PaperlayerView exact revision on arXiv

Source-linked AI summary

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, David L. Bentley, Mihaela van der Schaar

arXiv:2609.14882v1cs.LGcs.AI

TL;DR

Sequence models can predict complex biological outcomes, but black-box logic and unreliable post-hoc explanations leave hypotheses difficult to inspect, while transparent models are not designed for raw sequences. SeqMaestro connects sequences to interpretable models through varied feature representations and extraction strategies; across four biological questions, results indicate that no single representation is sufficient across tasks.

  • Problem

    Sequence models can predict complex biological outcomes, but black-box logic and unreliable post-hoc explanations leave hypotheses difficult to inspect, while transparent models are not designed for raw sequences.

    Source: Introduction · p. 2

  • Method

    SeqMaestro connects raw nucleotide sequences to interpretable models through positional, pattern-based, and programmatic features and complementary feature-generation strategies.

    Source: Introduction · p. 2

  • Results

    Across four biological questions, no single representation was uniformly preferable: positional features captured most signal in some tasks, while programmatic features retained substantially more information in others.

    Source: Discussion · p. 14

  • Takeaways & Limitations

    SeqMaestro produces compact, computable sequence hypotheses that can be inspected, challenged, and tested.

    Source: Discussion · p. 15

    SeqMaestro generates predictive hypotheses rather than establishing causality; associated sequence properties may be causal, proxies, or reflect dataset-construction confounding.

    Source: Discussion · p. 14

Abstract

from arXiv · show

Nucleotide sequence analysis is central to problems spanning regulatory genomics, evolutionary biology, and phenotype prediction. Classical bioinformatics methods extract interpretable sequence properties such as motifs and k-mer composition, but their flexibility is limited. In contrast, modern deep learning models can learn powerful predictive representations directly from raw sequences, yet their internal representations and decision mechanisms are difficult to inspect. Interpretable machine learning methods (e.g., sparse linear models and decision trees) provide human-understandable representations of predictive relationships but are not designed to operate directly on nucleotide sequences. Here, we introduce SeqMaestro, a machine learning framework that proposes biological hypotheses from nucleotide sequences using interpretable models. Our solution is centered around a two-layer interface that connects nucleotide sequences with the broader ecosystem of interpretable machine learning. SeqMaestro uses this interface to fit diverse combinations of interpretable models, feature representations, and extraction strategies, leveraging variability across transparent models to identify robust biological signals and richer predictive relationships than feature importance alone can provide. The system also supports data transformation and cleaning, model fitting, hyperparameter tuning, reliability analysis, and synthesis of results into a contextualized written report. By providing these capabilities through a no-code workflow, SeqMaestro is designed to make interpretable sequence analysis accessible to researchers without requiring extensive programming or machine learning expertise. SeqMaestro thereby provides an accessible route from nucleotide sequences to biological hypotheses.

Introduction

Sequence analysis has expanded beyond fixed feature vocabularies, but deep models and transparent models each leave a gap between prediction and inspectable sequence hypotheses. SeqMaestro connects sequences to interpretable models through expressive feature representations and varied extraction strategies.

Introduction

  • Classical methods capture meaningful sequence structure but require researchers to specify the expected signal form, potentially overlooking properties outside that vocabulary.Examples include homology search, motif discovery, hidden Markov models, k-mer representations, and gkm-SVM.Source: Introduction · p. 1
  • Deep models predict complex sequence-related outcomes, but their logic is difficult to inspect and post-hoc explanations can be unreliable, motivating models that express hypotheses directly.The paper notes that explanations may not faithfully reflect learned information, can be unstable under small perturbations, and may disagree.Source: Introduction · p. 2
  • SeqMaestro maps raw sequences into human-readable hypotheses using positional, pattern-based, and programmatic representations, with programmatic features expressing richer sequence structure.Programmatic computations can encode composition, positional relationships, motif organization, and higher-order logic while remaining inspectable.Source: Introduction · p. 2
  • Its second interface layer offers one-shot, iterative propose-fit-refine, and joint feature-generation strategies that couple hypothesis generation to empirical feedback to different degrees.These strategies can be combined with different transparent models and feature types.Source: Introduction · p. 2
  • SeqMaestro searches across transparent model classes because linear models, GAMs, and trees expose different relationships, enabling aggregation and robustness assessment across pipelines.The framework treats model class as part of the hypothesis search rather than limiting interpretation to individual feature importance.Source: Introduction · p. 2

1 Dataset preparation

The passages describe how transparent models can express richer sequence hypotheses and how SeqMaestro packages analysis steps into an end-to-end workflow.

1 Dataset preparation

  • Decision trees can reveal when a sequence property is informative only under another condition, while GAMs show how predictions change continuously across feature values.The model itself can therefore contribute to a biological hypothesis beyond an isolated predictive feature.Source: §1 · pp. 3–4
  • SeqMaestro's workflow includes data transformation and cleaning, model fitting, hyperparameter tuning, reliability analysis, and synthesis into a contextualized report.The no-code workflow is designed for researchers without extensive programming or machine-learning expertise.Source: §1 · p. 4

Results

The results evaluate SeqMaestro’s interpretable models across four DNA and RNA tasks, showing how sequence-feature representations can recover known sequence signals and yield task-specific hypotheses from sequences and labels.

Dataset preparation

Model fitting - SeqMaestro interface

  • The interface supports positional, pattern-based, and programmatic features, with language-model-driven one-shot, iterative, or joint feature generation for the large programmatic search space.Programmatic features compose sequence properties into human-readable hypotheses; the joint strategy is currently available for decision trees through an existing method [13].Source: Model fitting - SeqMaestro interface · pp. 4–6

Analysis and synthesis

  • For analysis and synthesis, SeqMaestro reduces models to ten features, refits them, assesses feature robustness in three tiers, and produces a written report.The tiers test feature stability under random subsampling, feature–target association, and whether removing a feature reduces model accuracy.Source: Analysis and synthesis · p. 6

Experimental Findings

  • Across four DNA and RNA questions, gradient-boosted trees did not meaningfully outperform interpretable models; AUROC was the primary performance metric.The case studies were chosen to span different informative sequence signals, and the analyses reported interpretable models and convergent concepts with supporting evidence.Source: Experimental Findings · p. 6

Polycomb nucleation

  • Six of nine Polycomb configurations selected reliable CpG-density measures, which separated hold-out segments at 0.84 AUROC; a tree rule identified cluster 7 with 97% of training segments.The rule required more than seven CpGs in 200 bases and over 49% G+C in the central 60 bases. Programmatic-feature scores held after reduction, while positional-only models fell from about 0.81 to about 0.64, supporting a window-wide signal and a threshold-like CpG-density hypothesis.Source: Figure 2 · p. 7, Polycomb nucleation · p. 8

eIF4E dependence

  • GC3 content separated hold-out eIF4E-resistant and sensitive transcripts at AUROC 0.82, and resistant transcripts showed higher baseline eIF4E binding than sensitive transcripts and the transcriptome overall.The analysis also recovered elevated G and C content immediately upstream of the start codon as an expected discriminator, while identifying coding-region codon usage as the strongest predictor of eIF4E sensitivity.Source: eIF4E dependence · p. 10, Figure 4 · p. 11

PTBP1 exon regulation

  • Programmatic features reached AUROC 0.875 on PTBP1 exon regulation, versus 0.656 for the best positional and pattern-based model.The signal was regional pyrimidine content weighted toward the exon; three reliable features independently separated hold-out exons at AUROC 0.81–0.85. The analysis also identified downstream G-richness as a possible predictor of PTBP1-enhanced exons.Source: Figure 5 · p. 12, PTBP1 exon regulation · p. 13
  • Across four tasks, SeqMaestro produced interpretable models with at most ten features and property lists supported by independent configurations and sequences.The properties included known biology recovered without human prompting, as well as new leads for further experiments.Source: PTBP1 exon regulation · p. 13
  • The analyses recovered known biology without human prompting, using positional features for pausing and LLM-driven feature generators for eIF4E that were not given the factor’s name.Source: PTBP1 exon regulation · p. 13
  • The four tasks yielded leads involving a nucleation threshold, an E. coli-like pausing translocation mechanism, codon-optimality coupling to eIF4E recruitment, and two candidate PTBP1 co-regulators binding the downstream intron.These leads were produced from sequence and labels alone and offer starting points for computational or bench experiments.Source: PTBP1 exon regulation · p. 13

Discussion

SeqMaestro searches across interpretable representations, feature-generation strategies, and model classes to produce compact predictive models and robust, inspectable sequence hypotheses. The discussion emphasizes that representation and model choice shape what can be learned, while the framework’s conclusions remain predictive rather than causal.

Discussion

  • Across four biological questions, SeqMaestro produced compact predictive models and sequence properties that could be inspected directly and evaluated for robustness.The resulting hypotheses include both the sequence property measured and the predictive relationship in which it participates.Source: Discussion · p. 14
  • Positional features suit location-specific signals, pattern-based features capture recurring local elements, and programmatic features combine regions, counts, composition, or other sequence properties.No representation was uniformly preferable: simple positional features sufficed in some tasks, while programmatic features retained more predictive information in others after reduction to few features.Source: Discussion · p. 14
  • Interpretable model families expose additive effects, response shapes, and conditional rules, making thresholds and interactions part of the hypothesis beyond feature importance.A ranked feature list can identify a predictive signal while leaving its predictive logic unspecified.Source: Discussion · p. 14
  • Language models may propose or refine programmatic features, but observed data and fixed statistical procedures determine retention, prediction, and reliability.The framework records information exposed to feature generators, allowing identifying biological context to be withheld when distinguishing discovery from reproduction of prior knowledge.Source: Discussion · p. 14
  • SeqMaestro generates predictive, testable hypotheses rather than establishing causal mechanisms; its current implementation addresses binary classification and resampling within one dataset is weaker evidence than independent replication.The experiments are not a benchmark against the full range of modern sequence foundation models or highly optimized deep sequence predictors.Source: Discussion · p. 14
  • The framework could be extended with additional transparent model families, independent-dataset replication, and iterative experimental workflows that feed new measurements into later searches.The proposed workflow would test hypotheses generated from one dataset and use resulting measurements to start the next search.Source: Discussion · p. 14
  • SeqMaestro frames machine-learning output as a compact, computable nucleotide-sequence statement that can be inspected, challenged, and tested.The output is more explicit than a prediction or attribution score alone.Source: Discussion · p. 15

Methods

SeqMaestro turns user-provided labeled sequences into prediction tasks, searches combinations of sequence representations and compatible models, then reports compact models alongside feature-reliability evidence. Its workflow includes task checks and a user checkpoint, a shuffled-label control, and distinct agent-driven and fixed-step execution routes.

Overview

  • The workflow serves both users seeking accurate small models and those seeking sequence properties that multiple independent methods support.A run includes task setup, model fitting, feature analysis, and a report of models, scores, and reliability findings.Source: Overview · p. 19

Stage 1: building the prediction task

  • Stage 1 defines the prediction target, sequence windows, and data splits, while checking sequence provenance, filtering, and class-specific losses before user sign-off.A simple compositional model provides a baseline, and sequences are grouped for splitting according to dependencies such as genes or clonal families.Source: Overview · p. 19, Stage 1: building the prediction task · p. 21

Two ways to run SeqMaestro

  • SeqMaestro can run through an agent with stage instructions or as a fixed sequence with documented defaults; both routes produce the same models, records, and checkpoints.The fixed route is less flexible because it cannot improvise, while the agent route retains a user-mediated checkpoint.Source: Figure 6 · p. 20, Two ways to run SeqMaestro · p. 24

Stage 2: fitting the representation-model battery

  • Stage 2 fits compatible combinations of five sequence representations and model families, including language-model-proposed programmatic features and shuffled-label controls.Programmatic proposals may be made at once, refined iteratively, or generated within a growing tree; the battery includes three interpretable families and a gradient-boosted comparison model.Source: Overview · p. 19, Stage 2: fitting the… · pp. 21–22

Stage 3A: reducing to a small model

  • Stage 3A reduces each configuration to at most ten features and presents the best-performing small model from each interpretable family.The feature-reduction procedure repeatedly drops the least-used feature and refits, recording validation performance as feature count falls.Source: Stage 3A: reducing to a small model · p. 22

Stage 3B: the reliability panel

  • Stage 3B assesses feature robustness through resampling selection frequency, separation of unseen classes, and change in hold-out log loss when features are removed.Tier 1 uses 100 half-samples; tiers 2 and 3 use hold-out sequences, and the panel also checks behavior on simulated setups with known answers.Source: Stage 3B: the reliability panel · pp. 22–23

The report

  • The report combines stored run results with explanations, separating traceable descriptions and numbers from AI-written advice or interpretation.Its contents include best small interpretable models and robust concepts or hypotheses.Source: Figure 6 · p. 20, The report · p. 23

Implementation

  • The reported experiments used the agent-driven route with GPT-5.6 Sol, while feature proposal and rating components used GPT-5.4 through Azure OpenAI.Scores use percentile bootstrap intervals from 200 evaluation-set resamples, and the run records code and environment provenance.Source: Implementation · p. 24

Appendix A The system at a glance

This appendix describes SeqMaestro’s configuration and feature-analysis procedures, including stability and hold-out association analyses, reliability assessment, and workflow safeguards. It also clarifies the use of Tj for hold-out association and the limits on interpreting per-feature p-values from multivariate importance.

Appendix A The system at a glance

  • The battery has 13 feature-source/model-class configurations, each fitted at five seeds with real and shuffled training labels; additive models omit pattern features, while boosted trees are uninterpreted accuracy references.Boosted-tree features are recorded but never interpreted.Source: Table 1 · p. 25, Appendix A · p. 25

Executing code that a language model wrote

  • Generated feature code passes syntax-tree inspection, restricted execution, and direct compilation of the inspected tree; these measures do not stop a determined adversary without process isolation.The safeguards make accidental or casual misuse difficult, but isolating the process was not attempted.Source: Executing code that a language model… · pp. 26–27

What is recorded about every feature

Matching features between configurations

  • Features match by absolute Spearman correlation of at least 0.8, with normalized-definition matches taking precedence; maximum-similarity assignment forms concepts but can transitively join distinct properties.The report flags concepts containing two features from the same configuration as a symptom of such chaining.Source: Matching features between configurations · p. 27

Reducing a configuration to a fixed number of features

  • Feature reduction repeatedly drops the lowest-importance features, removing zero-importance features together and dropping the weakest quarter while more than 30 remain, until the fixed budget is met.The final model is refitted on the surviving features.Source: Reducing a configuration to a fixed… · p. 28

Tier 1: would the selection return this feature again

  • With K = 10 reported features, Tier 1 tests selection stability over 100 complementary half-samples and reports thresholded feature- and cluster-level counts with a chance-selection bound.The bound’s usefulness depends on candidate count, so the report indicates when it is informative or uninformative; hold-out tiers follow Tier 1 and Stage 3A.Source: Notation · p. 28, Tier 1: would the selection return… · p. 29

Filtering and grouping, before any label is consulted

  • Before label-based tests, candidates are filtered by variance and prevalence, then clustered by complete-linkage absolute Spearman correlation at 0.8; results are reported per feature and cluster to address split votes.Mutually exclusive indicators at one position are also grouped together.Source: Filtering and grouping, before any… · pp. 28–29

Tier 2: association on the hold-out set

Tier 3: does the model need the feature

Tier 3 tests whether selected features are needed by removing them and measuring loss changes, while calibration checks and panel limitations qualify how its findings should be interpreted. The appendices compare method classes and describe the agent-run workflow.

Tier 3: does the model need the feature

  • Tier 3 refits each model without feature k and measures per-row log-loss change, with paired one-sided testing, a sign-flip check, and Holm correction across K tests.Clusters of two or more reported features are also removed and tested as a block.Source: Tier 3: does the model need the feature · p. 30

Calibration checks

  • Calibration pairs a permuted-label global-null check with a planted-signal recovery check to assess error control and whether effects of comparable size can be recovered.Under the null, Tier 1 agreement between half-samples is informative because the procedure returns K features regardless.Source: Calibration checks · p. 31

The limits of the panel

  • The panel describes rather than selects, requires fresh hold-out data for action, and cannot resolve dataset-to-dataset stability or confounding.Tier 2 is blind to features that matter only in combination; Tier 1 measures resampling stability on one dataset.Source: The limits of the panel · p. 31

Appendix D Comparison with related works

  • Appendix D presents Table 2 as a comparison of SeqMaestro with major classes of interpretable nucleotide-sequence methods.The table key uses ✓, ✗, and N/A for capability support, absence, and non-applicability; an asterisk marks model-class dependence.Source: Appendix D · p. 31, Table 2 · p. 31

Appendix E Running SeqMaestro under an agent

  • Appendix E has the agent use a tested Python library, stage-specific instructions and checkpoints, and a run folder that logs decisions.The agent cannot read hold-out sequences, labels, or scores until the Stage 3A models and stability tier are fixed.Source: Appendix E · pp. 31–32
  • Stage 1 produces sequence-label-group rows, a task description, and a sealed hold-out split, with window, grouping, and model inputs documented.The workflow directs use of checks for orientation, group-aware splitting, and split power.Source: Appendix E · p. 32

Appendix F The prediction tasks used in this study

Appendix F describes four biological questions across six windows, with three windows used for eIF4E transcripts to assess how much each region reveals. It also specifies the hold-out score comparisons and traces feature processing through the pipeline.

Appendix F The prediction tasks used in this study

  • The study uses four biological questions and six windows, including three eIF4E transcript windows to compare how much each region reveals before fitting.Table 3 describes each dataset's window and the variable held to one side of the split to avoid scoring relatives of training sequences.Source: Appendix F · p. 32, Table 3 · p. 33
  • Table 4 reports hold-out AUROC for base-composition baselines, interpretable models, gradient-boosted references, and a chance-separation floor.The caption identifies the strongest Stage 1 baseline and, where needed, the strongest baseline within a ten-feature allowance; it gives no cell values here.Source: Table 4 · p. 34

Appendix G Features and extraction mechanisms

  • Feature sources enter Stage 2 models, then Stage 3A reduces features and Stage 3B assesses robustness and groups similar features into concepts.The report can include best small models and a list of reliable properties; the illustrated feature names do not describe a particular run.Source: Appendix G · p. 34

Appendix H Computational requirements

Computational requirements include both local model-fitting time and language-model waiting, which behave differently as data and pipeline configurations change. Figures 7 and 8 depict feature flow and summarize runtime, fitting, token, and cost measurements.

Appendix H Computational requirements

  • Local fitting time rises with sequence and candidate-feature counts, while language-model call counts depend on the Stage 2 battery and remain roughly similar for datasets hundreds of times larger.The six study runs were sequential on a single ten-core workstation without a GPU.Source: Appendix H · p. 34
  • Figure 7 traces thirteen illustrative features through pipeline stages, from sequence-derived or language-model proposals to reduced models, reliability analysis, and report outputs.Its feature names borrow from the PTBP1 task and do not represent a particular run.Source: Figure 7 · p. 35
  • Figure 8 covers run time, fit time versus sequence count, feature-pool elimination, request tokens, and API costs across configurations and seed settings.The caption notes an extrapolation to one million sequences and a widened-pool comparison, but provides no plotted values here.Source: Figure 8 · p. 36

Wall-clock time

Most task setups finished within half a day, while the largest took two days; on smaller setups, API response waiting accounted for roughly three quarters of elapsed time.

Wall-clock time

  • Five of six task setups finished in 8–12 hours, while the 499,672-sequence RNA polymerase II pausing task took two days.The reliability panel took under two hours on the five smaller setups and thirteen hours on the largest, which used seven configurations and twenty half-samples.Source: Wall-clock time · p. 36
  • Roughly three quarters of elapsed time on smaller setups went to language-model waiting, and the pace was set by API latency rather than the processor.The iterative proposer ran thirty serial fits at five to eight minutes each; ten adaptive-tree fits ran concurrently and finished together in about eighty-five minutes.Source: Wall-clock time · p. 36

How fitting and the panel scale

Fit time grows with sequence count and candidate-feature count, while reliability selections cost several times more than a single fit.

How fitting and the panel scale

  • An additive fit on a million sequences is projected to take about seven minutes, while a hundred selections would take about 90 hours.The million-sequence estimate is twice the largest dataset in the study.Source: How fitting and the panel scale · pp. 36–37
  • A reliability selection costs four to seven times a single fit and grows with sequence count at much the same rate.One selection consists of a full-pool fit followed by elimination to ten features, about twenty-five fits on shrinking pools.Source: How fitting and the panel scale · p. 37
  • With 3,000 sequences, an additive fit takes four seconds on 404 per-position columns and three minutes on 5,864 columns after adding k-mer counts.Source: How fitting and the panel scale · p. 37

Tokens

The section reports the token and runtime costs of SeqMaestro and explains how its language-model roles, feature inputs, and interpretable model families fit together.

Tokens

  • Feature ratings take about 2,000 tokens, one-shot proposals about 13,000, and iterative refinement about 67,000 across twelve calls.A one-shot proposal serves all three models fitted on it.Source: Tokens · p. 37
  • A depth-5 tree with five refinement rounds and ten candidates per round costs about 5.3 million tokens for one fit; a depth-3 tree with one round costs about a seventh as much.Joint generation requests candidates at every expanded node.Source: Tokens · p. 37

Cost

  • Thirteen configurations at one seed use about 6 million tokens, costing $9–$56 at September 2026 list prices, or $1–$8 without joint generation.Five seeds with shuffled-label controls use about ten times as many tokens; the prices assume prompt caching.Source: Cost · p. 37

Appendix I A design agnostic to the language model and the agent harness

  • The agent harness needs an isolated working shell, instruction loading, and approval pauses; the same stage instructions work across environments, and library models are chosen independently.The agent-driving model and the library model that proposes and rates features can be selected separately.Source: Appendix I · pp. 37–38, Figure 9 · p. 38

Appendix J Understanding the interpretable models

  • Appendix J uses PTBP1 exon examples to explain three transparent model families; its tree figure shows feature splits, class shares, and an example exon’s path.Source: Appendix J · p. 38, Figure 12 · p. 41

What every model is given

  • Models receive numerical features computed from sequences, including positional indicators and programmatic measurements of composition, word counts, run lengths, or regional contrasts.Programmatic features can encode position, distance to a landmark, spacing, or conditional measurements.Source: What every model is given · p. 38

The sparse linear model

  • The sparse linear model sums standardized feature values multiplied by weights; its L1 penalty leaves most weights at zero, yielding a small set of directional additive associations.Weights measure log-odds change per standard deviation of a feature.Source: The sparse linear model · pp. 39–40

The additive model

  • The additive model replaces weights with feature-specific curves that can reveal gradients, thresholds, saturation, or reversals, though extreme curve regions may rest on few sequences.Each feature contributes its curve height, with no terms combining two features in the fitted model.Source: The additive model · pp. 40–41

The decision tree

A decision tree expresses predictions as paths of threshold tests, making context-dependent sequence rules visible while leaving tree selection uncertain across data splits.

The decision tree

  • Each tree split tests whether a feature is at or below a training-chosen threshold, and each sequence follows one yes-or-no path to a prediction.The threshold is selected to best separate the classes among sequences reaching that box.Source: The decision tree · p. 41
  • A tree path can express a context-dependent rule: a base 51 nt before the exon separates 48% silenced from 94% only among exons meeting specified CU and UCU conditions.The same base alone separates the classes hardly at all; repeated thresholds on different branches can describe graded effects or context-shifting thresholds.Source: The decision tree · pp. 41–42, Table 6 · p. 42
  • The table lays out four paths as conditions; its second and third rows are the comparison showing a base that matters only in one context.Source: The decision tree · p. 41, Table 6 · p. 42
  • A tree drawn from one sample is not unique: other trees can fit the same sequences about equally well, so the reliability panel checks whether features recur on another half of the data.Source: The decision tree · p. 42

Appendix K The full analysis reports

Six self-contained reports document analysis runs across the stages of the workflow and several biological tasks.

Appendix K The full analysis reports

  • Each run report covers preparation decisions and costs, configurations with shuffled-label controls, model figures and feature tables, reliability and calibration checks, and per-feature evidence.Source: Appendix K · p. 42
  • The listed reports cover Polycomb nucleation, RNA polymerase II pausing, PTBP1 exon regulation, and three eIF4E-dependence sequence windows.The eIF4E windows are the first fifty coding-region codons, the last thirty 5′ UTR bases, and the last hundred 3′ UTR bases.Source: Appendix K · p. 42
Loading 2609.14882v1…