Descriptions:
In this AI Engineer conference talk, Akram Baharlouei, a machine learning engineer at Altos Labs, explores the engineering challenges of applying transformer-based foundation models to single-cell biology. Altos Labs is a biotech startup focused on cellular rejuvenation — reversing aging at the cellular level — and Baharlouei frames the problem through the lens of Shinya Yamanaka’s Nobel Prize-winning discovery of transcription factors that can reprogram aged skin cells back to an embryonic-stem-cell-like state.
The talk covers why single-cell RNA sequencing (RNA-seq) has become the dominant modality for foundation model training, with datasets now scaling to 500 million and even one billion cells. Baharlouei is candid about the data quality challenges: RNA-seq measurements are highly noisy and heterogeneous across labs and machines, represent only snapshots of a dynamic biological process, and exhibit bursting behavior that confounds standard modeling assumptions. He surveys current state-of-the-art single-cell foundation models and discusses the gap in other modalities — proteomics, morphology, spatial transcriptomics — that need to advance before truly holistic virtual cell models become practical.
For ML engineers curious about applying large model architectures to biological domains, this talk provides an accessible but technically grounded entry point. It situates the work within larger initiatives like the Human Cell Atlas and the broader goal of building virtual tissues and organs to compress the notoriously expensive and failure-prone drug development pipeline.
📺 Source: AI Engineer · Published July 19, 2026
🏷️ Format: Deep Dive
