Mechanistic Interpretability of Biological Foundation Models: An Empirical Analysis of scGPT Gene-Embedding Geometry

Technical Summary: Mechanistic Interpretability of Biological Foundation Models

Problem Statement
While biological foundation models (such as scGPT) are widely claimed to learn structured biological knowledge, interpretability research often relies on qualitative examples or downstream task performance scores rather than direct analysis of internal representations. A critical gap exists in understanding whether these models encode specific, known biological structures—such as physical protein-protein interactions—within their static internal parameters before any cell-specific contextualization occurs. This paper addresses a narrow empirical question: Does the learned gene-token embedding space of the scGPT model encode known physical interaction structures among human gene products?

Methodology
The study performs an empirical analysis of the public scGPT human checkpoint, specifically isolating the static input gene-token embedding matrix (60,694 genes, 512 dimensions) prior to any transformer-layer computations or cell-specific expression values.

1. Data Sources:
   • Model: scGPT human checkpoint (Cui et al., 2024a).
   • Interaction Ground Truth: Two independent databases were used:
     • STRING v12.0: Human physical links mapped to gene symbols, resulting in 726,387 unique pairs across 18,441 genes.
     • BioGRID 5.0.259: Human physical interactions, resulting in 89,187 unique pairs across 11,250 genes.
2. Experimental Design:
   • Cosine Similarity Analysis: The authors computed pairwise cosine similarities between embedding vectors of known physical interaction partners versus non-interacting pairs.
   • Negative Sampling: To control for "hub" bias (where highly connected genes might artificially inflate similarity), negative pairs were sampled from the same gene universe with endpoint-frequency matching to the positive set.
   • Baselines: Gene-label permutation baselines were established by randomly shuffling embedding rows to ensure the signal was not an artifact of the sampling procedure or global geometry.
   • Stratification: STRING interactions were analyzed across confidence score bins (≥400, ≥700, ≥900) to test for a dose-response relationship.
   • Nearest-Neighbor Enrichment: The study calculated Precision@k (k=10, 50, 100) for known interaction partners among the nearest neighbors in the embedding space, comparing observed precision against random expectation based on gene degree.

Key Results
The analysis reveals a consistent, though moderate, signal indicating that scGPT's static embeddings contain physical interaction information:

• Edge Recovery (AUROC): Known physical interaction partners had significantly higher cosine similarity than matched nonedges.
  • STRING: AUROC increased with confidence, ranging from 0.622 (score ≥400) to 0.704 (score ≥900).
  • BioGRID: AUROC was 0.589.
  • Baselines: Permutation baselines consistently returned to ~0.50, confirming the signal depends on gene identity within the learned space.
• Monotonicity: There was a monotonic increase in mean embedding cosine as STRING confidence scores rose (from 0.011 for nonedges to 0.111 for score 900–1000 edges), suggesting the geometry tracks interaction confidence rather than just binary edge status.
• Nearest-Neighbor Enrichment: The local neighborhood structure is enriched for physical partners.
  • For STRING score ≥700, Precision@10 showed a 52.9-fold enrichment over random expectation.
  • For BioGRID, Precision@10 showed a 17.3-fold enrichment.
  • Despite high enrichment ratios, absolute precision remains low (e.g., <5% of top-10 neighbors are recorded physical partners for STRING ≥700).

Key Contributions

• Direct Empirical Evidence: The paper provides quantitative evidence that the static input embedding layer of a single-cell foundation model (scGPT) already encodes detectable physical interaction structure, independent of cell-contextualized activations.
• Convergent Validation: The signal is validated across two independent interaction resources (STRING and BioGRID), survives rigorous endpoint-frequency matching, and demonstrates a dose-response relationship with interaction confidence.
• Refined Interpretability Scope: The study delineates the boundary between "biological organization" and "causal mechanism," showing that while the model organizes genes by physical interaction, this does not equate to learning causal regulatory logic.

Significance and Claims
The authors frame their conclusions with deliberate modesty. The paper supports a limited mechanistic claim: the scGPT input gene-token geometry is biologically organized before contextual processing begins. This suggests that downstream transformer layers do not start from arbitrary gene identifiers but from vectors already weakly structured by molecular interaction data.

However, the authors explicitly state what the results do not prove:

• The signal is insufficient for standalone physical interaction prediction (AUROCs of 0.59–0.70 are not strong predictors).
• The presence of physical interaction signals in static embeddings does not imply the model has learned causal regulatory mechanisms or directionality.
• The study does not claim that scGPT uses these embeddings causally for perturbation prediction.

The significance lies in establishing a baseline for mechanistic interpretability in biological foundation models: the "lookup table" itself contains biological structure. Future work, the authors suggest, should focus on causal interventions and feature-level circuit analysis in deeper activations to understand how this input signal is utilized.