Relational Foundation Models for Enterprise Data with Jure Leskovec - #768

In this episode Jure Leskovec unveils a relational foundation model that can make accurate predictions on any enterprise database without any task‑specific training. By treating tables as nodes and foreign‑key links as edges, the model applies graph neural networks directly to raw relational data, eliminating the costly feature‑engineering pipelines that have dominated tabular machine learning for decades. This shift mirrors the breakthroughs seen in computer vision and natural language processing, where models now learn straight from raw pixels or tokens, and promises enterprises faster deployment and higher predictive performance across diverse use cases.

Leskovec also dives into his AI for science agenda, describing the AI Virtual Cell project that builds multi‑scale representations from proteins up to whole patients. Using single‑cell RNA‑seq—a 20,000‑dimensional snapshot of each cell’s protein abundance—combined with protein language models like ESM and structural tools such as AlphaFold, the team creates a digital twin from a single drop of blood. The approach is fully self‑supervised, letting biological hierarchies emerge from data rather than being hand‑coded, accelerating drug discovery, disease trajectory mapping, and personalized therapy design.

Practical applications of relational deep learning are already evident. Fraud detection, anti‑money‑laundering, churn prediction, and recommender systems all benefit from the model’s ability to ingest heterogeneous, multi‑table schemas and uncover hidden patterns that traditional XGBoost or linear models miss. By learning directly on the graph of entities—customers, products, transactions—these systems achieve double‑digit accuracy gains while reducing engineering overhead. As enterprises grapple with ever‑growing relational datasets, Leskovec’s work signals a new era where AI can unlock value from the full relational fabric without the bottlenecks of manual preprocessing.