Relational Foundation Models for Enterprise Data [Jure Leskovec] - 768

Why It Matters

Enterprises can unlock predictive power from existing databases instantly, cutting costs and accelerating decision‑making while achieving higher accuracy than legacy tabular models.

Key Takeaways

• Relational foundation model predicts across any database without training.
• Graph neural networks replace manual feature engineering for tabular data.
• Self‑supervised learning extracts biology from single‑cell RNA‑seq data.
• Multi‑table schemas become graphs, enabling end‑to‑end AI pipelines.
• Early results show double‑digit accuracy gains over traditional models.

Summary

The podcast introduces a new relational foundation model that can reason over structured relational data across any enterprise database without additional training.

By treating tables and foreign‑key links as a graph, the model applies graph neural networks, eliminating manual feature engineering and enabling end‑to‑end learning directly on raw relational inputs.

Examples include self‑supervised analysis of single‑cell RNA‑seq data to infer cellular states, and pandemic spread modeling using social‑media interaction graphs, both yielding double‑digit accuracy improvements over traditional pipelines.

For businesses, this means faster deployment of AI, reduced data‑engineering overhead, and more reliable predictions, potentially transforming how enterprises extract value from their most abundant asset—relational data.

Original Description

In this episode, Jure Leskovec, co-founder and chief scientist at Kumo and professor of computer science at Stanford, joins us to explore two fronts of his work: AI for science and relational deep learning. We begin with AI Virtual Cell, a multiscale effort to learn data-driven representations from proteins to cells to patients using single-cell RNA-seq data, protein language models like ESM, and structure models like AlphaFold—without hand-encoding biology. Jure then dives into relational deep learning, reframing enterprise databases as graphs and training neural networks directly on raw multi-table data. He explains Kumo’s Relational Foundation Model (RFM2), which performs in-context learning over subgraphs to make accurate predictions on new databases and tasks with no training, and how this approach benchmarks against RelBench and other multi-table datasets. We also discuss real-world deployments at companies like Reddit, DoorDash, and Coinbase, explainability via attention over tables and columns, integration with agentic systems, deployment options, and practical limitations.

🗒️ For the full list of resources for this episode, visit the show notes page: https://twimlai.com/go/768.

🔔 Subscribe to our channel for more great content just like this: https://youtube.com/twimlai?sub_confirmation=1

🗣️ CONNECT WITH US!

===============================

Subscribe to the TWIML AI Podcast: https://twimlai.com/podcast/twimlai/

Follow us on Twitter: https://twitter.com/twimlai

Follow us on LinkedIn: https://www.linkedin.com/company/twimlai/

Join our Slack Community: https://twimlai.com/community/

Subscribe to our newsletter: https://twimlai.com/newsletter/

Want to get in touch? Send us a message: https://twimlai.com/contact/

🔗 LINKS & RESOURCES

===============================

Kumo.AI - https://kumo.ai/

Relational Graph Transformer - https://arxiv.org/abs/2505.10960

Relational Deep Learning: Challenges, Foundations and Next-Generation Architectures - https://arxiv.org/abs/2506.16654

KumoRFM-2: Scaling Foundation Models for Relational Learning - https://arxiv.org/abs/2604.12596

Transformers for Tabular Data at Capital One with Bayan Bruss - 591 - https://twimlai.com/podcast/twimlai/transformers-for-tabular-data-at-capital-one

📸 Camera: https://amzn.to/3TQ3zsg

🎙️Microphone: https://amzn.to/3t5zXeV

🚦Lights: https://amzn.to/3TQlX49

🎛️ Audio Interface: https://amzn.to/3TVFAIq

🎚️ Stream Deck: https://amzn.to/3zzm7F5