Protein design is an opposite of protein folding, just last week we saw AlphaFold2 claim to solve protein folding (finding the best structure given the sequence) with Deep Learning. Now this here DeepChain claims to do inverse protein folding, that is finding the best protein _sequence_, given the structure. This could be quite exciting, if done right. Can't wait to get myself an account and try it out...

I am Marcin, a part of InstaDeep's DeepChain team. I have been working with computational modelling of proteins, protein complexes etc. for quite a while now and I am rather excited to see a platform like DeepChain come to be.

If there is anything you want to know about what we do, especially from the Computational Biology point of view, do ask away!

This seems really interesting and can prove extremely helpful to experimental biologists who want to explore (and usually struggle with) the computational side of protein design. Although, I wonder if this can work seamlessly for a number of significantly large protein sequences (>500 aa long).

One more thing I wonder about is how well does DeepChain scale? If I had a large experiment, for example hundreds (thousands) of mutations to scan. Or multiple enzymes to redesign... Will their infrastructure manage?

Using AI for protein should be the future, website offers free trial too

But there has to be some sort of a catch... They offer the services at a surprisingly low cost. I wonder if it works, actually.

Free version is quite limited in functionality. You can play with language models but cannot run MD simulations or AI assisted mutagenesis.

Looks really exciting and well polished. I wonder, if I upload my data there, what sort of guarantee do I have, that it won't leak to my competitors?

I was sure attention-based models would be a game changer in biology and protein design! I think that attention is all you need :)

Very interesting, I wonder if this is what InstaDeep used for their Covid paper in NeurIPS

Marcin at InstaDeep here, I am one of the people behind DeepChain, and very eager to answer all questions.

I am glad to see you noticed our work on protein design for Covid (https://arxiv.org/abs/2012.01736), that is currently presented at NeurIPS. It has been a lot of hard work, but we think it turned out well.

You are partially right - the coronavirus work we did earlier this year to a large extent influenced DeepChain development, but - despite overlapping quite a bit - both have slightly different focus. The work you refer to was very result driven, aiming at providing the best result possible, with lesser focus on the underlying computational resources. We conducted a lot of experiments, developed a large amount of novel science processes, and protocols. Some of these worked well, others... well, we learned a lot too :-)

DeepChain builds on the coronavirus project, as well as other work performed at InstaDeep and outside. Its focus lies on making quite involved computational biology tasks approachable, usable, and useful to a non-expert user. The AI Designer module does what we did in Covid paper, but does it in an optimised way, delivering actionable results in a reasonable amount of time. It is thought as an exploratory tool, and we are happy to work with you to help you explore your DeepChain findings further.

DeepChain performs also one-click, short, reasonable Molecular Dynamics simulations to validate newly designed proteins, These too are very similar to the validations we did for the Covid paper, but contrary to the paper here it is the user, who gets to decide how involved the simulation needs to be. It is not a platform to run half-a-year worth of simulation in one go, but it works great for quick (and painless) evaluation of multiple poses in a uniform, reproducible manner.

Finally, the Sequence Playground is a way to tap into the fascinating (for me, coming from very structure-oriented background) world of very large AI models of protein sequences, parameterised on pretty much all that we know about proteins (both these annotated and these coming from metagenomic sequencing!). Here, within seconds, one can get an objective estimate of evolutionary pressures onto your protein. As these come from global models, you can (for example) see that a single mutation may make another position more susceptible to adopt a different amino acid, which in turn may potentially turn your entire protein topsy-turvy. What we see is that protein regions being very stable (structurally and evolutionarily) also have a very "tight" probabilities. The Playground allows you to explore any sort of variation in sequence and see how well it works out, given out knowledge on billions of years of evolution. Then you can feed the most intriguing designs into Site-directed mutagenesis tool in AI Designer and see how the mutation works out from the structure and interaction point of view.

As you see, I am rather excited about the platform, so if you want to know anything else, ask away!

Looks promising! Hope this will soon help deliver new therapeutics

Amazing platform to explore protein sequences

this is what the biologists need to push the productivity to an other level

The future of protein design !

It's very kind of you to say so. We at InstaDeep believe in democratisation of discovery, because it is only when appropriate tools become widely available, that some of the great science gets to happen. We know, that future of medicine lies in protein-based therapeutics and are firmly convinced, that approachable platforms, like DeepChain, are the way to get there.

Great platform!!!

This is awesome!