[4/4] Our approach is by design a meta-method: you can use it with your favorite single-cell RNA-based GRN inference tool, and squeeze more insights out of your data! Check us out GitHub: github.com/aliaaz99/GRN....

#singlecell #GRN #LLM

[3/4] We use these embeddings to construct a prior graph and then further refine it with some known TF-target interactions as pre-training targets. Finally, we use this augmented prior graph jointly with a GRN inferred by *any* other method, in order to produce a final prediction.

[2/4] We use gene descriptions from NCBI Gene database and embed them into a high-dimensional space with a LLM (Qwen3-8B). This idea was inspired by GenePT (pubmed.ncbi.nlm.nih.gov/37905130/) and a great study on gene embeddings from @vyao.bsky.social's group (www.biorxiv.org/content/10.1...).

Nicolae Sapoval @nsapoval.bsky.social presented "Theoretical and empirical performance of pseudo-likelihood- based Bayesian inference of species trees under the multispecies coalescent"
A fantastic theory talk, offering intuitive insights!
Paper: doi.org/10.1101/2025.01.28.635282

As the next step, we aim to develop rigorous corrections to the pseudo-likelihood-based credibility intervals in order to further improve scalability and applicability of Baeysian phylogenomic inference.

In our work we explore suitability of pseudo-likelihood for Bayesian phylogenomic inference. We show that using pseudo-likelihood greatly reduces the computational burden of the Bayesian inference. However, the inferred credibility intervals are overconfident.

Likelihood-based phylogenomic inference is common, but it faces scalability issues. Hence, pseudo-likelihood has been previously proposed as a statistically consistent (for topology estimation) and scalable alternative: doi.org/10.1186/1471...

In case if you lose the URL (it’s not pretty), I have linked this on my website (nsapoval.github.io) as well.

This is aimed primarily at the people who are just starting their thesis-based masters or a PhD. However, it’s also an evolving document, so suggestions and ideas are welcome!