MI-PEFT: Mixture-of-Experts Integrated Parameter-Efficient Fine-Tuning Protein Language Models Improves Acidophilic Proteins Classification
q-bio.QM · cs.LG
Quick verdict
Proposes PEFT+DeepSeekMoE on ESM C-600M for acidophilic protein ID; promising, but figure-only results and unequal tuning leave SOTA claims under-supported.
Summary
This method paper introduces MI-PEFT, a framework for classifying acidophilic versus non-acidophilic proteins that pairs a frozen ESM C-600M backbone with one of six LoRA-style parameter-efficient fine-tuning adapters (LoRA, MiSS, C3A, OFT, RandLoRA, SHiRA) and replaces the dense classifier with a sparse DeepSeekMoE head (three routed experts, top-2 routing, plus a shared expert and load-balancing loss), followed by per-fold temperature scaling. Using a curated benchmark of 4,089 acidophilic and 1,654 non-acidophilic NCBI proteins (20% sequence-identity cap between train and test; 80/20 split), the author reports roughly 0.91–0.92 accuracy, 0.94–0.95 sensitivity, 0.84–0.85 specificity, 0.76–0.79 MCC and 0.95–0.96 AUC on the held-out test set, with trainable parameters kept below 1%. C3A is selected as best, chiefly on specificity and MCC, and outperforms classical ML, DNN/LSTM baselines and a reimplemented ACE model under UMAP and confidence-density comparisons. Main limitations: results appear only as figures with approximate ranges, without fold variance or significance tests; the ACE comparison changes the original frozen-embedding protocol; hyperparameters differ across adapters (C3A uses a tenfold higher learning rate); the recent AcidProNet method is not compared. Code and data are released on GitHub.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- All six LoRA-based PEFT variants combined with a DeepSeekMoE head on ESM C-600M outperform classical ML, deep-learning baselines and the previous SOTA (ACE) on acidophilic protein classification (§3.3–3.4).
- C3A is the best of the six adapters, distinguished mainly by specificity and MCC under class imbalance (§3.3).
- Replacing the dense head (≈1.06×10⁷ parameters) with a DeepSeekMoE head (1,221,896 parameters) improves all six metrics over MLP and Vanilla MoE heads at matched parameter scale (§3.5, Table 4).
- Constraining adaptation to low-rank updates preserves pretrained representations and improves negative-class discrimination relative to broader fine-tuning (§3.4, Fig 5A).
- Per-fold temperature scaling yields flatter, better-calibrated confidence distributions useful for prioritizing experimental validation (§2.5, Fig 5B).
Strengths
Systematic, controlled PEFT benchmark
Six LoRA-family adapters (LoRA, MiSS, C3A, OFT, RandLoRA, SHiRA) are evaluated under one protocol (same splits, same MoE head, same calibration); Table 1 documents forward mappings and initialization, Table 3 lists key hyperparameters, §2.7 fixes optimizer, schedule, batch and early-stopping.
Concrete efficiency accounting
Table 2 reports per-method trainable parameters (2,507,528–5,203,208; 0.43–0.90%), epoch times (76.1–102.7s) and allocated memory (25.1–27.8GB); Table 4 parameter-matches the head ablation (MLP 1,181,186 vs DeepSeekMoE 1,221,896 vs Vanilla MoE 1,363,434) so the comparison is not confounded by head size.
Class-imbalance-aware design and evaluation
Leakage control and calibration discipline
Backbone scaling insight
Weaknesses
No statistical support for comparative claims
SOTA comparison is not apples-to-apples and is incomplete
Opaque dataset curation
Hyperparameter asymmetry across adapters
Internal inconsistencies and editorial sloppiness
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This single-author method paper proposes MI-PEFT, a pipeline for classifying acidophilic versus non-acidophilic proteins: a frozen ESM C-600M backbone adapted by one of six LoRA-family PEFT methods (LoRA, MiSS, C3A, OFT, RandLoRA, SHiRA; §2.3, Table 1), a dense classification head replaced by a sparse DeepSeekMoE head (§2.4), and per-fold post-hoc temperature scaling (§2.5). Four claims carry the paper: (1) all six adapters outperform classical ML, deep-learning baselines and the previous SOTA, the ACE model of Susanty et al. (2024); (2) C3A is the best adapter, mainly on specificity and MCC under class imbalance (§3.3); (3) the DeepSeekMoE head shrinks classifier parameters from ≈1.06×10⁷ (a 2304×4608 dense mapping, §2.4) to 1,221,896 (Table 4) while beating MLP and Vanilla MoE heads at matched scale (§3.5, Fig 4A); (4) constraining adaptation to low-rank updates preserves pretrained representations and improves negative-class discrimination versus broader fine-tuning (§3.4, Fig 5A).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MI-PEFT?
MI-PEFT is a framework for identifying acidophilic proteins from amino-acid sequence. It adapts an ESM C-600M protein language model using parameter-efficient adapters (six LoRA-style methods were benchmarked), replaces the usual dense classifier with a sparse mixture-of-experts head (DeepSeekMoE with routed plus shared experts and a load-balancing loss), and applies temperature scaling to calibrate prediction confidences. The best configuration found, using the C3A adapter, is released with code and data on GitHub.
How accurate is MI-PEFT at identifying acidophilic proteins?
On the paper's independent test set (818 acidophilic and 331 non-acidophilic proteins), the author reports approximately 0.91–0.92 accuracy, 0.94–0.95 sensitivity, 0.84–0.85 specificity, 0.76–0.79 Matthews correlation coefficient, and 0.95–0.96 ROC-AUC. These values are given as figure-based approximations rather than exact tables, and specificity is the weakest metric because acidophilic proteins outnumber negatives roughly 2.5 to 1.
Is the MI-PEFT code and dataset publicly available?
Yes. The paper states that the complete dataset and framework are available at github.com/Duncan-SHH/MI-PEFT-for-Acidophilic-Protein-Classification. The dataset comprises 4,089 acidophilic and 1,654 non-acidophilic proteins curated from NCBI, split 80/20 into training and independent test sets with a 20% sequence-identity cap between them.
Why is class imbalance a problem for acidophilic protein classification?
In this benchmark, acidophilic proteins outnumber non-acidophilic ones by roughly 2.5 to 1 (4,089 vs 1,654). Models optimized naively for accuracy tend to over-predict the majority (positive) class, producing many false positives and depressing specificity. The paper responds with specificity/MCC-centered evaluation, an auxiliary load-balancing loss in its mixture-of-experts head, and post-hoc temperature scaling to improve confidence calibration.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All quantitative results exist only inside figures (Fig 2–5); the PDF extraction carries text only, so exact metric values could not be cross-checked and the review relies on the approximate ranges quoted in the prose.
- Table 3's hyperparameter columns were scrambled by extraction (values detached from method names), so per-method rank/α assignments were deliberately not asserted in this review; the learning-rate asymmetry is cited from §2.7 prose instead.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.