Intermediate Layers Encode Optimal Biological Representations in Single-Cell Foundation Models

TL;DR

This study shows that in single-cell foundation models, the most informative features depend on the task and cell state, with optimal layers varying rather than always being the final layer.

Contribution

It systematically evaluates layer-wise representations, revealing that optimal feature extraction layers are task- and context-dependent, challenging the common practice of using final-layer embeddings.

Findings

Optimal layers for trajectory inference peak at 60% depth.

Perturbation response optimal layers shift significantly across cell states.

First-layer embeddings outperform deeper layers in quiescent cells.

Abstract

Current single-cell foundation model benchmarks universally extract final layer embeddings, assuming these represent optimal feature spaces. We systematically evaluate layer-wise representations from scFoundation (100M parameters) and Tahoe-X1 (1.3B parameters) across trajectory inference and perturbation response prediction. Our analysis reveals that optimal layers are task-dependent (trajectory peaks at 60% depth, 31% above final layers) and context-dependent (perturbation optima shift 0-96% across T cell activation states). Notably, first-layer embeddings outperform all deeper layers in quiescent cells, challenging assumptions about hierarchical feature abstraction. These findings demonstrate that "where" to extract features matters as much as "what" the model learns, necessitating systematic layer evaluation tailored to biological task and cellular context rather than defaulting to final-layer embeddings.