Better docking with AI?

Welcome to Plenty of Room!

Happy Nobel week! Today, we are taking a closer look at using machine learning to improve computational drug discovery. Just to stay with the theme of the Nobel in Physics and Chemistry

Plenty of Room is your guide to the cutting-edge news related to molecular machines. New here? Just go ahead and subscribe! Already subscribed? Help a friend or a colleague save some time and share this with them!

Let’s get into it now.

Better docking with AI?

Discovering new drugs is hard work. To simplify the process, scientists use structure-based virtual screening to sift through massive libraries of potential compounds, narrowing them down to those worth pursuing. And the libraries screened are enormous: billions of compounds, with a “B”! Or at least, this would be the dream: unfortunately, screening libraries this big is incredibly challenging, and only a handful of times has it been successfully pulled off. The real bottleneck is the accuracy of the ligand docking programs used to predict how small molecules (ligands) bind to their target proteins. For those unfamiliar, molecular docking is a computational technique that predicts the most stable pose of a ligand in a protein's binding site, helping scientists figure out which compounds might be potential drugs.

That brings us to today’s paper, where the authors set out to overcome these limitations and build a method capable of screening multi-billion compound libraries. The authors presented RosettaVS, a new open access virtual screening platform, based on some previous tools and lot of cool optimization to speed up the hunt for drug candidates while improving docking accuracy.

The new platform combines traditional structure-based virtual screening methods with cutting-edge machine learning and artificial intelligence (AI):

• Active learning to prioritize compounds: Rather than docking every single molecule (which would take forever), RosettaVS uses AI-driven active learning to prioritize compounds for simulations. It starts with a small subset of molecules, learns from the docking results, and improves its ability to predict which compounds are most likely to bind successfully. This way, the platform balances computational efficiency and accuracy, saving time and resources.

• Modeling protein flexibility: Unlike traditional docking methods that assume proteins are rigid, RosettaVS accounts for protein flexibility during ligand binding. This makes it more effective at predicting binding interactions, especially when conformational changes during binding are involved. RosettaVS leverages the Rosetta General Force Field (RosettaGenFF-VS) to simulate receptor flexibility and improve docking precision.

This smart blend of machine learning and optimized molecular docking allows RosettaVS to scale efficiently to multi-billion compound (amazing!) by integrating with high-performance computing clusters. Plus, the platform excels in accurately docking ligands to small or polar protein pockets, which are often challenging for traditional docking algorithms. Finally, RosettaVS is available as an open-source tool, making it accessible to the broader scientific community and ensuring that it can be widely adopted!

To demonstrate the effectiveness of RosettaVS, the team focused on two drug targets with therapeutic potential:

1. KLHDC2: this is a human ubiquitin ligase, an enzyme involved in tagging proteins for cellular degradation. Targeting this type of enzyme has been catching a lot of attention in drug discovery, especially for cancer therapies. RosettaVS screened billions of compounds for potential binders and identified seven hit compounds with binding affinities in the micromolar range. One of these hits was further validated using X-ray crystallography, confirming that RosettaVS's model accurately predicted the ligand’s binding pose.

2. NaV1.7: this is a human voltage-gated sodium channel, playing a key role in pain signaling. Blocking this channel is a promising strategy for developing non-opioid pain medications. RosettaVS screened identified four potent compounds, with one compound exhibiting an IC50 of 1.3 μM (a measure of the compound's ability to inhibit the target’s function).

So, to wrap things up, this was a very exciting application of AI in drug discovery. I actually really liked that machine learning here is not used to replace existing methods, but to enhance and supercharge them. It’s also great to see a very robust validation of the compounds identified!
I would be remiss if I didn’t properly highlight other two aspects:

• Speed: RosettaVS did all these massive screenings in under seven days for both targets. This in an insane improvement from traditional methods, which can take months!

•  Accuracy: RosettaVS outperformed standard virtual screening tools, delivering high accuracy in binding affinities and ligand poses.

So, fast and accurate, what else can you want! But there is more in the paper, so I recommend you read it!

In other news:

• And that’s a FACT: Continuing with our series on terrible puns, this study shows how Chd1 and the FACT complex work together to maintain nucleosome organization during transcription. At the 5′ ends of yeast genes, RNA polymerase II generates hexasomes. Chd1 shifts these hexasome-nucleosome complexes, and FACT restores missing histone dimers, activating Chd1 to remodel the chromatin. Pretty cool stuff!

• Better modifications for nucleotides: Modified nucleotides are increasingly used in biotech and therapeutics (the Covid vaccine anyone?). But making them is not easy: so, this study presents a new method for synthesizing chemically modified oligonucleotides using template-dependent DNA ligation with T3 DNA ligase. It offers improved scalability and flexibility, producing oligonucleotides of varying lengths (20–120 nucleotides) and tolerating diverse modifications.

• Synthetic flagellar motors: Bio-inspired chemistry is always cool, and this paper doesn’t disappoint. The team introduced a novel mechanism for converting chemical energy into mechanical work at the nanoscale, inspired by the ATP-powered archaeal flagellar motor. The researchers designed micron-long peptide ribbons that catalyze a chemical reaction, driving the ribbons to morph from flat to helical structures, eventually forming spinning tubes. This is an interesting direction to develop new micro- and nanoscale machines!

• Not yet a subscriber to Plenty of Room? Sign up today — it’s free!

• You think a friend or a colleague might enjoy reading this? Don’t hesitate to share it with them!

• Have a tip or story idea you want to share? Email me — I’d love to hear from you!

• You have something you would love me to cover? Just reach out here or on my social!