Explainability of Protein Deep Learning Models
Bibliographic record
Abstract
Protein embeddings are the new main source of information about proteins, producing state-of-the-art solutions to many problems, including protein interaction prediction, a fundamental issue in proteomics. Understanding the embeddings and what causes the interactions is very important, as these models lack transparency due to their black-box nature. In the first study of its kind, we investigate the inner workings of these models using XAI (explainable AI) approaches. We perform extensive testing (3.3 TB of total data) involving nine of the best-known XAI methods on two problems: (i) the prediction of protein interaction sites using the current top method, Seq-InSite, and (ii) the production of protein embedding vectors using three methods, ProtBERT, ProtT5, and Ankh. The results are evaluated in terms of their ability to correlate with six basic amino acid properties-aromaticity, acidity/basicity, hydrophobicity, molecular mass, van der Waals volume, and dipole moment-as well as the propensity for interaction with other proteins, the impact of distant residues, and the infidelity scores of the XAI methods. The results are unexpected. Some XAI methods are much better than others at discovering essential information. Simple methods can be as good as advanced ones. Different protein embedding vectors can capture distinct properties, indicating significant room for improvement in embedding quality.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.011 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".