GRAPE

#GraphML 🍇 preprint is on @arxiv! ❤️ to @LucaCappellett6, @Zommiommy, Casiraghi, @RavanmehrVida, @Tiff_callahan, @realmarcin, @chrismungall, @pnrobins, @justaddcoffee & Valentini 📍@MonarchInit, @ConsorzioCINI, @LaStatale, @jacksonlab & @BerkeleyLab 📜:arxiv.org/pdf/2110.06196…

Woohoo! #GRAPE just hit 100 stars on GitHub! Thank you to all the amazing developers who have supported our graph representation learning library. We couldn't have done it without you! 🍇💜🍇 #opensource #machinelearning

@cthoyt @keenuniverse @mnick Absolutely, HolE is more suited for KGs! This is just a quick example of how to use the one-liner. Nevertheless, Cora has two edge types in this instance: paper-to-paper and paper-to-word. First, I wanted to do an HPO example, but my GPU needs more memory.

I've been asked how to use 🍇 to run and visualize @keenuniverse's implementation of @mnick's HolE model, so here's the one-liner!

Pushing the ✉️ of 🍇's @psresnik score implementation by computing 3T, i.e. 3*10^12, scores from @NCBI Taxonomy (2438821 nodes) upper triangular matrix. This is heavily parallelized and takes ≈3h on a 💻 with 8GBs of RAM and 96 cores.

You can now use 🍇 to compute @hp_ontology pairwise @psresnik scores, ~270M, on your notebook in about 1 second ⚡

@DanZiemianowicz @MKoutrouli @phanein @LucaCappellett6 @zommiommy I suggest to join either the telegram group or the discord server to discuss these topics in details, without limiting the content in 140 characters. You can see on our account the game of Thrones link prediction, for instance.

@GRAPElib @MKoutrouli @phanein @LucaCappellett6 @zommiommy Interesting. I’m sceptical this is possible from such a network, based on data with uncertain reliability (oh the things I’ve seen), but I trust your efforts and want to learn more. For your two ex., can you recommend specific tutorials? They can be analogous scenarios.

🎥>>🗣️ #8: @MKoutrouli's FAVA functional association networks, embedded using @phanein's DeepWalk + SkipGram with Right Laplacian sampling by @LucaCappellett6 & @zommiommy Done in ~2m on my desktop! ⚡ The edge prediction looks excellent (holdout 70/30)! ❤️

@DanZiemianowicz @MKoutrouli @phanein @LucaCappellett6 @zommiommy An example is predicting protein interactions that are missing in a network or, vice-versa, identifying those that may be out of place. Protein function prediction may be a node-label task, and how different topologies alter the function of the same proteins may be explored.

@GRAPElib @MKoutrouli @phanein @LucaCappellett6 @zommiommy to be pragmatic: why or how can one use this tool to better understand the biology, and the perturbation thereof, of an organism?

@DanZiemianowicz @MKoutrouli @phanein @LucaCappellett6 @zommiommy Super briefly: 1. Node embedding algos aim to create matrices capturing nodes' topological & structural info 2. Used for 🎥, node-label & edge prediction 3. Edge properties are captured by DeepWalk SkipGram in the 🎥 4. 🎀 You can put the 🎥 in Google Slides, g8 for conferences!

@GRAPElib @MKoutrouli @phanein @LucaCappellett6 @zommiommy As someone new to the field but primarily a wet lab experimentalist (proteomics) with basic graph knowledge, can you explain a bit more what we’re looking at? I’ll certainly check out your extensive 🙏tutorial library once I wrap my head around it

@MKoutrouli @phanein @LucaCappellett6 @zommiommy We are happy to facilitate access and processing of FAIR data! Let us know whether you'd like to run a task on your graph!

@GRAPElib @phanein @LucaCappellett6 @zommiommy Looks AMAZING! Thanks for considering our network 🙏

@larsjuhljensen @MKoutrouli @phanein @LucaCappellett6 @zommiommy Absolutely! In figures f, g and j, the traditional edge metrics Adamic-Adar, Jaccard Coefficient and Resource Allocation Index are highly predictive. A different behaviour would be peculiar to this type of network.

@GRAPElib @MKoutrouli @phanein @LucaCappellett6 @zommiommy I'm guessing edge prediction is quite easy in this network due to it being a co-expression network. If A correlates with B, and B correlates with C, there's a pretty good chance that A correlates with C too.

🎥>>🗣️ #7: @justaddcoffee's KGCOVID19 node & edge labels + properties, embedded using First-order LINE ⚡

@cthoyt My newbie goal was to see whether we could 🔮 the encounter between Night King & Arya, but I learned that in the 📚 the Night King does not exist at all :p Nevertheless, the predictions are surprisingly good!

#GraphML & #MachineLearning to 🔮 #GameOfThrones 5th book using characters' interactions from all previous 📚!

Edge types of NPKG. The training is unsupervised. 💻 Tutorial:github.com/AnacletoLAB/gr… 💻 NPKG #github repo:github.com/sanyabt/np-kg

🎥>>🗣️ #6! Edge prediction of @sanyabt11's NPKG (745.51K N, 7.25M E) embedded using 🍇's #Rust + #Python implementation of @tangjianpku's First-order LINE, ~30s on my PC Test set, 70/30 split 💽 from @zenodo_org ❤️ Edge types & 💻 in the 1st comment! #GraphML #MachineLearning

The Twitter handle of 🍇's PI, Prof. Valentini, is @gg_valentini

#GraphML 🍇 preprint is on @arxiv! ❤️ to @LucaCappellett6, @Zommiommy, Casiraghi, @RavanmehrVida, @Tiff_callahan, @realmarcin, @chrismungall, @pnrobins, @justaddcoffee & Valentini 📍@MonarchInit, @ConsorzioCINI, @LaStatale, @jacksonlab & @BerkeleyLab 📜:arxiv.org/pdf/2110.06196…

Visualization of edge prediction on PharMeBINet, connected holdout with 70/30 split. Good separation between existing and non-existing edges is achieved, suggesting an edge prediction task could achieve good performance. 💻:github.com/AnacletoLAB/gr…

🎥>>🗣️ #5! Edge types of @AstrorEnales's PharMeBINet (2.65M N, 15.88M E) embedded using 🍇's #Rust + #Python implementation of @tangjianpku's First-order LINE, ~1m on my PC 💽 from @zenodo_org❤️ Edge prediction & 💻 in the 1st comment! #GraphML #MachineLearning