08:06.21 And this is where we build in the concept of what a library is.
00:08:09.29 So, for example, what are the sets that would be of interest.
00:08:15.07 Well, the logical sets, an obvious way to group amino acids,
00:08:19.00 is to group them by their codons, their nucleic acid codons, that code for the amino acid and,
00:08:27.20 in terms of sets, we can use different sets of degenerate codons
00:08:31.11 to code for different sets of amino acids.
00:08:37.22 So the sets that we use for what I'll show going forward
00:08:40.23 are the different degenerate amino acid codons
00:08:43.29 and the implied list of amino acids that goes with those.
00:08:48.18 Again, we need to have one body version of that energy and the two body version of that energy.
00:08:54.27 Now, when we run our optimization -- it's the same optimization calculation,
00:09:00.04 we still have a matrix of energies and we still have this long vector of energies,
00:09:02.14 so you can run the same optimization protocol --
00:09:05.13 but now the optimization is actually optimizing across degenerate codons.
00:09:11.10 So the sequence that comes out is a sequence of degenerate codons
00:09:12.22 which then actually is the library.
00:09:17.28 If we synthesize the nucleic acid
00:09:18.04 and dope in the appropriate nucleotides at the targeted positions,
00:09:23.10 we'll end up then with a library of proteins that reflective of our calculation.
00:09:31.04 So rather than ending up with a single sequence, we end up with a library.
00:09:35.05 Okay, so we wanted to test this idea on a model system
00:09:39.04 to see whether or not it was a viable concept.
00:09:42.26 And the model system we selected, for convenience, was Green Fluorescent Protein (GFP).
00:09:48.21 Here's an image of Green Fluorescent Protein.
00:09:52.12 This is a naturally occurring protein that has
00:09:54.25 an interesting chromophore which you can see in the middle of the molecule.
00:09:59.08 And that chromophore is contained on this central alpha helix.
00:10:04.28 What we wanted to do in this design calculation is to design libraries of GFP
00:10:11.05 based on a variety of different computational protocols.
00:10:17.00 So we're actually using a variant of GFP that has a point mutation relative to the wild type
00:10:22.22 that's not important for this discussion,
00:10:26.01 and we're targeting 15 positions on this protein.
00:10:28.13 There are 15 positions that flank the actual chromophore.
00:10:35.14 Positions 57-72, but excluding position 65,
00:10:41.24 which is actually the position where the chromophore is.
00:10:44.05 Now, this region of the protein is actually completely buried and normal library construction
00:10:49.03 on completely buried segments of proteins would be very difficult because of the
00:10:53.13 detailed packing and hydrophobic nature of the core it's hard to make random libraries
00:11:01.13 that are going to be viable when you're looking at completely buried positions.
00:11:06.15 Okay, so the property of interest here is fluorescence
00:11:09.09 and so when we think about preservation of function,
00:11:14.01 we want to make sure that the members of our library all fluoresce.
00:11:19.00 In terms of diversity of function, what we're interested in looking at is
00:11:23.08 the change in emissions maximum peak position.
00:11:29.24 So GFP has an emissions maximum at some wavelength and we'd like to think about
00:11:34.23 shifting that peak position around.
00:11:37.23 Perhaps we want something that's slightly redder than GFP
00:11:39.11 or maybe something that's slightly bluer than GFP.
00:11:43.10 So the preservation of function is scored by integrated fluorescence intensity
00:11:48.03 and the diversity of function is scored by peak position of the emissions maximum.
00:11:56.05 And then there are going to be a number of libraries that we're going to evaluate,
00:11:59.07 most of which are design libraries.
00:12:03.11 So on the top is this DBIS method which I just described
00:12:06.22 using the ORBIT scoring function and
00:12:09.12 the calculation is run in a way such that we're going to force the calculation to use
00:12:16.27 2-fold degenerate codons at 9 of the 15 positions.
00:12:24.12 So the calculation is going to decide,
00:12:28.02 of the 15, which 9 positions are going to get mutated and,
00:12:31.14 for each of the 9 positions that will get mutated,
00:12:34.06 it's going to decide which 2-fold degenerate codon is going to be used.
00:12:36.15 An additional restraint on the calculation is that
00:12:39.09 we always want to preserve the wild type amino acid.
00:12:43.04 So the calculation then will be restricted to the wild type amino acid
00:12:48.13 and an additional amino acid that can be coded for by a 2-fold degenerate codon.
00:12:58.10 So this will give us a very high mutation rate: 9 out of 15 positions will be mutated/targeted.
00:13:06.19 We also want to run a DBIS calculation which is smaller in terms of mutation rate.
00:13:13.05 So for the DBIS 4^4 calculation, we'll let the program decide
00:13:19.18 which 4 of the 15 positions will get mutated
00:13:24.02 and in this case, the program will select 4-fold degenerate codons,
00:13:25.11 again with the constraint that the wild type amino acid has to be included.
00:13:30.29 The third method, SCMF (Self-Consistent Mean-Field) is based on a protocol that was developed
00:13:38.24 a number of years ago in the lab.
00:13:43.00 Here, what we do is we run a mean-field calculation in order to identify
00:13:48.19 the 2 positions out of the 15
00:13:52.10 that are predicted by the calculation to have the highest tolerance to mutation
00:13:57.03 and then at those positions we do saturation mutagenesis.
00:14:02.07 In the actual experiment, we use 32-fold degenerate codons to do that.
00:14:07.13 So the first three are computationally directed libraries:
00:14:11.00 two with the new method I just described and
00:14:13.25 one with a previously published method that's
00:14:17.04 focused on trying to find positions tolerant to mutation.
00:14:22.09 The fourth library is a pseudo-random library as a control.
00:14:29.13 So in our pseudo-random library,
00:14:30.28 we're going to take the 9 positions that were determined by the first calculation,
00:14:35.02 and pick a random 2-fold degenerate codon for that position.
00:14:42.16 And this is pseudo-random because we're using the calculation to target which 9 positions.
00:14:48.14 A fully random library would probably not give any useful results
00:14:53.03 for the reasons I stated earlier.
00:14:54.15 This is a very core position of the protein.
00:14:57.21 We also have another control, an error prone PCR control,
00:15:02.03 targeted at the entire gene. You'll see why we did that momentarily.
00:15:08.15 Okay, so here are the libraries:
00:15:13.10 the three computational libraries and then the random control.
00:15:18.00 And you can see, as I described, that the DBIS library,
00:15:26.11 in what looks like blue/dark color is the wild type amino acid,
00:15:30.00 and in red are the variants that were selected by the calculation.
00:15:35.14 And you'll see that 9 of the 15 positions were selected.
00:15:39.10 The DBIS 4^4 calculation targeted 4 of the 15 positions and then used 4-fold degenerate codons.
00:15:46.08 So you can see the set of amino acids that are being selected at those positions.
00:15:51.03 The mean-field (SCMF) calculation, as I said earlier,
00:15:53.21 is going to find the 2 positions that it thinks
00:15:55.04 are the most tolerant to mutation, and do saturation mutagenesis at those 2 positions.
00:16:00.11 And then the random control is the same 9 positions as the first DBIS calculation,
00:16:06.19 but now with a randomly selected 2-fold degenerate codon
00:16:10.27 but also with the constraint that we preserve the wild type amino acid.
00:16:14.17 Okay, so the first property of interest --
00:16:16.24 so we can make those libraries in the laboratory and then
00:16:22.06 pick colonies, make protein, and measure optical properties --
00:16:25.23 so the first property of interest is the integrated fluorescence intensity,
00:16:31.27 which we're using as a measure of the preservation of function.
00:16:36.21 So the libraries are ordered from top to bottom by their mutation rate.
00:16:42.29 The most heavily mutagenized library has an average mutation rate of 4.5.
00:16:48.23 This is actually a phenomenally high mutation rate because you have to remember that
00:16:53.26 this is 4.5 mutations over a stretch of 15 amino acids, so very heavily mutagenized.
00:16:59.00 This DBIS 4^4 library has a slightly lower mutation rate of 3.0.
00:17:03.28 The mean-field calculation library has a mutation rate of roughly 2,
00:17:08.07 and of course the pseudo-random control has the same mutation rate as the top library
00:17:13.22 because that's how we set it up.
00:17:17.03 The error prone PCR control has a very low mutation rate:
00:17:21.17 it's 2.5 mutations, but directed across the entire gene.
00:17:27.26 So a relatively low mutation rate; the entire gene is on the order of 300 amino acids.
00:17:31.14 So we go from very high mutation rate for the design libraries to slightly lower mutation rates,
00:17:35.07 and then a very low mutation rate for this error prone PCR control.
00:17:41.06 Shown in this column is the number sampled.
00:17:46.00 So we expressed the library,
00:17:47.18 obviously we don't know what the proteins are because they're randomly distributed,
00:17:49.20 picked the colonies, isolate protein,
00:17:52.29 and measure the fluorescent properties.
00:17:55.11 So that column just shows how many different samples we evaluated
00:17:59.14 for each of the different libraries -
00:18:01.02 on the order of 1500 samples, which is roughly 3-fold oversampling of the theoretical library size.
00:18:12.01 We're going to score for preservation of function,
00:18:14.27 as measured by integrated fluorescence intensity.
00:18:18.00 For the library with the highest mutation rate, what we can see is that
00:18:22.04 roughly 50% of the library maintains
00:18:27.26 an integrated fluorescence intensity that's at least 1/50 that of the wild type protein.
00:18:32.29 And as much as 12% of the library maintains an integrated fluorescence intensity
00:18:38.12 that's at least half of the wild type protein.
00:18:42.15 So this is actually pretty remarkable: we have an enormous mutation rate
00:18:48.24 and yet a significant fraction of the library still shows fluorescent properties,
00:18:53.04 which is good, which is what we want.
00:18:56.07 As we now look at libraries with lower mutation rates,
00:19:00.09 we see an interesting trend.
00:19:04.20 In fact a very counterintuitive trend.
00:19:08.27 At the average mutation rate of 3.0, across this stretch of 15,
00:19:13.03 the preservation of function actually goes down:
00:19:16.03 with the lower mutation rate we actually lose preservation of function.
00:19:20.13 And the mean-field library, which is now saturation mutagenesis at the 2 positions
00:19:23.27 predicted to be most tolerant to mutation,
00:19:27.25 has the lowest preservation of function.
00:19:32.08 Now this is actually quite counterintuitive.
00:19:33.25 I told you earlier in the talk that typically as we lower the mutation rate
00:19:38.02 we would expect to have higher preservation of function.
00:19:41.07 That is, if we're beating on the protein less,
00:19:45.10 we expect more of the library to actually survive the mutagenesis.
00:19:49.02 But here what we see is that high mutation rate
00:19:52.05 gives us the greatest preservation of function.
00:19:54.24 This is exactly what we wanted to achieve in the design calculation.
00:19:58.05 We want the design calculation to actually show us
00:20:01.08 where to put the mutations,
00:20:05.00 constrained for a given library size,
00:20:08.10 and to select the amino acids that will go there
00:20:11.26 with the highest probability of stabilizing the fold.
00:20:14.07 So that's a great result.
00:20:16.27 The random control gives the expected result: very low preservation of function.
00:20:20.24 If we don't put any chemical intuition --
00:20:25.19 a calculation that scores for chemical functionality or chemical stability --
00:20:30.22 of course we're going to get libraries that aren't very good.
00:20:36.05 Now the interesting result is the error prone PCR control, which has a very low mutation rate,
00:20:40.13 remember 2.5 across the entire gene.
00:20:44.10 It has a high preservation of function and this is exactly what we would expect:
00:20:49.02 if we mutagenize at a low level, on a random basis,
00:20:53.01 we would expect most of the function of the library to be preserved,
00:21:00.28 to be very similar to wild type.
00:21:03.13 That's exactly what we got with the error prone PCR control.
00:21:04.16 Now we'll see on the next slide, when we think about diversity of function,
00:21:08.18 that the PCR result will also be consistent with what our expectations would have been.
00:21:15.10 Okay, so let's think about diversity of function.
00:21:18.09 We're going to score, for this chart, diversity of function
00:21:24.05 of those clones that have at least 50% of the integrated fluorescence of wild type.
00:21:32.01 And so these are the numbers of samples that are like that from each of the libraries,
00:21:35.27 and again the libraries are ordered from high mutation rate to low.
00:21:40.27 And we're looking at diversity of function as the
00:21:43.28 difference in peak position between wild type and the variants.
00:21:47.14 Okay, so you'll see at 0 would be wild type, no change in peak position.
00:21:54.10 This is change in peak position in nanometers (nm).
00:21:57.20 The white bar, in the middle of the red bar, is the median peak position of the samples.
00:22:03.21 And you can see for the top design library that it's roughly 2-3 nm shifted.
00:22:15.07 The red bar is +/- 50% of the samples --
00:22:19.15 sorry +/- 25% of the samples --
00:22:23.08 so the red bar encompasses half of the data.
00:22:27.08 And the black tick marks are the actual data.
00:22:30.25 So you can see, for the library with the highest mutation rate, we get what we would expect:
00:22:35.18 high mutation rate gives us fairly good diversity of function,
00:22:40.06 in this case represented by changes in peak position.
00:22:44.19 And the peak positions are almost uniformly distributed over the range,
00:22:47.13 from the high end to the low end.
00:22:50.27 If we go into the libraries that have lower mutation rates,
00:22:55.20 we get the expected result, the intuitively obvious result:
00:23:01.13 lower mutation rate means that we're going to have less diversity of function
00:23:05.12 because most of the molecules are going to look like wild type.
00:23:09.08 And so you can see that by the time we're down to the mean-field library,
00:23:13.21 all the data are sitting right at 0.
00:23:18.08 So, we're basically having no effect on the diversity of function in the protein:
00:23:25.28 they're all wild type-like.
00:23:28.27 The random control is also similarly very constricted and shows the expected result:
00:23:35.07 lower diversity of function.
00:23:37.01 And the error prone PCR control is also consistent with what we would have expected.
00:23:40.16 So recall that the white bar is the median.
00:23:45.00 The median peak position for the PCR library is right at 0,
00:23:50.05 that is we basically haven't changed the functional characteristics of the library at all.
00:23:54.23 And the red bar, which you may not be able to see because it's right around 0,
00:24:01.20 most of the data for that library are right at 0, that is no shift in peak position at all.
00:24:08.19 And so we have a very sharply peaked gaussian or gaussian-like distribution at 0,
00:24:12.28 consistent with what we would expect:
00:24:15.05 low mutation rate, distributed randomly over the entire gene,
00:24:18.06 leads to little change in the functional characteristics of the protein.
00:24:23.07 Okay, so that's great, we have a library design method that shows high preservation of function,
00:24:34.23 which means we can go to high mutation rate, which then leads to high diversity of function.
00:24:41.08 Now, this combination of high preservation of function and high diversity of function
00:24:46.02 in many respects corroborates our idea, our hypothesis,
00:24:50.08 that libraries can be designed based on fold stability
00:24:54.16 and ignoring the direct calculation of chemical function.
00:24:59.09 So the Fold before Function hypothesis is supported by the data that I showed you.
00:25:05.15 And we can conclude that stability is a good surrogate for function in protein library design.
00:25:11.19 Okay, so now let's move on to a an example of applying this to a problem
00:25:16.13 that might actually be interesting to some of you.
00:25:22.11 So, we want to now see if we can apply this methodology to generate
00:25:28.13 a variant of an existing Blue Fluorescent Protein that has enhanced optical properties.
00:25:34.00 So, Blue Fluorescent Protein, or BFP.
00:25:36.23 And the idea here is that in many experiments,
00:25:40.25 BFPs are used because they're not very stable, that is they're not photostable --
00:25:47.01 they photobleach very readily and they have short half-lives --
00:25:51.00 and they have very weak fluorescence intensities.
00:25:56.05 So they're very difficult to use in a cell biology experiment, for example.
00:25:59.28 What we'd like to do is take BFP,
00:26:02.12 design a library using the methods I just described,
00:26:05.14 design a library targeting residue positions around the chromophore.
00:26:10.23 The chromophore is shown in blue in this structure
00:26:14.26 and the positions that we've selected are a set of positions shown in red
00:26:19.07 that are geometrically close to the chromophore.
00:26:24.28 And the idea then is to try to find variants
00:26:27.04 that actually have improved fluorescent properties and in particular,
00:26:30.07 we're interesting in improving the quantum yield to get better brightness
00:26:35.14 and improving the photobleaching properties of the protein
00:26:38.04 so that the fluorescence lasts for longer periods of time.
00:26:43.21 So we selected 12 design positions for this study and set the calculation up to generate a library
00:26:51.10 with a diversity of 10^5, roughly.
00:26:55.21 That's a huge reduction in the saturation mutagenesis diversity,
00:27:00.08 which would be on the order of 10^15 I believe.
00:27:05.04 And so, we've already achieve enormous reduction in combinatorial complexity
00:27:09.27 versus the saturation mutagenesis experiment which obviously you could never do
00:27:14.28 because you could never screen a library of that size.
00:27:21.06 Alright, experimentally, again, the library is going to be encoded using degenerate codons
00:27:25.14 and experimentally we're going to screen the library
00:27:29.00 using Fluorescence Activated Cell Sorting (FACS).
00:27:35.12 The chart on the upper right shows the design positions,
00:27:42.24 the wild type amino acid for each of the design positions,
00:27:47.22 and the codon that is seen in the wild type gene.
00:27:53.26 As I mentioned earlier, the output of the design calculation, in a case like this,
00:27:56.28 is in fact a list of degenerate codons.
00:28:00.04 And that list is shown here in the row which is Design.
00:28:09.07 You should be able to see the footnote on the table which shows how those degenerate codons
00:28:13.23 are actually translated into normal, everyday familiar, nucleotides.
00:28:23.03 And so the list of amino acids that is pulled out of the calculation is shown,
00:28:28.03 and you can see that the calculation is
00:28:32.26 putting diversity at different positions in the set of the design positions,
00:28:37.13 and it's distributing diversity in a non-uniform way.
00:28:42.06 So in this case, we didn't actually restrict the program to use uniform diversity,
00:28:46.27 we let it choose what it wanted.
00:28:48.22 You can see in some positions it's using a large number of amino acids
00:28:52.07 and other positions it's choosing wild type amino acid
00:28:54.19 and other positions still it may be only using two amino acids.
00:29:00.10 This would be hard to imagine coming up with on your own by looking at the structure.
00:29:03.16 So, there's the library. We then screen the library,
00:29:07.21 again using FACS analysis,
00:29:11.20 and selected variants that had interesting properties.
00:29:17.13 One variant that we pulled out, a variant that we canned Azurite,
00:29:20.19 is a triple mutant of wild type.
00:29:23.27 It's got a phenylalanine to leucine mutation,
00:29:26.14 a valine to isoleucine mutation,
00:29:28.29 and a valine to arginine mutation.
00:29:31.26 And interestingly, as desired, it has a much improved photobleaching half-life.
00:29:40.04 So you have see in this graph, wild type BFP and a previously known variant,
00:29:45.07 show a rapid photobleaching.
00:29:50.11 The Azurite variant on the other hand, shows photobleaching that is much improved,
00:29:54.25 quantitatively by about 40-50 fold.
00:30:01.18 It's also much brighter than wild type BFP.
00:30:06.04 So, shown on the right are panels that show fluorescence in 293T cells
00:30:13.07 and you can see in the bottom of that panel that
00:30:18.15 BFP at 0s is barely visible and of course if photobleaches rapidly
00:30:24.07 so as you go out in time it becomes even less visible.
00:30:28.20 Azurite, on the other hand, shown on top, has markedly improved fluorescence intensity and,
00:30:35.29 as the previous graph indicates, and as you can see over time,
00:30:42.16 retains that fluorescence intensity as the experiments are carried out.
00:30:47.00 So that's actually very nice.
00:30:50.17 So in terms of quantum yield, the quantum yield goes from 0.34 to 0.55.
00:30:55.22 So we get an improvement in quantum yield which translates into improved brightness
00:30:59.09 and an improvement in photobleaching half-life which means that this variant
00:31:03.04 can actually then be used in real experiments.
00:31:05.14 So we take this as an interesting success.
00:31:09.03 Now I should point out, however, that
00:31:11.23 one of the mutations of the three turns out to have arrived serendipitously.
00:31:16.08 So it's amazing that, as in many library experiments,
00:31:20.09 you know, you rely on luck at some level, this is computationally-directed luck,
00:31:24.26 but even in a computationally-designed library, we still have some element
00:31:30.09 of just blind luck.
00:31:32.19 And so, in the design calculation at position 224,
00:31:36.12 you can see in the table above,
00:31:40.10 the program actually wanted a 2-fold degenerate codon
00:31:44.29 that was just methionine and valine.
00:31:52.05 But when the oligos were ordered --
00:31:54.07 they were either ordered incorrectly or synthesized incorrectly --
00:31:57.08 we ended up with a degenerate codon that was actually much more inclusive
00:32:01.27 in terms of the number of amino acids that were allowed.
00:32:05.12 And as fate would have it, one of those amino acids ends up in the final solution, an arginine.
00:32:11.24 So on the one hand, the design calculation recognizes that it wants to put something big there,
00:32:16.21 methionine was one of the large amino acids that was selected by the design calculation,
00:32:20.23 experimentally it chooses a large amino acid (arginine),
00:32:25.13 but what we missed in the design calculation is that the real experiment, the protein itself,
00:32:30.27 really wanted to have a large polar amino acid.
00:32:36.22 And our force field methodology, for this particular calculation,
00:32:39.19 does not accommodate the burial of a large polar amino acid at that position.
00:32:45.11 So even though this is a very positive result, it also points out the need to continue
00:32:49.20 to think about how to modify these methods,
00:32:54.02 in particular with respect to developing better force fields that would accommodate, for example,
00:33:00.24 the burial of a large polar amino acid in the core of this particular protein.
00:33:08.04 Okay, so let me acknowledge the people who did the work:
00:33:11.23 the design method, development, and experimental validation
00:33:15.12 was done by Thomas Treynor, who was a postdoc in the lab at the time,
00:33:19.25 with assistance from Christina Vizcarra, who is a graduate student in the lab,
00:33:24.18 and by Daniel Nedelcu, who was an undergraduate student in the lab.
00:33:30.13 The enhanced BFP screen was done
00:33:33.06 in collaboration with Patrick Daugherty's lab at UC Santa Barbara
00:33:35.28 and Marco Mena, a grad student in his lab, actually carried out the work.
00:33:41.05 Thank you - I hope you enjoyed it.

Videos in this Talk

Part 1: Protein Design by Computation
Audience:
- Researcher
- Educators of Adv. Undergrad / Grad
Part 2: Designing Protein Libraries and an Example of Making a Better Fluorescent Protein
Audience:
- Researcher
- Educators of Adv. Undergrad / Grad

Speaker: Stephen Mayo

Total Duration: 1:05:35

Recorded: January 2008

All Talks in Biophysics

Talk Overview

In this lecture, I discuss the challenges of designing new proteins that fold into a particular structure or perform a particular function. Using computational protein design, one can design a protein based solely upon our knowledge of amino acids and protein folding, a hard task but one which has had recent successes.

Another approach is to screen combinatorial libraries of proteins for a desired function. In particular, I will discuss how structure-based computational methods can allow for high mutation rate (to explore a diversity of function) while maintaining stably folded structures (a necessity to preserve stability and function). In the last part of my talk, I discuss how we have used protein libraries to design a new blue fluorescent protein.

Speaker Bio

Stephen Mayo

Stephen L. Mayo is currently Vice Provost (research) and Bren Professor of Biology and Chemistry at Caltech, where he has been a faculty since 1992. Dr. Mayor earned his B.S. in Chemistry at the Pennsylvania State University, and received a Ph.D. in Chemistry from Caltech in 1987. He did postdoctoral work as a Miller Fellow… Continue Reading