Michael Brocidiacono

PhD Student at UNC Eshelman School of Pharmacy. I teach chemistry to computers.

[Michael Brocidiacono @brocidiacono]

Overall, I'm glad to finally put this out. It's a cool method and will hopefully inspire future work. The whole MM/GBSA being almost as good as implicit solvent ABFE is also interesting in itself.

[Michael Brocidiacono @brocidiacono]

Excited to release our latest paper, Binding Free Energies without Alchemy: https://t.co/KhovIy2ZUD. Here we show that you can directly estimate protein-ligand binding free energies using only end-state simulation data if you use an implicit solvent model.

[Michael Brocidiacono @brocidiacono]

Ultimately, it turns out implicit solvent ABFE isn't much of an improvement over MM/GBSA on these ligand systems in general. Turns out explicit treatment of water is much more important than conformational entropy!

[Michael Brocidiacono @brocidiacono]

Lots of caveats ofc (no method did particularly well here -- lots of room for improvement!) If you're at NeurIPS and want to chat, please reach out!

[Michael Brocidiacono @brocidiacono]

I'll be at NeurIPS starting tomorrow! I'll be giving a poster at the MLSB workshop on Sunday on simulating training data for ML virtual screening models: https://t.co/OeMj3hy1ie

[Michael Brocidiacono @brocidiacono]

TL;DR -- our model helps when used as part of an ensemble. We're also beating Boltz-2 despite running ~100x faster.

[Michael Brocidiacono @brocidiacono]

@alantomusiak Thank you 🙏

[Michael Brocidiacono @brocidiacono]

Exciting to announce our latest paper on Ligand Force-Matching! Rapidly predicting how well a small molecule will bind to a protein is the bottleneck in virtual screening. ML methods can work well on targets similar to their training set, but often fail to generalize. (1/)

[Michael Brocidiacono @brocidiacono]

Check it out here: https://t.co/IGJQTsPuCF

[Michael Brocidiacono @brocidiacono]

Using this workflow, we're getting competitive performance at virtual screening across 6 targets -- but this is just the beginning. This is the first AI method that has the potential to work on even the weirdest target classes, such as disordered proteins or even RNA! (3/)

[Michael Brocidiacono @brocidiacono]

@richardfuisz I haven't looked into this too much, but light-modulated drug release is also pretty cool. Eg: https://t.co/7UzmWsjbuv

[Ian Dunn, PhD @ian_dunn_]

New preprint! “Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation” https://t.co/ciW5EYKwqj FlowMol is a flow matching model that jointly samples the geometric and topological structure of molecules. We also present surprising results on flow matching for categorical data!

[Michael Brocidiacono @brocidiacono]

I'll be presenting this work at @ACSCOMP this week - check it out if you're here!

[Michael Brocidiacono @brocidiacono]

Excited to share our latest paper! https://t.co/vF9Fn4nAjd We propose a new metric that can better assess model enrichment in realistic virtual screening scenarios. We then made the BayesBind benchmarking set, which we hope will become a standard VS benchmark in the field

[Michael Brocidiacono @brocidiacono]

I'd love to see what the community will do with this! I will personally buy drinks for the first group that gets >50 median enrichment (EFBmax) on this set