So excited about OnSIDES! This resource is desperately needed….

RT @HarvardDBMI: We're in a polar vortex now, but brighter days will come to Boston. along with the 21st year of our Summer Institute in….

What do you want to know?.

Kudos to @Forbes for sharing our work on TxGNN so broadly!.

RT @KexinHuang5: 📢 Super excited to share our new study @NatureMedicine on developing and validating an explainable graph-based foundation….

✅ Check out TxGNN’s predictions using our interactive visualization tool at I’m incredibly grateful to our amazing team @marinkazitnik, @KexinHuang5, @WangQianwenToo, Shreyas Havaldar, @AkhilVaidMD, @jure, @girish_nadkarni, and @BenGlicksberg!.

👀Interpretability matters! TxGNN offers interpretable multi-hop paths through its Explainer module, revealing the model's rationale and relevant biological entities behind each drug prediction. This transparency is critical for building trust with clinicians.

🔎 To ensure real-world relevance, we validated TxGNN’s predictions using a large EHR database @MountSinaiNYC. Here, TxGNN demonstrated strong alignment with clinicians’ off-label prescriptions for 1M+ patients.

🌟What makes TxGNN unique? Its *disease pooling module* aggregates sparse data using mechanistic similarities between diseases. This enables TxGNN to perform exceptionally well even when the data is limited!

💥While other AI models fall short when tested on diseases with no existing treatments, TxGNN shows up to 49.2% improvement in predicting drug indications and 35.1% in contraindications in challenging *zero-shot* tests across 11 diverse, real-world disease splits.

💡TxGNN addresses this challenge! By doing self-supervised learning on a massive biomedical knowledge graph, TxGNN is trained to predict both drug *indications* and *contraindications* for 17,080 diseases, including many that have zero FDA-approved treatments.

🔬 Globally, there are over 7,000 rare and undiagnosed diseases, yet only 5-7% have treatments, leaving the majority untreated. Zero-shot drug repurposing—identifying new uses for approved drugs—has the potential to identify therapies for diseases with no therapies.

📣Excited to share our new study in @NatureMedicine that @KexinHuang5 , @marinkazitnik, and I have been working on for the last four years! TxGNN is a graph-based foundation model for zero- shot drug repurposing that can find therapies for diseases with zero treatment options.

So grateful to have a wonderful mentor and advisor in @zakkohane and so excited to share this work with @HarvardDBMI community! 🙏🏼.

This year at #ML4H2024, I’m helping in bringing a new Perspectives Track! 🧠✨ We’re excited to feature insights from pioneers in the field. Our themes this year are:. 🤖 Foundation Models .🚀 Deployment. Who do you want to hear from? Share your suggestions!💡💯.

Starting my internship today at Apple Health AI in New York! Will be working on self supervised learning for time series.

RT @AIHealthMIT: Most self-supervised learning (SSL) methods for clinical time series data only use one data type e.g. vital signs or ECGs.….

🧪🧬🪐 Excited to share our perspective on how AI will transform science in the next decade! What a wonderful experience to collaborate with so many incredible co-authors across the globe 🥂.

RT @marinkazitnik: Excited to share our @Nature paper on the role of AI in scientific discovery 🌟🔬 #AI4Science. AI is transforming discover….

🥂Thrilled to showcase our #ICML2023 work on "Sequential Multi-Dimensional Self-Supervised Learning for Clinical Time Series" with Aniruddh Raghu! SMD-SSL is a pre-training objective for complex multimodal time series in healthcare. Come by poster 727 at 1:30 today to discuss!

RT @EricTopol: Upending the model of #AI adoption in healthcare:.Examples of how low and middle income countries are out in front .Our late….