Today at 11am I’ll present some of the exciting work in AI x Bio we’ve been up to at @Latent_Labs at the @RenPhilanthropy booth (Hall A/B, booth 1343 @NeurIPSConf - straight through the entrance at the back left). Come say hi!

Congratulations David! This picture was taken just 10 days before he got the @NobelPrize - I’m glad I got this one signed 😅 — I’ve never seen a lab like the @UWproteindesign Such an inspiring place ✨ And congratulations of course also to John Jumper and Demis Hassabis!

Come with us about enzyme engineering with AI and ultra high throughput screening at our poster at @gembioworkshop at #ICLR2024 with @MaxGantz_

Conditioning diffusion-based generative models for motif scaffolding? Come check out our talk at 11.05 in Room 242 at @AI4D3, or visit our poster in the break. @DidiKieran Francisco Vargas @vdutor @MathieuEmile @julia_tweeting_

Complementary to PoseChecker, my friend Martin Buttenschön released “PoseBusters” — where he critically examines the outputs of AI small-molecule docking models! Super insightful- also highly recommend a read! (Plus what an awesome name 👻) arxiv.org/abs/2308.05777

Enjoying @icml_compbio in Hawaii 🌺 with @antonia_boca, a fantastic Master’s student in our lab! She has a poster on her super cool work on protein engineering, so come check it out! 👀 – at Honolulu, HI

12/11 Some nitty-gritty details for those who stayed till the end of this thread 😉 GWL focuses on models which adhere to a "reference-free" approach. I.e. GNNs with local message passing that does not fix a canonical frame for the full graph or for each neighbourhood.

🥳 Thank you to all who made this paper possible @chaitjo @crisbodnar @TacoCohen @pl219_Cambridge and @chaitjo for assembling this great team! It was an insightful and fun journey!

11/11 In summary, we can place various models on the map of these 3 key design choices, allowing us to better understand what a given geometric GNN can or cannot do. For details: Paper (arxiv.org/abs/2301.09308), Github (github.com/chaitjo/geometric…), Talk (youtube.com/watch?v=OutT6qGs…)

10/11 Finally, our third experiment uses the examples from @spozdn to show how *single layers* which build features of higher body order can distinguish trickier local neighbourhoods.

9/11 Our second experiment studies whether an *individual layer* can distinguish the relative orientation of two neighbourhoods with high rotational symmetry, demonstrating the relevance of the higher angular resolution of higher-order equivariant representations (l>1).

8/11 Our first experiment ("k-chains") studies how fast and reliable information is passed on through stacking multiple layers of message passing.

6/11 💡 It shows that equivariant geometric GNNs are strictly more powerful than their invariant counterparts (unless the graph is fully connected). Intuitively, the equivariant information tells us how different neighbourhoods should be "stitched together" to give the full graph

5/11 We therefore derived a Geometric WL test (GWL) which generalises WL to the geometric setting. GWL upper bounds what graphs a geometric GNN can distinguish, and therefore in principle produce different outputs for.