GNNome was published in @genomeresearch! This is a novel paradigm for de novo genome assembly based on GNNs. Without explicitly implementing any simplification strategies, it can achieve results comparable or higher than other SOTA tools. Paper, code, and overview are 👇 [1/8]

RT @KexinHuang5: 📢 Introducing Biomni - the first general-purpose biomedical AI agent. Biomni is built on the first unified environment fo….

RT @avapamini: How well do single-cell foundation models perform w/o finetuning?. Our work @GenomeBiology shows that in zero-shot settings,….

RT @chaitjo: Equivariance is dead - long live equivariance?. Scaling pure Transformers is the future of molecular generative models?. I thi….

RT @michael_galkin: 📣 Our spicy ICML 2025 position paper: “Graph Learning Will Lose Relevance Due To Poor Benchmarks”. Graph learning is le….

There are also a few notable limitations worth mentioning: GNNome is limited to OLC assembly graphs constructed by other tools. Moreover, the focus was only on haploid assembly with a single sequencing technology. We leave these for future work. [8/8].

To achieve this, we leveraged symmetries available in the assembly graphs in various steps of the pipeline: model architecture, loss function, graph partitioning, node masking, and decoding all employ symmetrical strategies, which were critical for the success of GNNome. [7/8]

We also show that the model trained only on HiFi data can be directly transferred to ONT assembly graphs generated with Raven. Though such a direct transfer is suboptimal, the assemblies are still of reasonable quality and often better than those of other tools. [6/8]

Particularly interesting were the cases where GNNome managed to correctly assemble genomic regions, with no fragmentation or repetitiveness, in which some other assemblers failed. One such was observed in the case of a human chromosome 11. [5/8]

By only predicting which edges lead to the optimal reconstruction, GNNome achieves results comparable or higher than other SOTA tools when paired with hifiasm and PacBio HiFi data. By not implementing any simplifications explicitly, GNNome is flexible and adaptable! [4/8]

While the previous version of the tools was limited to graphs generated with the Raven assembler and was evaluated on individual human chromosomes, now it can take any OLC assembly graph as input and reconstruct genomes of entire organisms. [3/8].

First, I would like to thank everyone without whom this work would not be possible: @msikic, @xbresson, @Martin_fschmitz, Thomas Laurent, and Kenji Kawaguchi. Paper: Code: 🧵 You can find the brief overview below! [2/8].

RT @msikic: Cannot agree more. Though-provoking. Recently, we showed the implementation of GNNs on genome assembly (huge graphs). https://t….

RT @ai4na_workshop: 🏆 We would once again like to thank our sponsors, who made it possible for us to provide you with lunch, and some of yo….

If you have any feedback for us, we would love to hear it! Please take a few minutes to tell us what you think:.

Organizing @ai4na_workshop was as challenging as it was rewarding. The room was full, the talks and posters interesting, and the panel engaging. This would not have been possible without @TinVlasic, @im50603 and @msikic. For the first-timers, I think we did a good job 😌.

I was very happy to see that AI conferences realized the world is bigger than the US, and what makes me happy now is how many people discovered how wonderful Singapore is :).

The account seems to be inactive, but this has never been more true.

RT @DBahdanau: Adam deserves the award, but in Singapore everyone still uses SGD.

RT @ai4na_workshop: Looking forward to seeing you tomorrow @iclr_conf ! Our workshop will start at 9 am and will take place in rooms Opal 1….