Mutual information estimates in protein models show high variability and limits
When Can You Trust Neural Critics? Auditing Mutual Information Estimates of Higher-Order Epistasis in Protein Language Models
Machine Learning
Summary
Mutual information (MI) measures how much two things are related, but estimating MI from neural networks can be tricky and unreliable, especially for complex interactions. The authors check past MI estimates done on protein models and find that results vary a lot depending on the specifics of the calculation. They show that even with improved methods, the estimates often miss important signals unless given extra information. This means claims about detailed protein interactions based on these estimates should be viewed cautiously for now.
What this means in practice
- •For protein engineers: Avoid relying solely on blind mutual information estimates to analyze complex protein sequence interactions in models like ESM-2, reducing risk of overinterpreting noisy signals.
- •For machine learning engineers: Use the provided diagnostic protocol to audit and improve stability of mutual information estimates when evaluating neural network representations in high-dimensional settings.
Authors
Maryam Rahimimovassagh, Waldemar Karwowski, Niloofar Yousefi
Abstract
Mutual information (MI) estimation is used to probe what neural representations encode, including in protein language models, but variational MI estimators have variance that scales at least exponentially in the true MI. We audit a prior neural-MI study of order-resolved epistasis in ESM-2 on the four-site GB1 landscape. Run-to-run differences show no systematic bias and a magnitude comparable to ordinary cross-seed retraining variability, and the evaluation-pairing scheme shifts scores substantially. At the original dimensionality, two of three alternative estimator families stabilize the estimate across seeds and one does not. In a synthetic calibration at the representation dimensionalities of two other protein language models (a Gaussian channel with known MI; no model representations), all three families recover materially less signal than required, despite correct signal ordering and stable seeds. A control critic given the true signal direction, which also reduces its input to one dimension, recovers 93.3% and 98.7% of the analytic MI at gain 1.5 (at 5,120 and 4,096 dimensions), so the signal is recoverable with privileged information, but the control does not identify why blind estimation fails. We package the checks into a reusable diagnostic protocol and conclude that even a stabilized estimator is not currently sufficient for order-resolved dependence claims made under blind search at this dimensionality and sample size.