#jobim2024 Idriss André, on graph representation learning and semantic distribution, with application to omics expression data.

Multi-omics integration of sample-related data, with "fusion" [better choice than "integration", indee., IMHO] algorithms. Introduces Knowledge Graphs. Nodes and edges have labels [important nuance].

Graph Embeddings: graph → vectors. Several approaches possible, uses WalkLM. Basically random walks the graph to build up a text sequence from labels. Then input to a LLM

#jobim2024 Toni Gabaldón, on tracing genomics adapations in emerging fungal pathogens.

Starts with a (rather sad) story of a patient dying from a rare fungus infection at the hospital. Sequenced in the lab, proved to be a new Pichia species. It is an hybrid!

This brings many scientific issues (negleoting fungus, opportunistic pathogens, drug resistance, difficult taxonomy and diagnosis, emergent species, …).

Convinced by the presentation of the Peer Community In system at #jobim2024, a free recommendation process of scientific
preprints based on peer reviews and a journal → https://peercommunityin.org/

They are also on Mastodon: @PeerCommunityIn

Paul Ficek's keynote about using data science to build a bridge from model systems to human disease, in #jobim2024

Starts with history of mouse models. Jackson ("JAX") lab founded in 1929 (by Clarence C. Little), with the objective to study genetic diseases on newly bred JAX mice®️ (sic). Continues to do so today. Data science is —like mice— scientific enablers.

Think about the mouse as a genome model. But different from human, but complementary. Mouse models for rare diseases are very good.

#jobim2024 Aurélien Beaude talks about
AttOmics: Attention-based architecture for diagnosis and prognosis from Omics data.
https://easychair.org/smart-program/JOBIM2024/2024-06-26.html#talk:261667

About personalized medicine, from omics data, using DL.

SoA notably talks about using ontological layers before auto-encoder layer, a method that I find interesting as well :-)

Introduces using self-attention, to pick relevant information from input. But a costly (quadratic) method! So proposes to group features and apply self-attention on groups.

Gwenn Guichaoua talks about drug-target interactions predictions, at #jobim2024.

Targets drug discovery (de-orphanising drugs, anticipate off-targets).

Use large database of interacting molecules, then do supervised ML (binary classification).

Introduces the new "LCIdb" database, built from consensus across 5 DBs (thousands of proteins, hundred of thousands of molecules).
Has good coverage of human proteins.

Prediction pipeline relies on features computations for proteins and molecules.