Big data and big computing have made leaps and bounds in drug discovery - but data quality and reference to experimental evidence are key to grounding these approaches. From ML training to validating approaches, experimental and computational techniques have to work together for progress - especially as we move toward more complex targets and drugs.
Day One: Wednesday 9th June
Inferring intramolecular potentials from crystallographic structure observations
Carl presented machine learning (ML) technologies used as part of the workflows in Astex Pharmaceuticals' TorML project. Astex Pharmaceuticals discovers and develops novel therapeutics with a focus on cancer and central nervous system (CNS) disorders. TorML examines conformational preferences of molecules during ligand-based drug discovery to highlight unusual conformations in a structure.
“We like our structures to be as usual as possible, as unusual as necessary!” Carl said, noting that unusual can indeed be usual in drug discovery.
TorML builds on the ability of Mogul to show torsion angles from known structures in the Cambridge Structural Database (CSD) but extends the capability with custom calculations to make use of proprietary data. This model was trained with CSD structures but uses different fitting to predict conformations. He showed specific examples of molecules where Mogul and TorML agree closely, and where they differed.
In his future work, he surmises that ML potentials can improve generalization. Further developments will include showing features other than torsion and showing the provenance of a prediction (included in Mogul by design). He stressed the importance of calibrating models with “ground truth.”
Blending experimental and computational approaches to assess protein target druggability
Natalie presented her recent work from the Cancer Research UK Newcastle Drug Discovery Unit. The centre has a long history of collaborations across industry and academia to drive the best outcomes and has seen successes with Rucaparib and Erdafitinib launched to market. The work presented focused on Fraglites, and their application to investigate undrugged or difficult to drug targets - often in unpredicted or previously unexplored sites.
A Fraglite is a small, halogenated probe used to perform rapid crystallographic screening and identify even weak or low-occupancy binding. Here she presented work that showed how Fraglites can differentiate between two very similar targets.
“These domains are structurally similar but sequentially divergent, and we see chemistry-based scores rank their druggability very differently. Not all domains are created equally druggable,” she said.
The team then expanded on their Fraglite research – utilizing molecular dynamics simulations that predict novel binding sites in the same way as Fraglites. This approach revealed under- and over-prediction issues as well as bias towards sites explored by experimental work. Future work will look to blend computational and experimental approaches further, with more high-quality fragment data and a computational pipeline to better inform the methods.
Day Two: Thursday, 10th June
Moving from the Newtonian to the Quantum Realm in docking and scoring
Robert presented recent work with Jimmy Stewart, Jason Cole and Mihaela Smilova that asks whether semi-empirical quantum mechanics (QM) can improve current docking methods. This sprang from efforts to find active molecules against COVID-19 MPro through a global crowdsourcing effort (the Covid-19 Moonshot).
The initial work highlighted deficiencies in current scoring paradigms. Typically docking uses a variety of algorithms and places ligands in poses which are then evaluated with an empirical scoring function. However these approaches do not take into account the electronics of the system, and the team wanted to examine if semi-empirical QM approaches could produce better results. Results were shared from several test cases and analyses, which employed MOPAC to evaluate the principally enthalpic contributions to ligand-protein binding. So far this approach has shown some promising and some puzzling results.
Their future work will look to speed up the protocol and explain when and why the approach works and the causes of outliers.
Fragment Hotspot Maps to Drive the Progression of Fragment Hits
Mihaela Smilova - University of Oxford
Mihaela presented recent work examining how the established fragment hotspots tool could be applied to ensembles of fragment-bound protein crystal structures. Fragment-based drug design is a powerful and experimentally well-established approach, but progressing hits to higher potency leads can still take years. By better understanding the binding sites and modes available, this could be expedited.
Her analyses involve using the Fragment Hotspot Maps method. This method uses SuperStar to generate a propensity map of the surface of the protein
– showing sites where apolar, H-bond acceptors and H-bond donors are available on the protein surface. These are further refined, accounting for pocket knowledge and molecular context, to give a fragment hotspot map.
The presented work looked at combining these maps for ensembles of structures of closely related proteins, to see if they can detect differences. The results are soon to be published but broadly find that such an approach is a good way to summarise binding information and generate hypotheses.
Watch Mihaela’s talk
Artificial Intelligence in Drug Discovery: What is Realistic, What are Illusions?
Andreas Bender - Cambridge University and Nuvisan Pharmaceutical Services
Watch Andreas' talk
Andreas Bender spoke in the use of AI in Drug Discovery, asking where are we now and what is needed to advance further?
From hype and predictions, to our present reality the presentation explains why it is more complex than it seems to apply artificial intelligence to pharmaceutical discovery and development.
Once again we extend a big thank you to all of our speakers for the fantastic presentations.
We hope you can join us at our next meeting -
so you don't miss out. subscribe to our monthly newsletter