SeqMaestro: From nucleotide sequences to biological hypotheses through interpretable machine learning
Evgeny S. Saveliev, Krzysztof Kacprzyk, Charlotte Capitanchik, Neelanjan Mukherjee, Kate Matlin, Ryan Sheridan, Srinivas Ramachandran, Jernej Ule, et al.
cs.LG · cs.AI
Quick verdict
SeqMaestro makes sequence ML interpretable via LLM-written features; four genomics tasks recover known biology; closed-LLM reliance limits reproducibility.
Summary
SeqMaestro is a framework that connects nucleotide sequences to interpretable machine learning through three feature representations: per-position indicators, pattern-based k-mers/motifs, and programmatic features written by a language model (proposed one-shot, iteratively with fit feedback, or jointly during tree growth). It fits batteries of transparent models (sparse logistic regression, additive models, decision trees, plus XGBoost as a black-box reference), reduces each model to ten features, and scores candidates with a three-tier reliability panel: resampling stability (100 half-samples), hold-out association (Mann-Whitney AUROC with 10^5 permutation calibration and multiple-testing corrections), and ablation necessity, alongside shuffled-label and planted-signal controls. On four tasks (883 to ~500,000 sequences) spanning Polycomb nucleation, RNA polymerase II pausing, eIF4E-dependent translation, and PTBP1 splicing, the system reproduces expected signals (CpG density, pause-associated G positions, pyrimidine asymmetry, GC3 codon usage) and proposes testable new hypotheses such as G-rich pause-site signatures. Gradient-boosted references never meaningfully surpassed interpretable models on any task. Main limitations: binary classification only, dependence on closed-weight LLM APIs, and validation restricted to four tasks without independent-dataset replication.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
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
- A two-layer interface (three feature representation families x three LLM-driven extraction strategies) lets intrinsically interpretable models operate directly on nucleotide sequences.
- Fitting diverse transparent pipelines and aggregating them via a three-tier reliability panel yields more robust and richer biological hypotheses than feature importance alone.
- On four genomics tasks the framework recovers established signals (CpG density, pause-site G positions, pyrimidine asymmetry, GC3) and proposes new testable hypotheses (e.g., G enrichment at -10/-2/-3/-7/-8 around pause sites) with hold-out support.
- Interpretable models in the battery lose little accuracy relative to gradient-boosted black-box references, which never meaningfully surpass them on any task studied.
Strengths
Well-designed reliability panel with strong built-in controls
Stage 3B combines Tier 1 stability (selection re-run on 100 half-samples drawn as 50 complementary pairs), Tier 2 hold-out association (Mann-Whitney AUROC calibrated against 10^5 reshufflings, reported under three multiplicity corrections plus a direct false-discovery-proportion estimate), and Tier 3 necessity (log-loss increase on hold-out when refit without a feature, Holm-corrected across ten features), with shuffled-label and planted-signal sanity checks; the report itself separates AI-written interpretation from traceable numbers.
Honest treatment of the black-box baseline
Gradient-boosted trees are included as reference and the text states they 'do not meaningfully surpass interpretable models on any task'; Appendix Table 4 openly reports cases where the black box is better (e.g., Pol II pausing 0.981 vs 0.960 best interpretable ten-feature model), demonstrating the accuracy cost of transparency is usually small.
External/orthogonal validation of discoveries
Leakage-aware evaluation discipline
Unusually transparent reporting of costs and provenance
Weaknesses
Reproducibility depends on closed-weight, versioned commercial LLMs
LLM prior-knowledge leakage is controlled in only one of four tasks
Selection pressure from many tries on one dataset
Potential design confound in the pausing task
Scope limited to binary classification, four curated tasks
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SeqMaestro (arXiv:2609.14882) is a method paper proposing a bridge between nucleotide sequences and intrinsically interpretable machine learning. The conceptual contribution is a two-layer interface: a taxonomy of feature representations (positional, pattern-based k-mers/motifs, and programmatic features defined as human-readable computations) and three extraction strategies for programmatic features (one-shot LLM proposal, iterative propose-fit-refine, and joint generation inside tree growth via Huynh et al. [13]). The system claim is end-to-end automation - data cleaning, battery fitting, hyperparameter discipline, reliability analysis, and a written report - through a no-code or agent-driven workflow. Validation is four biological case studies (two DNA, two RNA): Polycomb nucleation (~5,789 segments), RNA polymerase II pausing (499,672 x 101 nt sites), eIF4E dependence (~2,300-2,400 transcripts, three windows), and PTBP1 exon regulation (883 exons). Central claims are that transparent pipelines recover known biology (CpG density for nucleation, G at pause-proximal positions, pyrimidine asymmetry upstream of PTBP1-regulated exons, GC3 codon usage for eIF4E), expose richer logic than feature importance (thresholds, response curves, conditional rules), and lose little accuracy versus gradient-boosted references.
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 SeqMaestro?
SeqMaestro is a machine-learning framework that takes labelled nucleotide sequences (a binary classification task) and returns interpretable models plus a list of sequence properties that are robust enough to pursue as biological hypotheses, presented with a written report. Its distinctive element is a 'programmatic feature' layer: a language model writes short human-readable Python functions of the sequence (e.g., CpG density, distance between motifs), which are then fitted into transparent models such as sparse logistic regression, additive models (EBM), and decision trees.
Does SeqMaestro predict sequences better than deep learning?
No. It targets interpretable hypothesis generation, not raw predictive accuracy. The paper benchmarks gradient-boosted trees on the same feature columns as a black-box reference and states these 'do not meaningfully surpass interpretable models on any task'; deep sequence models (DeepSEA, Enformer) are discussed as related work rather than as accuracy competitors, and no head-to-head accuracy comparison with them is reported.
Which biological tasks did SeqMaestro demonstrate?
Four binary tasks spanning DNA and RNA: Polycomb nucleation in mouse ES cells (200 nt windows, ~5,789 sequences, hold-out AUROC around 0.86-0.87), RNA polymerase II pausing in human genes (499,672 sites, 101 nt windows, hold-out AUROC 0.92-0.96), eIF4E-dependent translation (three windows of ~2,200-2,400 transcripts, up to about 0.88), and PTBP1 exon regulation in K562 cells (883 exons, hold-out AUROC about 0.86-0.88).
How does SeqMaestro decide which discovered features are reliable?
A three-tier reliability panel: Tier 1 re-runs the whole feature-selection procedure on 100 half-samples to check the feature is selected again; Tier 2 tests whether the feature alone separates the classes on the untouched hold-out set (permutation-calibrated Mann-Whitney AUROC with multiple-testing corrections); Tier 3 removes each feature and measures the accuracy loss. Features that compute near-identical values are grouped into 'concepts' before testing. Shuffled-label and planted-signal controls verify the panel claims nothing under the null and recovers a known effect.
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.
- Large portions of the extracted text are omitted mid-document (part of the Polycomb results between Fig. 2 and Fig. 3, the end of the PTBP1 section including Fig. 5, part of the Limitations section, and appendices after Appendix K), so figure-level judgments rely on embedded captions and narration only.
- Figures are not analyzable as images; all numeric claims come from the extracted text and caption text.
- No explicit code or data availability statement appears in the visible text; codeAvailable is set to false based on absence, not on a stated refusal.
- Baseline counts and exact numbers are taken from Tables 1, 3, 4 and the Methods as rendered; minor OCR artifacts (e.g., reference numbering, figure-caption duplication) were tolerated.
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.