This work marks a key milestone in my PhD journey, where we've used AI to gain insights into cancer — a small step toward understanding this disease better and offering hope for the future. 🧬🤖

There are a lot more results detailed in our full text, showcasing the depth of our work. Immense thanks to all coauthors for their exceptional contributions. All code is available on tinyurl.com/5n6dtjmw. In addition a web server version is accessible at cellhit.bioinfolab.sns.it

We computationally validate our predictions on TCGA and experimentally on Glioblastoma (in collaboration with Fondazione Pisana per le Scienze) and on PDAC (in partnership with the European Institute of Oncology).

This approach includes leveraging proprietary models like GPT4 from #OpenAI but also deploying a locally executable pipeline using freely available Mixtral Instruct from #MistralAI, along with guidance and quantization strategies.

Using LLMs, we've created a pipeline to generate, starting from available metadata, drug descriptions which are then linked to @reactome.org 's biological processes. This novel method extends drug annotations and deepens MOA understanding via AI-enhanced biological pathway exploration

Leveraging the power of XAI techniques, we systematically asses whether trained models can uncover underlying molecular determinants and mechanisms of action of drugs. This serves a dual purpose: deriving new biological insights and ensuring the accuracy of existing knowledge 🔍

Two key diff: 1) This version uses PyTorch instead of TF, which I find easier to use (an error in TF on my setup prevented me from using the original umap-learn module). 2) The code is streamlined for a minimal yet efficient implementation. Simpler and easier to adapt (but missing some features).

The implementation closely follows the original work by @timsainburg.bsky.social and @lelandmcinnes.bsky.social , for which I am deeply thankful. While still a work in progress and a bit rough around the edges, it is already functional and ready to use

Key features include a straightforward Python implementation enhanced by PyTorch for efficient modeling and GPU acceleration, along with fast neighbor computation powered by FAISS. It supports batched training for scalability and provides a user-friendly, scikit-learn-style API for easy use

What's Parametric UMAP? It's like regular UMAP but learns a neural network that can map new data points without recomputing the entire embedding. Perfect for large datasets or when you need to embed new samples