Probing regulatory logic
SAE features are annotated with biological programs, and steering them moves gene expression in sensible directions. But does AIDO.Cell encode the causal direction of gene regulation — does perturbing a transcription factor propagate to its targets more strongly than the reverse?
If yes, we have a real regulatory model inside the scFM. If no, it’s encoding co-expression memory (which genes go together) without a notion of who drives whom. That’s a big distinction for any downstream perturbation- prediction story, so we tested it directly.
Test setup
- TF: TBX21 (master Th1 regulator).
- Feature: f3092 — PR ≈ 11 on CD8 T cells, top genes include CXCR3, IFNG, CCL5, TBX21, TNF — a coherent Th1 module.
- Cell type: CD8 T cells (n = 316).
- Ground truth: CollecTRI signed regulon — 53 literature-curated TBX21 targets.
- Control targets (in-feature, in-regulon): CXCR3, IFNG, TNF.
- Random pool (in-feature, not in-regulon): CCL5, KPNA2, OASL.
For each cell, alpha, and target gene \(T_i\):
- Forward effect \(F_i\) = change in \(\text{logit}(T_i)\) when we steer f3092 at the TBX21 token.
- Reverse effect \(R_i\) = change in \(\text{logit}(\text{TBX21})\) when we steer f3092 at the \(T_i\) token.
If TBX21 causally drives its targets, \(|F_i| \gg |R_i|\).
Result: no directional asymmetry
Across all three targets and all tested alphas (±2, ±5, ±15), paired permutation tests on \(|F_i| - |R_i|\) fail almost everywhere. The forward and reverse effects are roughly symmetric — perturbing TBX21 moves its targets about the same amount as perturbing a target moves TBX21.
Interpretation
Under this probe, AIDO.Cell’s representation of the TBX21-Th1 module looks like a bidirectional co-expression cluster, not a directed TF-to-target graph. Two caveats before drawing grand conclusions:
- This is one TF with ~3 targets in one feature. The negative is for this system, not necessarily all of regulatory biology.
- The steering is at the feature level, not a direct perturbation of the TF’s token. A more direct test would intervene on TBX21’s own gene embedding and measure propagation.
That said: if the scFM did encode directionality, we’d expect to see a signal on the Th1 module with CollecTRI ground truth. We don’t.
Why this matters
It suggests that “steering a feature” works as a transcriptional-state operator, not as a causal intervention on a regulatory network. For interpretable perturbation prediction this is good news (state-level steering is tractable and interpretable) and bad news (we shouldn’t claim the scFM has learned causal regulation from expression data alone). The field has been suspicious of this for a while; this is a small, direct data point.
Further reading
- Full experimental design, statistics, and per-target figures:
reports/regulatory_logic_experiment.md.