Searcharxiv⌕ Search

arXiv · 2611.16881

Protein Language Models Fail To Capture Contact Predictions Under Invariant Relabeling

Abstract

Protein language models (pLMs) produce highly accurate contact maps from single-sequence inputs. People often take this accuracy as proof that these models implicitly learn invariant coevolutionary statistics, extracting the same signals as classical alignment-based methods like Potts. In this work, we introduce a diagnostic framework around a global bijective relabeling of the amino-acid alphabet. This transformation permutes residue letters while leaving all alignment-level information, column entropies, and pairwise edit distances unchanged. Pseudolikelihood Potts models and inverse-covariance estimators remain invariant under this transformation, but pLMs fail the test. Across multiple architectures (ESM-2, MSA Transformer, AMPLIFY) and model scales (8M to 3B parameters), pLM contact maps collapse under arbitrary relabeling. Using Hamming distance controls and biochemical grouping interventions, we attempt to isolate what likely causes this degradation and find that these models rely on explicit, canonical residue identities (``alphabet anchoring'') rather than pure coevolutionary statistics. Increasing scale makes models more responsive to biochemical classes, but their performance still degrades under relabeling. Our findings show a gap between how pLMs and classical coevolutionary models compute contacts. More broadly, our work shows how diagnostic perturbations that leave classical models provably unchanged can expose hidden representation biases and prevent over-reliance on neural black boxes in biological modeling.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Paarth Batra, Aryan Keluskar. 2026-11-12. Protein Language Models Fail To Capture Contact Predictions Under Invariant Relabeling. https://searcharxiv.com/abs/2611.16881

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Optimising Recruitment and Sample Allocation in Multi-Site Rare Disease Studies

Clinical research in rare diseases is challenged by the limited availability and broad geographical distribution of eligible patients. This necessitates multi-centre collaboration and strategic selection of analytical laboratories based on capacity, cost, and capability. We present a mathematical framework based on cardinality optimisation to support key operational decisions in study design, balancing clinical relevance, feasibility, budgetary constraints, and logistical factors. The framework was implemented in the Reconstruction and Computational Modelling for Inherited Metabolic Diseases (Recon4IMD) project, aiming to recruit 1,900 diagnosed patients across 24 healthcare centres and eight metabolomic laboratories. Compared to both proportional heuristic and greedy marginal allocation baselines, the proposed optimisation model consistently achieved higher recruitment yields in simulated multi-centre scenarios and reduced metabolomic analysis costs by approximately 30% under unconstrained conditions. Notably, both the heuristic and greedy allocation baselines became infeasible under progressively tighter laboratory capacity constraints, whereas the optimised model maintained feasibility and cost control. These findings highlight the value of integrated optimisation strategies for improving recruitment, platform selection, and sample allocation in complex, resource-limited research settings.

q-bio.QM↗

Analysis choices can produce apparent rich-club differences: an Alzheimer's disease EEG case study

Background: The weighted rich-club coefficient tests whether a network's most strongly connected nodes are preferentially interconnected. Its interpretation depends on two often-implicit choices: how the randomised comparison networks are constructed, and whether the hubs are re-identified in each comparison network or held fixed at those of the observed network. When hubs are defined by connection strength, these choices change what the coefficient tests. New method: We examined how these choices affect inference using resting-state EEG from 20 people with Alzheimer's disease and 20 controls. Holding the data fixed, we compared four analysis pipelines, crossing two ways of building the comparison networks (global weight reshuffle or strength-conditioned null) with the two hub choices, and ran a calibration study using data constructed to contain no preferential hub coupling beyond nodal strength. Results: An apparent eyes-open group difference in rich-club organisation appeared only under the pipeline combining a global weight reshuffle with re-identified hubs. Under the empirical resampling ensemble, that pipeline reported a group difference in 43% of pseudo-studies, always higher in the Alzheimer's disease group, while the other three stayed near the expected 5%. Comparison with existing methods: Unlike global weight reshuffling, strength-conditioned nulls approximately preserve nodal strengths; fixing observed hubs also retains the identity of the core being tested. Conclusions: The same data can support or dissolve an apparent rich-club difference depending on these choices. No single choice is universally correct, but reporting which were made, and why they fit the question being asked, is essential to interpreting the result.

q-bio.QM↗

Quantum Cognition Machine Learning for Forecasting Chromosomal Instability

The accurate prediction of chromosomal instability from the morphology of circulating tumor cells (CTCs) enables real-time detection of CTCs with high metastatic potential in the context of liquid biopsy diagnostics. However, it presents a significant challenge due to the high dimensionality and complexity of single-cell digital pathology data. Here, we introduce the application of Quantum Cognition Machine Learning (QCML), a quantum-inspired computational framework, to estimate morphology-predicted chromosomal instability in CTCs from patients with metastatic breast cancer. QCML leverages quantum mechanical principles to represent data as state vectors in a Hilbert space, enabling context-aware feature modeling, dimensionality reduction, and enhanced generalization without requiring curated feature selection. QCML outperforms conventional machine learning methods when tested on out of sample verification CTCs, achieving higher accuracy in identifying predicted large-scale state transitions (pLST) status from CTC-derived morphology features. These preliminary findings support the application of QCML as a novel machine learning tool with superior performance in high-dimensional, low-sample-size biomedical contexts. QCML enables the simulation of cognition-like learning for the identification of biologically meaningful prediction of chromosomal instability from CTC morphology, offering a novel tool for CTC classification in liquid biopsy.

q-bio.QM↗