Probing isn't Causality in Molecular Models
- 1Shiv Nadar University
Abstract
Linear probes establish whether a chemical variable can be recovered from a molecular representation; they do not establish how that variable contributes to subsequent predictions. We replace a concept-associated scalar component at an intermediate block and measure the change through a separately fitted downstream outcome readout. Comparisons with random directions and architecture-matched random encoders distinguish recoverability, intervention selectivity, and a pretraining-associated selectivity advantage. In Uni-Mol, block-12 total-distortion edits predict the observed energy-change sign for 97.6\% of held-out pairs, with Spearman ρ=0.787. Structure-clustered intervals for the primary direction and encoder contrasts exclude zero. A stricter 20-direction sensitivity analysis does not survive correction across six cells. A post-review analysis finds substantial concept--outcome alignment; intervention effects persist after removal of the same-block outcome direction, but this does not establish that subsequent encoder computation is necessary. In MolCLR, pretraining improves dipole recovery by 0.080 R2, whereas concept-minus-random sign-accuracy contrasts are $0.029$ and −0.015 at the two tested blocks. Their paired uncertainty is unresolved, so they do not establish an absence of selective effects. We propose an operational evidence ladder whose conclusions are conditional on the concept, intervention, control family, and fitted readout.
Related
- CPD-YOLOv8: A Causality-Driven Physical Modality Fusion Framework for Early Micro-Fire Detection to Protect Complex Wildland Vegetation
- Time versus causal inferences: Action-outcome learning and prediction differentially shape the experience of causality and simultaneity
- Improving Brain-Behaviour Causality with Neurofeedback
- From Translation to Causality: A Two-Axis Map of RNA-of-Origin and Output Attribution in Translated Non-Coding RNAs During Skeletal Muscle Myogenesis and Regeneration
- Cortical thinning and altered directed connectivity in Parkinson’s disease: a cortical thickness–guided voxel-wise granger causality study
BibTeX
@misc{srivastava2026probing,
title = {{Probing isn't Causality in Molecular Models}},
author = {S. P. Srivastava and S. C. Krishna and H. Singh and R. K. SINGH},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.05.756781},
url = {https://www.biorxiv.org/content/10.64898/2026.10.05.756781v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Probing isn't Causality in Molecular Models
AU - S. P. Srivastava
AU - S. C. Krishna
AU - H. Singh
AU - R. K. SINGH
PY - 2026
DA - 2026/10/05
PB - bioRxiv
DO - 10.64898/2026.10.05.756781
UR - https://www.biorxiv.org/content/10.64898/2026.10.05.756781v1
AB - Linear probes establish whether a chemical variable can be recovered from a molecular representation; they do not establish how that variable contributes to subsequent predictions. We replace a concept-associated scalar component at an intermediate block and measure the change through a separately fitted downstream outcome readout. Comparisons with random directions and architecture-matched random encoders distinguish recoverability, intervention selectivity, and a pretraining-associated selectivity advantage. In Uni-Mol, block-12 total-distortion edits predict the observed energy-change sign for 97.6\% of held-out pairs, with Spearman ρ=0.787. Structure-clustered intervals for the primary direction and encoder contrasts exclude zero. A stricter 20-direction sensitivity analysis does not survive correction across six cells. A post-review analysis finds substantial concept--outcome alignment; intervention effects persist after removal of the same-block outcome direction, but this does not establish that subsequent encoder computation is necessary. In MolCLR, pretraining improves dipole recovery by 0.080 R^2, whereas concept-minus-random sign-accuracy contrasts are $0.029$ and −0.015 at the two tested blocks. Their paired uncertainty is unresolved, so they do not establish an absence of selective effects. We propose an operational evidence ladder whose conclusions are conditional on the concept, intervention, control family, and fitted readout.
ER -