In Silico Docking Used in First LabDAO Community Preprint

In October of 2022, a tweet from Gabriele Corso caught the attention of the LabDAO community. A new protein-ligand docking method called DiffDock was now available, with purported improvements on its predecessor algorithm, EquiBind. In an edited interview with LabDAO co-founder Jocelynn Pearl, Niklas Rindtorff shares the story of how LabDAO’s first scientific paper came to be.

[embed]

Jocelynn

How did you first hear about the DiffDock tool?

Niklas

The first time I heard about DiffDock was on Twitter. The first author, Gabriele Corso, is a scientist at MIT, co-advised by Regina Barzilay and Tommi Jaakkola, had already published in the space before. EquiBind, published by Hannes Stärk, was one of the first graph neural networks for protein-ligand docking. It was already on my radar that this was a group that was working on some really interesting machine learning tools for computational biology.

Kelvin Wallace, a gifted undergraduate student, and I had EquiBind running internally and we were already playing around with it. We were trying to generate hypotheses around the docking position of BIOIO-1001, a molecule that was being studied by BIOIO at the Washington University in St. Louis, but ran into some well-known accuracy problems with EquiBind.

When DiffDock came out, it was clear very quickly that we needed to containerize this tool, make it available to the community, and use it for this manuscript.

We got onto DiffDock and it was way more accurate compared to all the tools we benchmarked it against.

Containerizing these research-grade tools is pretty tricky. You need a lot of frustration resilience to work through all the little package incompatibilities. But Kelvin did an awesome job getting this tool to run and we were able to help the group of researchers around Tim Peterson to publish the paper.

Predicted small molecule docking pose 278 of BIOIO-1001 with SIRT3. Source: Li et al, Biorxiv

Jocelynn

That’s awesome. Is this tool now available to anyone?

Niklas

Yeah. The DiffDock container is publicly available for everyone on GitHub. Gabriele Corso maintains the main repository on Github, and LabDAO is maintaining a fork as well, so if you go to github.com/LabDAO/diffdock, you can find that container. There is a Dockerfile or you can directly download the latest container and run it yourself. And it’s actually been downloaded already a couple of times.

Example view of a ligand docked to a protein structure using DiffDock

Niklas

Here you can see an example interaction that we have generated. During early stage drug discovery, you want to predict how a small molecule and a protein target can interact. One downside is that running one job usually takes ten minutes — which is still faster than most other tools. That’s why we also implement EquiBind, which is even faster.

Jocelynn

And how many of these interactions can you map at once? Just one? One to one?

Niklas

No, the cool thing is you can do many to many relatively easily. We’re currently building PLEX, which is a command line client that makes scientific tools easily accessible on the Lab Exchange compute network. We built PLEX in a way that you can just point it to any directory and it will automatically dock all files it can access . We are, however, currently running into problems with scalability. Right now we just have one set of GPU nodes; the queue gets built up quickly, so we need more people contributing some of their GPUs.

Jocelynn

So our hope is that as a community we’ll have more resources as far as GPUs to run things like this?

Niklas

Yes, if you have a GPU right now and are familiar with Linux and the command line, you can help us with your GPU today. There are installation instructions to contribute a node on our Github.

Jocelynn

What does this docking information mean? How does one interpret the results?

Niklas

Docking results generally should be interpreted with great care. The tools, no matter how good they are right now, are still inferior to any experiment. That said, it’s pretty exciting how good they have gotten lately.

DiffDock retains a large part of its accuracy even when run using computationally predicted protein structures. Source: Gabriele Corso, Twitter

The team that developed DiffDock shared new comparison data to other tools using predicted protein structures a couple days ago.

DiffDock is basically outperforming not only EquiBind, which was developed in that same group, but also state-of-the-art docking tools like SMINA, a bread-and-butter open source tool.

GLIDE, another tool, is maintained by Schrödinger, a for-profit company that is selling licenses for this tool. As you can see in the figure above, It turns out that GLIDE has lower accuracy and it takes about 1,000 seconds to run on their setup, while DiffDock runs in 10–40 seconds and is far more accurate.

Jocelynn

Wow.

Niklas

This table is why I can’t sleep at night. We’re currently living through this moment where the tools that have been necessary to develop drugs using computers are suddenly getting way better and they’re also open source. I think we’re seeing a step-change where accuracy is improving at a rate that will have some really interesting second-order effects in the coming months.

Surely, we need to be careful because these models are trained on data that might limit their ability to generalise very well. A lot of models are trained and tested on a part of a big data set that’s called PDBind.

To summarise, we still need to look at these results with some caution — but overall, I can’t help but get pretty excited about the next couple of months and what they will mean for people that want to use open source tools to accelerate drug discovery.

Jocelynn

Wow. That’s exciting. Can you share how folks who are interested in this type of work can get involved within the LabDAO community?

Niklas

Right now, we are very much still focused on computational tools. That said, we love it when non-computational groups come up with a really applied question that they need help with from a computational side.

We have been building partnerships with CRO aggregators like Science Exchange and Cromatic; these are marketplaces that can help you find physical laboratories when you need them.

If you are a computational scientist and you want to run a computational job, we’d love to get your feedback on what we’ve been building. If you’re excited about the new models that are coming out just as much as we are, then you can also take a stab at containerizing them with us so that we can share them with everybody in the community.

Jocelynn

Awesome. One last question. Will there be more research papers coming out of LabDAO in the future?

Niklas

Yeah, absolutely, we’re just getting started.