Luyện nói tiếng Anh bằng Shadowing qua video: [TALK 13] Structure Prediction and Design using AlphaFold – Sami Chaaban

Chọn chế độ luyện tập

Mỗi chế độ mở một phòng luyện tập toàn màn hình.

606 câu

Nếu các câu quá ngắn hoặc quá dài, hãy bấm Edit để chỉnh sửa.

Alright, yeah, so let's talk about AlphaFold.

I'll be talking about structured prediction using AlphaFold, and then at the end, I'm going to talk about design.

I'm going to give more of a higher level kind of view of AlphaFold

and give more of like a user -centric focus and kind of as a user and at the LMB,

you know, how to use it and what you need to know in order to properly interpret the results.

So, yeah, this is the breakdown.

We'll start with the overview, talk about how to write predictions, especially at the LMB.

I'll talk about some tips and then some things that you can use alphabetical prediction for.

And at the end, we'll talk about designing your own protein binders.

All right.

Well, first of all, why would you need structural prediction?

Well, you might just want to know what your protein looks like, right?

I mean, it's very important to give this structural context to your biological questions, right?

But really, I mean, a lot of the applications have been to generate models that help interpret structural data like cryo -EM, crystallography,

and other structural data that either you want to use as a starting point

or you want to help interpret lower resolution features.

You might want to use it for protein engineering for designing constructs.

How do I cut my proteins so that I can purify better?

And we'll talk about that later.

But really what you want to do is quickly generate these hypotheses

that can be tested by these other techniques that we've been talking about.

And this progress towards predicting a structure purely from its sequence has been monitored by this competition called CASP.

And, you know, over the years, the developments in neural networks and how, you know, proteins can be predicted.

And, you know, around 2018 is when AlphaFold came out, which did really well.

And then 2020, you know, AlphaFold2 came out and it did extremely well.

And, you know, anecdotally, 90 % is where you would consider comparable to experiment different.

In any case, these are the results that they were showing.

You see these targets.

So these are proteins that had not been seen before by these algorithms, but then were predicted with quite good confidence.

So you're seeing the global distance test here, just an overall measure of how close you are to the actual structure.

And so then over the years, we've now come out with alpha -fold multimer and alpha -fold 3.

And kind of alpha -fold is an aggregate term for these things, right?

So we have the ability to predict multimeric proteins or now with other molecules like DNA and RNA.

Okay, so how does it work?

The insight comes from available structures on the PDB.

So this is a database online where experimentalists will submit their structures.

And now we have a sequence to structure relationship.

And there's many on the PDB.

And this is what can be used by these neural networks to train a model

that will be able to predict the structure of a new sequence it hasn't seen before.

But one of the insights that was really important, which had been also already in development over the years, is insight from a multiple sequence alignment.

So you look at your sequence of interest, but you also look at the related sequences.

You can see there are residues that are evolving in complementary ways.

And so that's a hint that these might be close to each other in space.

And so this co -evolution is a very important signal in these structure predictors.

And in a way, you can think about the multiple sequence alignment as actually being the input rather than just the sequence.

So that's what we're what I'm calling the multiple sequence alignment structure predictors.

So we have alpha fold, which I'll be mainly talking about.

There's also Rosetta fold from the Institute of Protein Design from David Baker, which is also very accurate.

And again, these you're generating a sequence alignment and you're getting out structure.

You should be aware that there's also these protein language model structure predictors.

This is like ESM fold or mega fold where it doesn't actually generate multiple sequence alignment, but rather directly generates a structure.

And you can think about it as the evolutionary information has essentially been embedded in the model.

So it may suffer from the same things, but it runs a lot faster because you don't have to generate sequence alignment.

All right.

So how does that fold work?

Just a high level view.

I'm not going to get too deep into it.

But like I said, we have to generate a multiple sequence alignment.

And the way to do that is to search, you know, database.

So we have things like your F90, Magnify, or BFD.

These are either metagenomic databases or some clustered uniprot databases

and from those we can extract related sequences to our sequence of interest, generate multiple sequence alignment that then goes into this first model,

the first part of the algorithm, which is the MSA module.

And by extracting those relationships, we can now just generate an initial estimate of where these XY coordinates of the different residues should be.

Another thing that can be input into this module is, you know, examples of structures that are similar to your structure right

so these are templates it turns out that these are not as important

but you should know that they're also something that helpful can fetch from the pdb let's say and insert into

this MSA module.

Now, after we get our initial X, Y, Z coordinates for residues, then it goes through some other modules.

So this is a structural module that then gives you something that looks like a protein.

And for AlphaFold2, this was a structural module, but then for AlphaFold3, this is now the diffusion module.

So instead of treating the residues as proper geometries, it's basically a denoising step that's turning it into something that looks like a protein.

And because of this difference, actually, you end up with sometimes quite different results between AlphaFold2 and AlphaFold3.

And some of the things, you know, the caveats of FF3 might be coming from this difference as well, which we'll talk about later.

But essentially, once you get that, and it's not accurate at first, so it goes through this recycling step,

so it gets put back in for a few recycles until it basically converges on something

that is as confident as it can get.

All right.

And you can slap me at any point, by the way, if you have any questions.

really all you have to worry about.

Oh, I should talk about ColabFold.

Yes, so this is the AlphaFold framework in general.

Now, you probably have heard of ColabFold.

And essentially, ColabFold is an implementation of AlphaFold that is on the Google Colab notebook, which is why it's called ColabFold.

But also, it has a multiple sequence alignment search that uses MMCs,

which performs a clustering step and also a pre -filtering step.

In any case, it's a faster version of obtaining

that multiple sequence alignment alignment and that's what we have implemented at the LMB locally.

So when you hear call AlphaFold, it's essentially AlphaFold with the MMSeq database search.

But really all you need to worry about as a user is, you know, your sequence.

That's what's nice about these things.

You give it your sequence and you get your structure.

Now with AlphaFold you'll get five structures because they train five different models in slightly different ways.

And so these are five different attempts at getting the right answer

and then it will rank them and so hopefully one of them will be correct.

You'll get your multiple sequence alignment.

So this is what was generated in the pipeline.

Here's an example from this run.

100

I'll talk about it in a second.

101

And then really importantly, you'll get these confidence metrics.

102

For example, we get global metrics like the predicted template modeling score or the interface predicted template modeling score.

103

This is basically one number to say how accurate we feel this structure is and where higher is better.

104

But then you also have local confidence metrics like PLDDT or PAE, which we'll get into.

105

And these are, you know, what is my confidence about the secondary structure or about two residues interacting?

106

So I think these are really important.

107

So we'll talk about them in detail.

108

And something that I just want to highlight here is that, you know, for things like PTM and IPLDT,

109

these are numbers that are better when they're higher.

110

The PAE is an error.

111

So a lower number is better, right?

112

Lower error is better.

113

I see people make this mistake quite a bit.

114

But yeah, we'll get into those.

115

So here is an example of a protein that I work on.

116

So I work on a motor protein, dynein, and this is a dynein -dynactin complex.

117

And just part of dynein is this, you know, subunit called the light intermediate chain.

118

It's a good example because it has this globular domain.

119

It has some alpha helices on either end.

120

And from which you predict just this protein on its own, you see you get five structures from alpha fold.

121

You can see these alpha helices are kind of floating in space around the globular domain.

122

You know, one shouldn't necessarily take those positions of those alpha helices seriously.

123

If they're, you know, stuck on a string, it just kind of plays randomly.

124

But these five structures are ranked by the algorithm.

125

And over the years, different ranking metrics have been used.

126

For example, average PLDDT or just that PTM number or the IPTM number or some combination of these things,

127

80 % of this, 20 % of that.

128

so in general what i recommend is that you actually look at all the five of those

129

predictions and i'll talk about some special cases where that's important

130

but yeah i mean typically users will just look at the top ranked one

131

and it might not be able to give them the answer

132

that they want the other thing that's generated like i said is the multiple sequence alignment

133

so here you're seeing all of the sequence

134

that were pulled out of those databases across the position of the protein and here's the concept of msa depth right

135

So the more sequences in the multiple sequence alignment, the more the depth, so the more robust the prediction is going to be in that region.

136

You can see those regions in the NNC termini that have less depth, so they may not be predicted as robustly.

137

So that's good to look at.

138

I mean, typically you would want more than 30 sequences, although you can see here, you'll get thousands in many cases.

139

Now, the first important metric is this PLDDT.

140

So this is a local confidence metric

141

so it's you know c alpha distances within 15 angstroms of the residue

142

and so this is a really good way of figuring out

143

if you have confident secondary structure so one rule of thumb is below 50

144

and it's probably disordered so there's no secondary structure

145

if it's above 50 then probably the secondary structure is created with confidence so in this example.

146

You have this region of the protein that is disordered.

147

You can see the spaghetti is colored in red, and those are the low PLDDT regions, whereas these alpha helices are high PLDDT.

148

So if you're asking questions about secondary structure, it's better to look at the PLDDT.

149

And I'm just showing what AlphaFo's recommendations are for what these numbers mean above.

150

So yeah, 0 to 50 is disordered for bad prediction.

151

Now, if you're interested in how residues are interacting across a protein, then you want to look at the predicted aligned error.

152

So here's each one of the five ranked predictions with their PAE plus below.

153

And so let's just focus on the top ranked one here.

154

So what you're asking with the PAE is what's the expected error of residue one

155

if you align the predicted and true structure at residue two,

156

which is a bit confusing because we don't actually have the true structure in this case.

157

So that was the power of Alpha Fold is

158

that they were able to train this metric based on when they did have the true structure.

159

And basically, how does that confidence turn into an error?

160

And then can we apply this to things that we haven't seen?

161

It turns out to do quite well.

162

And let's talk about this plot a bit more specifics.

163

So if you look at this part of the plot, we have a question of what is the confidence of residue 198 versus itself?

164

Of course, it's very high, which is why you end up with this diagonal across the PAE.

165

A more interesting question is, for example, what's the confidence of 198 versus 113?

166

It's actually not that interesting, but it's a good example of where you have very high error.

167

So the confidence of the position of these two is actually not very good, which makes sense.

168

I mean, they're very far apart.

169

Now, it gets more interesting when you're asking things within a domain.

170

For example, here I have two residues

171

that are close to each other whether i'm asking 324 versus 256

172

or vice versa you can see that we're getting a very low error

173

so it's very confident in the relative position of these two residues

174

so we can you know make a hypothesis

175

that these two may be in a domain or interacting with each other or something like this

176

and so what's interesting is that

177

when you have these domains then you start to become confident in the position of your neighbor

178

and your neighbor's neighbor and your neighbor's neighbor's neighbor

179

and then you get this extension away from the diagonal of this low error region

180

so that's when you get domains so you get these like square um regions or you know extended regions

181

if we keep going down you see this loop there's not much to say within the pae

182

and we have our next sub domain and

183

when you have two sub domains that are interacting with each other

184

that are far away in sequence space you end up with these blocks

185

that are away from the diagonal right

186

so here we have this region of the protein is interacting with

187

that region of the protein

188

so you get you end up with these regions of low error that are outside of the deck

189

for things like secondary structures it's not really that informative you can see

190

that the extension away from the diagonal is actually not

191

that large because the alpha helices are quite small and again

192

if you're asking questions about secondary structure you'll be wanting to look at the plddt rather than the pae

193

so what about protein protein interactions it's the same logic

194

so here we have an example of this subunit

195

that i showed you before and this other part of the dining tail and you know

196

if you give any implementation of alpha fold more than one

197

chain essentially what's going to happen is it's going to use alpha fold multimer

198

so you usually don't have to worry about

199

which model is going to be used is alpha fold is it multimer typically it's already handled for you.

200

And for this protein -protein interaction question, we have these different PAEs for each.

201

We look at one.

202

It's essentially the same thing, right?

203

So we have protein A with itself.

204

This is identical to what I showed before.

205

It has its diagonal.

206

It has extensions from the diagonal.

207

So these are subdomains interacting with each other.

208

Same thing for B.

209

But what we're really interested in are these quadrants of protein A interacting with B or B interacting with A.

210

so that's what you're getting out here right

211

so we have regions of low error um between the two different proteins

212

which is the same logic now you might already be asking yourself why is it

213

that when we're asking a versus b

214

and b versus a are we getting slightly different results like why isn't this perfectly a perfectly symmetric plot

215

and here's a dramatic example of this where we have this little peptide

216

and this um you know globular domain and you can see you know if we're asking B versus A, we're getting a lot of error,

217

but if we're asking A versus B, we're getting less error.

218

And so what's happening here?

219

That's really not clear what's going on.

220

I mean, these things are generally black boxes.

221

But the way that I've understood this is that

222

if you have one protein that has a lot of variability or low confidence in its position,

223

whereas another one has a lot of confidence in its position,

224

like a bigger domain, then sometimes if you ask a question of one versus the other, it'll be quite high error because it's moving a lot.

225

Whereas Yes.

226

The big one versus the small one, you'll get less error.

227

So what I would do as a user is just treat this as a red flag.

228

Maybe this prediction should be taken with a grain of salt.

229

Right.

230

And of course, all of these things need to be tested by other techniques,

231

but I should I should flag this because I get this question a lot.

232

And so this is I mean, that this is probably the most important part of the whole talks.

233

If there are any questions, I'm happy to take them now on the PAE. Or the.

234

I'm going to do that.

235

Good.

236

Now, sometimes it's hard to kind of read these plots.

237

So I find things like ChimeraX where you can highlight certain regions of PAE really useful

238

and then it'll map where that is on the structure.

239

So this is just implemented in ChimeraX.

240

There's also this PAE viewer web app where you can load your structure

241

and you can also load cross -linking mass spec data onto this, which is really nice.

242

You can see which are satisfied enough.

243

All right, so those are the output.

244

We have a structure is a multiple sequence alignment and our metrics, and depending what kind of claims you're making for your project,

245

you'll want to submit different things for manuscript.

246

If you're making interpreting the interaction between different proteins and then continuing further in your project based on those,

247

then you'll want to present the plot for this prediction.

248

You're making claims about secondary structure want PLDDT we don't usually submit things like you know the multiple sequence alignments

249

or the other

250

ranked predictions but yeah please submit your PAEs

251

if you're talking about protein protein interactions in your manuscripts it's also nice to actually deposit these predictions

252

so you know people can actually go

253

and see what they look like it's not really appropriate to submit them to the PDB

254

because those are experimentally derived structures and so something like mono archive

255

which is a nice place that you can deposit these things

256

so here's just a recent example from kelly's lab where they deposited this nice structure from alpha fold 3 prediction

257

and um you know so that other people can check it out all right

258

so how do you actually run your own predictions actually the simplest way is to just fetch a prediction right

259

so alpha fold has run essentially all human proteins through its algorithms

260

and submitted it into this database or I've released it into this database, but it's also integrated into the Uniprod.

261

So if you just go in the structure section of Uniprod, you can see the alpha fold, what your alpha, the structure is.

262

Now, like for something like, you know, RL3 here, this globular protein, you see this one looks like in Uniprod,

263

you know, it makes sense to just fetch it from there or look at it on the website.

264

But for things that are, you know, dimers or, you know, multi -protein assemblies, you can end up with something like this,

265

where it's just like an extended alpha helix whereas we know that this is a coiled, coiled protein.

266

So if you run, you know, a prediction yourself, you know, you'll get the actual structure of this cold coil.

267

And so yeah, typically you'll want to run your own structures,

268

the prediction sorry, and so a simple way to do that is just to go on the Google Colab notebook.

269

Remember I mentioned Colab fold.

270

This is a Google Colab notebook,

271

so hosted by Google where you can put your sequence and have it run on Google's GPUs and then you know, obtain the prediction.

272

um and you know if you just visit um

273

that website then you'll see all the different options it's not just colapful

274

that they have the problem with this is

275

that you have to kind of stay on the page

276

and the tip of the page might time out and

277

if you know

278

if you want to use their nice gpus

279

you have to pay a little bit of money now what's nice about the lmd is

280

that we have a local installation of colapful and

281

if you're not familiar with scientific computing then um i encourage

282

you to visit i mean you should visit the scientific computing uh website read their guidelines on how to use the cluster,

283

how to submit jobs, and your practices.

284

But essentially, once you have all of that in place, you just need to add these lines to your .cshrc or .vashrc,

285

and then you'll have access to the colapfold command, where you can just provide it a FASTA file with the sequences that you want.

286

Just run colapfold to that FASTA file.

287

You'll submit it to the cluster, do its thing, and then sometime later, you'll get your predictions in that directory.

288

And there's a lot of options which you can check out

289

if you just run Quolap Fold on its own to see what your options are for submission.

290

Some of the things that you can play with, for example, is playing with the number of recycles.

291

I mentioned that recycling is getting more attempts at getting the right structure.

292

This is usually handled internally where it reaches some tolerance or if it gets to 20 recycles, usually you don't mess with that.

293

You can also provide your own template.

294

So maybe you have something, a starting point for AlkaFold and you can provide it as a PDB.

295

This hasn't worked really that well for me, but for some people it has worked.

296

So that's an option that you should know about.

297

I didn't show this, but this is just an example of recycles where you can see that the

298

prediction kind of plateaus at around 10 to 11 recycles.

299

And so in this case, that was 11 was sufficient, but usually you just kind of let it handle it internally.

300

Something interesting you can do is actually sample confirmations or try to sample confirmations.

301

Alpha -folds will sometimes just predict the open state of a protein or the closed state.

302

Sometimes you want to get it out of that and try to predict other confirmations.

303

And something you can do is just by introducing uncertainty into the model.

304

So like I said, the deeper the MSA, then the more robust the prediction.

305

But if you want to actually introduce uncertainty and get it out of some local minimum,

306

then you can just reduce the number of sequences in the multiple sequence alignment.

307

And then hopefully that uncertainty will lead it to one another structure getting that other confirmation.

308

You can also increase the number of seeds, which is basically a random initialization of the model.

309

And if you have different seeds, like two seeds, then you'll get two per model array

310

so you get 10 structures in the end and then you can start sampling a lot of different um

311

starting points or you can just randomly switch off parts of the network apparently

312

that works as well for some people here's an example with many seeds with this dropout option

313

you can see this protein has kind of sampled these different conformations

314

so it does sometime work

315

and you know you just have to try these things right um an important one

316

if you're doing like prior oem model map building models into maps let's say

317

you might want to use amber relaxation

318

which is you know helping the sidechains get into the proper geometry

319

because you know alpha fold might have things that clash aren't quite correct

320

and so amber will help them kind of relax into a place

321

that is a good starting point for for model building all right

322

so that's alpha fold two alpha fold three

323

there's this really nice online server from deep mind where you can just put your sequence in

324

and and let it run on their servers.

325

You just need a Google account to be able to do this.

326

The only caveat here is that you only have 30 jobs that you can do per day.

327

But it's really nice because you can essentially add different types of molecules.

328

You can add DNA, RNA, ligands, ions.

329

Here's, for example, an ADP molecule that I've added to this protein.

330

If you're interested in different types of molecules for structure prediction, you should also check out RositaFoldOatom.

331

This one has like this nice web interface that you can use.

332

And then you get this prediction out.

333

This is a really large protein that AlphaFloat3 web server lets you run.

334

And you can see the PAE plot.

335

You can see what we talked about here, the PLDDT colored, the IPTM and the PTM scores, all the things that we talked about.

336

And you can download this, the five structures that it's predicted.

337

The only thing that you can really play with here on the online thing is the seed, right? of the different random initialization of the model.

338

So you can try different seeds if you wanted to see if you could get different results.

339

So let's talk about the difference between the two.

340

I mean, at a higher level.

341

So what's nice about AlphaFull 3, of course, is that it handles different types of molecules, like I mentioned, whereas AlphaFull 2 doesn't.

342

It's extremely fast.

343

They've sped up the MSA portion, and the diffusion module that I mentioned is also quite fast.

344

and this web server of course is running on google's um server

345

so they're also very fast and

346

so you can run up to 5000 amino acids as well um

347

which is really nice of course it has limits you can only run up to 30 jobs

348

and it can produce a few more hallucinations so especially in regions

349

that are disordered it tends to make them ordered

350

they give you alpha helices where there should be i know a lot of people have noticed this

351

so alpha full 2 is also you know still very nice to run I don't think it's necessarily been replaced,

352

but of course it only handles proteins.

353

But yeah, it can be more accurate for multiverse.

354

Here's just some anecdotes about alpha -3.

355

This person is saying it's the end of the world as we know it.

356

After this transcription factor has been correctly predicted to bind to this DNA.

357

But then again, you have people that say, you know, it's not able to do predictions properly,

358

but rather has predicted a ballet 2 -2 structure for this routine complex.

359

So I think you just have to try different things.

360

Something I want to mention also is another injection of information into AlphaFold, which is in this software called AlphaLink,

361

which is essentially taking cross -linking mass spec data and introducing that into the pipeline.

362

So if you think about cross -linking mass spec data, you're essentially getting information about the relationship between residues right the positional relationship

363

and so if you just put

364

that in as weights to the msa module you're essentially improving what it's already picking up from the msas

365

so here's an example for

366

this protein complex where alpha fold has not predicted these in proximity

367

yet with alpha link due to the cross links that were predict that were experimentally measured

368

we can uh alpha can more correctly predict this structure

369

so we have this implemented here

370

if you just do collab link 2 um your fasta

371

but now you also give it your cross links um you can also just run this and uh

372

if you just run collab link 2 without arguments you can see what the um what the options are for

373

that all right so i got some tips so like i mentioned

374

if you have really big proteins

375

you still have to start thinking about what you're submitting to

376

a full three server like i mentioned has up to 5 000 residues the notebook

377

you can get to one to two thousand

378

um on so the gpus on our cluster you can get you know two

379

and a half three and a half thousand depending on

380

which cluster node you submit to but i mean even for the some things

381

that i studied like the dynein heavy chain is four and a half thousand residues, right?

382

And so you could try getting, you know, very close to the maximum limit, but then something like alpha 3 will,

383

you know, predict this shrimp -like structure for dynein, which we know is not correct.

384

And so typically what you want to do is segment out your big proteins into smaller portions, smaller domains, smaller regions of interest,

385

and then maybe stitch them together if that's your goal.

386

So this is what I've done for this large protein.

387

I mean, sometimes it can even help with accuracy.

388

You can see here this cold protein that has this spaghetti structure between these two cold coils.

389

But if you segment it out and then do predictions of different segments,

390

you can correctly get the intermediate cold coils that we know are there from other techniques.

391

Something to watch out for.

392

I mean, false negatives are always going to be a thing, right?

393

but here's an example of this little peptide binding to this long -closed coil

394

that we have experimental evidence for that is not predicted.

395

I wanted to mention this

396

because in this case we just truncated the protein a little bit more

397

and all of a sudden it's able to predict this interaction.

398

So sometimes you get some weird phenotypes like this that you want to play around with, you know, where you're truncating things.

399

Much more scary are the false positives.

400

Here's an example where I've taken this protein in green and I've just chopped it in half.

401

So basically I'm exposing what would be varying residues.

402

And I have another protein that's basically not related.

403

And you can see that it's basically taking this beta strand

404

and appended it to the core of this protein with a very nice PAE.

405

So we're interpreting here, right, this PAE plot, very nice in these quadrants, whereas this is impossible.

406

So you have to be really careful about these things, you know, where you're chopping things.

407

And of course, all of these things need to be further verified by other methods, like the ones that we've been talking about.

408

I didn't mention that you can get model -to -model variability, right?

409

You get five predictions.

410

So here, four of these predictions are showing that this peptide is interacting with the cold coil here.

411

But then if you look at the fifth one, we're getting a different part of the peptide interacting with a different part of this cold coil.

412

And there's some NMR data suggesting that this peptide in its entirety can interact

413

and so maybe this is something that warrants further investigation right

414

so this is something that came out not just from the first ranked result

415

but also from the other results

416

or you get cases like this where you know one of

417

them will be open close well some of them will be closed closed

418

so just something you got to watch out for when you're

419

when you're running these things all right let's talk about some applications

420

so i think the simplest to understand is when you're doing structural biology

421

and you have lower resolution data you want a complementary method to see

422

if you can interpret the interactions i think alpha fold is great for this here's something

423

that clinton um put together for one of the structures he was working on back in 2015 of dining

424

so this is again part of this complex we have this big dynamic structure it's like a filament -like structure

425

and they were had a hard time fitting this p150 subunit into this density

426

And so what he did is he chopped it into different pieces.

427

He predicted the structure of each of these pieces, and then he could very nicely place them into this lower resolution map.

428

Now he has something he can work with, because now he can make mutations and test if this is actually where it docs using other methods.

429

Of course, it also helps with higher resolution structures.

430

I mean, here is, for example, a part of a high resolution structure that's not well predicted or not well resolved, like this little loop.

431

And, you know, if you just predict a structure with Amber Relax on, maybe it will help you get the starting point for this little loop.

432

And then you can maybe work with that and see if there's something you can build into it.

433

I didn't mention protein engineering early on.

434

Here's that cold coil that I mentioned.

435

And here I was trying to purify this protein and accidentally truncated it at this four helical bundle.

436

And it turned out that based on alpha -folk prediction,

437

I had cut right into this region that would actually be exposing part of the cold coil core.

438

And so then I just changed the truncation to this other region, and all of a sudden it became soluble.

439

So it doesn't mean that's why it worked, but just something it's nice to actually see where you're truncating your proteins when you're making these designs.

440

All right.

441

Something else you can do is screen for interactions if you have a database of,

442

let's say, So you have a cross -linking mass spec data set, or you have interaction screens deposited online from the BioGrid.

443

There's something like AlphaPullDown that lets you then do alphaful predictions of everything

444

and see if you can get new predictions out or something that you can run further experiments on.

445

I've written something called AlphaScreen, which just takes things from the unit prod and it makes little FASTA files for you to submit to our cluster.

446

Admittedly, this is a bad name given the AlphaScreen that was mentioned this morning, which is an actual essay.

447

Anyway, when you're doing these things,

448

then it starts to become more important how you rank these proteins because you might have hundreds or thousands of predictions then.

449

And, you know, there's things like the IPTM.

450

There's a PAE in that quadrant.

451

Is that a good thing to look at?

452

There's this local interaction score, which is some aggregate thing.

453

Then you have these docu.

454

I think the way to think about this is that by ranking, you're essentially just shortlisting these things.

455

And then really you need to use your experience

456

and carefully look at each structure to actually see which ones are worth further investigation.

457

Something else to think about with structure prediction is the ability to look at structural homology.

458

So we have, now if you have a domain

459

and you want to see all of the things that look like this domain but don't necessarily share the same sequence,

460

they're not homologous, you can use something like progress, which comes from the LMB, from Joe Greener, and Bold Mason and the Valley,

461

for example, where you can run these things through an alpha -fold database,

462

and maybe there's something that hasn't been solved that you can then further investigate.

463

Here's an example of a hook domain where you can see the alpha -fold database has been used to extract structural homologs.

464

So that's something that might be interesting to you.

465

I'll end this section with just some limitations and some advanced use cases.

466

I mean, it's important to think about isoforms when you're thinking about structure prediction.

467

You might have an isoform that is predicted to bind, and then another isoform that you don't think should bind, but is still predicted to bind.

468

Because the way that these things work is that you have an interaction interface.

469

If there's a lot of homology at the interface, then small changes between isoforms that should break this interaction actually might not be picked up.

470

And so you might get false positives from isoforms in the family, which we've seen.

471

Well, the numerization state isn't always obvious, right?

472

So this protein, is it a cold coil or is it a three -way cold coil?

473

It's not clear, right?

474

There's not an obvious metric you can use to say, actually, this should be a dimer, not a trimer.

475

We talked about alternate conformations and playing with the MSA.

476

There's also a way you can cluster the MSA, right?

477

You can take sequences that look like each other and then put those and use those as input,

478

and then take other sequences that look like each other

479

and use those and maybe those relate to different conformations of the protein right

480

so you can play around with the msa as a way of getting those out

481

and that's something like an alpha cluster let's say

482

of course post -translution modifications aren't going to be there right

483

there's not going to be a phosphorylation from alpha fold

484

so sort of might be very important mutations may not break

485

an interaction for the same reasons the isoforms won't right

486

because one residue will not change the a co -evolutionary signal of that interface necessarily.

487

So I've tried this with computations that I know should break the interaction where alpha fold will still predict interaction.

488

So generally, one should not use alpha fold to predict if mutation will break an interaction.

489

There's some evidence that rosetta fold might do better for this, but I mean, you should be careful with these things.

490

Ligens and nucleic acids are now handled really well in alpha fold three

491

and rosetta fold all atoms so that's not something

492

that we have to worry about anymore now whether

493

or not these things are accurate we're still you know it's still early

494

but it seems like it's working for many things

495

antibodies are difficult

496

because you have these rapidly evolving regions of protein of an antibody

497

that's trying to you know bind to its antigen and

498

so these rapidly evolving bits aren't really well handled by things like alpha fold

499

that use you know co -evolutionary information over long you know, over a long historical time, essentially.

500

And so there are dedicated things that you can try to run, but this tends to be a limitation.

501

And for proteins with very low homology, which essentially means that they don't have a lot of related sequences,

502

so the MSA depth is going to be very low.

503

Those things also...

504

are limited by this method.

505

You can try things like curating your own multiple sequence alignment, which can help.

506

We're using these protein language models, for sure predictors, which don't need an MSA.

507

But they may suffer from the same caveat, since the evolutionary relationships are essentially embedded in them.

508

So it might not actually solve the problem, but worth trying.

509

All right.

510

Are there any questions on this part before I go to find your site?

511

Go ahead.

512

okay

513

yeah we work with a lot of clothes i have seen that before i don't know why

514

in some cases it does it doesn't in the wrong orientation

515

but yeah something to be careful about yeah no not really um you could you do

516

yeah no i don't have any i haven't i haven't had

517

that problem enough times to have to think about a trick for

518

that um you could try different lengths i mean that's the only thing i could think of right

519

if you give it a longer full coil to work with maybe it'll then properly place it but that's yeah anecdotal

520

anything else all right

521

so what comes out of the alpha fold era is this ability to hallucinate proteins now

522

and that's essentially what binder design is you're trying to design a protein

523

that binds another protein using the kind of algorithms

524

that have been developed for alpha fold

525

and this is the general pipeline right you have a target

526

you want to bind something to where those yellow residues are

527

so those are hot spots and you want a certain length of protein let's say 60 residues.

528

And then what these pipelines are doing is hallucinating a backbone.

529

So either running it through alpha fold and just hallucinating a protein, seeing if it binds, and then generating some loss function so

530

that it continues iterating until it actually figures something out that might actually fit that space or something like RF diffusion.

531

And then once you have a backbone, you can optimize the residues within it using something like protein MPNN.

532

And then that will give you something

533

that actually looks like a protein is soluble maybe you can express

534

and purify it and then scoring

535

that okay does this look like a protein protein interaction is

536

the very surface area good is our predicted energy of binding right I mean are the alpha fold metrics

537

you know reasonable

538

and then you can use your top rank ones to then test using my physical techniques

539

and I think now what's nice is

540

that a lot of the difficult parts of this have been package together

541

so that you can just run your target through something like

542

Minecraft or World Shin and then you just get your list out on the other side

543

and not have to worry about you know necessarily installing each one of the individual components and so for something like

544

Minecraft

545

which uses this alpha fold hallucination here

546

I'll just show you the figure from the paper right

547

we have our target protein again we're using alpha fold to hallucinate some backbone

548

we're using this mpnn to optimize the sequence

549

and then we're running through alpha fold again to say okay does this look like a protein

550

you know will it bind and what's nice is

551

that you know more recently the success rate of these binder designs have actually gone up quite a lot

552

so we have now nanomolar affinity binders in you know in these cases at least you're getting quite high

553

success right

554

so I mean in some some cases

555

it is you know like a hundred percent with this case of course

556

this it depends on your target some targets are very difficult

557

and very hard to design binders for which I've also experienced

558

but what's nice about this is at least a simple input you know there's minimal intervention involved

559

and depending on your target and you know the success rate might be quite good

560

and at the MLB you know you can run this quite easily

561

if you just run bindcraft

562

and then a json file you'll essentially run it on our cluster

563

and this just has you know where you want to run

564

it what it's called what the actual target is the pb file

565

which chains where on

566

that chain you want to actually target this binder how long

567

you want this binder to be the range of lengths

568

and the number of final designs before it actually ranks them

569

and so here's something to think about is how big is my target how big is my binder

570

You don't want to make it too big because then it's going to just take too long to get enough binders.

571

So you want to keep it less than 500, let's say.

572

And there's also other parameters you can play with.

573

If you run the Bindcraft just by itself, you can see where these advanced and filter parameters are if you want to play around with them.

574

We can talk about it if you want.

575

But essentially, you'll get, let's say, hundreds of designs.

576

Here are some examples of ones that I run, just a subset of them.

577

And, you know, some of the things that I've noticed is that there are,

578

you know, sticky parts of proteins that binders will like to stick to, especially with this Minecraft implementation.

579

You know, it will like to bind to a hydrophobic region of your protein, for example.

580

And so, you know, sometimes you're kind of stuck.

581

Sometimes you can play with different parameters.

582

Like if your hotspots are ignored, you can maybe mutate those non -hotspot regions to lysines and see if you can trick it.

583

Or you can try playing with the interchain contact weights and then, you know, see if that will work.

584

But essentially, you'll end up with this long list and, you know, you got to cross, you have to, you know,

585

purify these proteins across your fingers if they actually bind, right?

586

you can use additional metrics you know like alpha 3 predictions

587

if alpha 3 doesn't predict this interaction of the binder maybe

588

it's not worth investigating further for example um i've used

589

that in my pipeline you know recently i had you know

590

i ordered 20 binders out of a list of like 600

591

and about half of them bound

592

so this actually it seems to be working for some targets for others i can i could get any problems

593

but it's a nice tool so i you know encourage you to try

594

if That's something that might be useful for either alternating confirmation of a protein, blocking a site, and all the things that you might want,

595

labeling, for example, all the things you might want a binder for.

596

All righty.

597

That is what I wanted to talk about.

598

These are just some resources, some things that I found useful.

599

For example, how to interpret outfold structures from the MPL is very nice.

600

This is the original talk about Colabfold, where they naturally outline this MMSeq and the alpha pipeline.

601

The illustrated alpha fold is an excellent write -up of alpha fold 3 in great detail

602

and a comparison with alpha fold 2.

603

And then Rosetta Commons has this course online, this YouTube course, about structure prediction and machine learning approaches and structure

604

structural biology and I would encourage you to look at that if that's something you're interested in.

605

We also have this listserv if you want to talk about alpha fold really good things.

606

All right, thank you.

Tải Ứng Dụng

Mọi thứ bạn cần để nói trôi chảy

Chấm điểm AIChấm từng câu nói

Luyện IPAThành thạo mọi âm

Từ vựngMở rộng vốn từ

Game Từ vựngHọc qua chơi

Quét QR

Download on theApp Store Get it onGoogle Play

FOR YOU

Video gợi ý

75 câu

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

Daily ConversationUniversity Students

53 câu

System Design Was HARD - Until You Knew the Trade-Offs

Daily ConversationUniversity StudentsWorking Adults

83 câu

SOFI: Sequence-based, cross-sectoral, One-health surveillance of food-borne infections | IMMEM XIV

Daily ConversationUniversity StudentsWorking Adults

108 câu

LTEC 782: Design-Based Research in Education (Session 01)

Daily ConversationUniversity Students

125 câu

Introduction to statistical prediction models

Daily ConversationUniversity Students

182 câu

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

Daily ConversationUniversity StudentsWorking Adults

346 câu

Trust the Process: How to Design Teams That Actually Work

Daily ConversationUniversity StudentsWorking Adults

113 câu

Architectural Design Process Explained

Daily ConversationUniversity StudentsWorking Adults

Từ vựng và ghi chú luyện nói cho bài học này

Bài luyện nói này dựa trên video “[TALK 13] Structure Prediction and Design using AlphaFold – Sami Chaaban”. Những từ được nhắc lại nhiều nhất trong bài: protein, structure, alpha, fold, sequence. Video này có 606 câu và 7888 từ để luyện shadowing. Phần lời nói dài 41:41. Người nói nói nhanh, khoảng 189 từ mỗi phút, nên sẽ có nhiều chỗ nối âm và nuốt âm. Chỉ 81% số từ nằm trong 3.000 từ tiếng Anh thông dụng nhất, nên từ vựng khá khó.

Từ vựng quan trọng trong video

15 từ đáng học trong video, kèm phiên âm và nghĩa:

Từ	Phiên âm	Nghĩa
fold động từ	/ˈfəʊld/	gấp, gập
sequence danh từ	/ˈsiː.kwəns/	dãy
predict động từ	/pɹɪˈdɪkt/	dự báo
residue danh từ	/ˈɹɛzɪduː/	bã
domain danh từ	/ˌdəʊˈmeɪ̯n/	đất đai, tài sản
coil danh từ	/kɔɪl/	cuộn cảm
rank danh từ	/ˈɹæŋk/	hàng
database danh từ	/ˈdeɪtəˌbeɪs/	cơ sở dữ liệu, CSDL
generate động từ	/ˈd͡ʒɛn.ə.ɹeɪt/	sinh ra, gây ra
metric tính từ	/ˈmɛt.ɹɪk/	mét
submit động từ	/səbˈmɪt/	đệ trình, nộp
interact động từ	/ɪn.təˈɹækt/	tương tác
versus	/ˈvɝsəs/	đối, đối đầu
module danh từ	/ˈmɑd͡ʒul/	mô-đun, bộ phận rời
interpret động từ	/ɪnˈtɜː.pɹɪt/	diễn giải

Cụm động từ bạn sẽ nghe

Từ	Nghĩa
check out động từ	tính tiền, trả phòng
go up động từ	lên
pick up động từ	nhặt, lượm
switch off động từ	tắt

Những câu đáng luyện lại

Các câu ngắn, trọn ý trong video mà bạn có thể dùng lại khi giao tiếp hằng ngày:

How do I cut my proteins so that I can purify better?
I'm not going to get too deep into it.
You'll get your multiple sequence alignment.
I'll talk about it in a second.
I mean, they're very far apart.

Ngữ pháp trong video

Những cấu trúc người nói dùng nhiều nhất, kèm đúng cụm từ trong video:

Cấu trúc	Trong video
Thì hiện tại hoàn thành tiếp diễn have/has been + -ing — việc bắt đầu từ trước và vẫn đang tiếp tục	we've been talking
Thì hiện tại hoàn thành have/has + quá khứ phân từ — việc đã xảy ra nhưng còn liên quan đến hiện tại	have been · has been monitored · we've now come
Câu bị động be + quá khứ phân từ — nhấn vào việc xảy ra, không phải người làm	can be tested · has been monitored · can be predicted
Mệnh đề quan hệ who / which + mệnh đề — thêm thông tin về người hoặc vật	Baker, which is · algorithm, which is · notebook, which is

Phát âm cần chú ý

Người nói dùng 149 dạng rút gọn, ví dụ you're, you'll, we're. Hãy nói theo dạng ngắn đúng như bạn nghe.

Âm “sh” và “zh”: prediction /pɹɪˈdɪkʃən/, essentially /ɪˈsɛnʃəli/, interaction /ˌɪn.təˈɹæk.ʃən/, structural /ˈstɹʌk(t͡)ʃəɹəl/, confirmation /ˌkɑn.fəɹˈmeɪ.ʃən/
Từ dài — đặt trọng âm cho đúng: essentially /ɪˈsɛnʃəli/, interaction /ˌɪn.təˈɹæk.ʃən/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, confirmation /ˌkɑn.fəɹˈmeɪ.ʃən/, speculation /ˌspɛk.jəˈleɪ.ʃən/

Những âm người Việt thường gặp khó:

Cụm phụ âm cuối — đọc đủ từng âm, đừng nuốt âm cuối: fold /ˈfəʊld/, sequence /ˈsiː.kwəns/, predict /pɹɪˈdɪkt/, alignment /əˈlaɪn.mənt/, bind /ˈbaɪ̯nd/
Âm /tʃ/, /dʒ/ và /ʒ/ — tiếng Việt không có, đừng đọc thành “ch” hay “gi”: generate /ˈd͡ʒɛn.ə.ɹeɪt/, module /ˈmɑd͡ʒul/, encourage /ɪnˈkʌɹ.ɪd͡ʒ/, fetch /fɛt͡ʃ/, biology /baɪˈɑ.lə.d͡ʒi/
Âm /æ/ — mở miệng rộng hơn âm “e”: alpha /ˈæl.fə/, interaction /ˌɪn.təˈɹæk.ʃən/, rank /ˈɹæŋk/, interact /ɪn.təˈɹækt/, whereas /(h)wɛɹˈæz/

Cách luyện với video này

Nghe hết video một lần, chưa cần nói, và ghi lại những từ bạn chưa biết.
Bắt đầu ở tốc độ 0,75×, nói đuổi từng câu, rồi quay lại tốc độ bình thường khi đã quen.
Ghi âm giọng mình rồi so với bản gốc, chú ý các từ như fold, sequence, predict.

Phương Pháp Shadowing Là Gì?

Shadowing là kỹ thuật học ngôn ngữ có cơ sở khoa học, ban đầu được phát triển cho chương trình đào tạo phiên dịch viên chuyên nghiệp và được phổ biến rộng rãi bởi nhà đa ngôn ngữ học Dr. Alexander Arguelles. Nguyên lý cốt lõi đơn giản nhưng cực kỳ hiệu quả: bạn nghe tiếng Anh của người bản xứ và lặp lại to ngay lập tức — như một "cái bóng" (shadow) đuổi theo người nói với độ trễ chỉ 1–2 giây. Khác với luyện ngữ pháp hay học từ vựng bị động, Shadowing buộc não bộ và cơ miệng phải đồng thời xử lý và tái tạo ngôn ngữ thực tế. Các nghiên cứu khoa học xác nhận phương pháp này cải thiện đáng kể phát âm, ngữ điệu, nhịp điệu, nối âm, kỹ năng nghe và độ lưu loát khi nói — đặc biệt hiệu quả cho người luyện IELTS Speaking và muốn giao tiếp tiếng Anh tự nhiên như người bản ngữ.

Phương pháp shadowing: đọc hướng dẫn từng bước đầy đủ →

Luyện nói tiếng Anh bằng Shadowing qua video: [TALK 13] Structure Prediction and Design using AlphaFold – Sami Chaaban

Chọn chế độ luyện tập

Tải Ứng Dụng

What AI can (and can't) currently do in a real lab | Jennifer Doudna

Executing Monitoring, Evaluation and Learning (MEAL) | Complete Practical Guide

Sec Growth DataScience staff meeting Sep 14 2022

Cell Biology | Cell Structure & Function

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

System Design Was HARD - Until You Knew the Trade-Offs

SOFI: Sequence-based, cross-sectoral, One-health surveillance of food-borne infections | IMMEM XIV

LTEC 782: Design-Based Research in Education (Session 01)

Introduction to statistical prediction models

We Let $200 GPT-Pro Think for 7 Days on Unsolved Math

Trust the Process: How to Design Teams That Actually Work

Architectural Design Process Explained

Từ vựng và ghi chú luyện nói cho bài học này

Từ vựng quan trọng trong video

Cụm động từ bạn sẽ nghe

Những câu đáng luyện lại

Ngữ pháp trong video

Phát âm cần chú ý

Cách luyện với video này

Phương Pháp Shadowing Là Gì?