PaperlayerView exact revision on arXiv

Source-linked AI summary

BaseCamp --- An Agentic AI Framework for Automating DNA Sequencing Data Pipelines

Eranga Bandara, Xueping Liang, Asanga Gunaratna, Tharaka Hewa, Abdul Rahman, Peter Foytik, Safdar H. Bouk, Sachini Rajapakse, Isurunima Kularathna, Pramoda Karunarathna, Chalani Rajapakse, Ng Wee Keong, Kasun De Zoysa, Amin Hass, Wathsala Herath, Ross Gore, Ravi Mukkamala, Nihal Siriwardanagea, Gihan Siriwardanagea, Aruna Withanage, Nilaan Loganathan, Sachin Shetty

arXiv:2609.28557v1cs.AIq-bio.GN

TL;DR

Sequencing workflow managers reliably execute established tools, but leave key sample-specific decisions to people. BaseCamp assigns that decision layer to six human-supervised agents while tools perform sequence analysis. Its proof of concept reports expert-concordant configurations, an inspectable filtering ledger, and cross-stage anomaly detection.

  • Problem

    Workflow managers execute specified pipelines reproducibly but leave sample-specific interpretation, variant adjudication, and failure response to human operators.

    Source: §3.1 · p. 13, §4 · p. 18

  • Method

    Six specialized agents coordinate the sequencing decision layer under human supervision, configuring and interpreting established tools rather than performing sequence analysis.

    Source: §7 · p. 41

  • Results

    99.1% call set concordance with expert configuration, alongside an inspectable filtering ledger and detection of cross-stage anomalies that execution monitoring misses.

    Source: §6.5 · p. 40, §7 · p. 41

  • Takeaways & Limitations

    The filtering ledger makes removed calls inspectable, while monitoring surfaces jointly implausible samples that pass individual checks.

    Source: §7 · p. 41

    The evaluation does not establish that agentic configuration produces more accurate variant calls than expert manual configuration across sample types and study designs.

    Source: §6 · p. 31

Abstract

from arXiv · show

DNA sequencing pipelines, spanning quality control, alignment, variant calling, and annotation, are now reliably executed by workflow management systems that orchestrate established bioinformatics tools at scale. What remains manual is the decision layer surrounding that execution: selecting quality thresholds appropriate to a sample and platform, adjudicating borderline variant calls, diagnosing anomalies, and determining which findings warrant expert review. These decisions are repetitive, judgment-intensive, inconsistent across operators, and frequently undocumented. This paper introduces BaseCamp, a novel agentic AI framework for automating the decision layer of DNA sequencing pipelines. The framework decomposes the pipeline into six specialized AI agents, covering sample intake and quality control, alignment, variant calling, annotation, cross-stage monitoring, and reporting. Critically, BaseCamp agents do not perform sequence analysis: established tools execute alignment, calling, and annotation, while the agents select among them, configure them, interpret their output, and decide what follows. This confines language model reasoning to the judgment layer where it is reliable and preserves the reproducibility existing tooling guarantees. Agent reasoning is powered by a consortium of fine-tuned, domain-specialized large language models coordinated by a central reasoning LLM, executing locally so no sequencing data leaves the operating environment, under human-in-the-loop orchestration. Evaluation shows agent-generated configurations are concordant with expert practice, that an explicit filtering ledger renders inspectable what filtering otherwise removes without trace, and that cross-stage anomaly detection surfaces conditions execution monitoring misses. BaseCamp offers a generalizable blueprint for agentic automation of scientific data pipelines.

1. Introduction

Sequencing tools execute mature, reproducible workflows, but consequential choices around thresholds, borderline calls, failures, and review remain manual. BaseCamp proposes a human-supervised, multi-agent framework for that decision layer, with a real-world proof of concept.

1. Introduction

  • Despite mature stage-specific tools and reproducible workflow managers, sample- and platform-specific thresholds, borderline calls, failures, and review priorities remain manual and inconsistent.These decisions depend on bioinformatician availability and vary across operators and sequencing runs.Source: §1 · p. 3
  • Existing genomics LLM applications target isolated sequence or literature tasks, leaving coordinated reasoning from raw reads through reviewed, reported variants an open challenge.The paper identifies the integration of multiple autonomous agents across the sequencing lifecycle as an unresolved domain challenge.Source: §1 · pp. 3–4
  • Sequencing automation needs a framework for agent roles, tool coordination, and human oversight because pipeline decisions can affect clinical, breeding, or diagnostic outcomes.Conventional automation handles deterministic execution but not contextual judgments such as assessing borderline calls or unusual quality profiles.Source: §1 · p. 4
  • BaseCamp decomposes the decision layer into specialized agents supervised through MCP-enabled checkpoints and supported by fine-tuned domain models coordinated by a reasoning LLM.The agents reason over operational artifacts; established bioinformatics tools remain responsible for the underlying computation.Source: §1 · pp. 4–5
  • A real-world proof of concept demonstrates BaseCamp’s practical applicability and reports reduced manual configuration overhead and improved anomaly detection across sequencing runs.The paper presents this validation as one of its four primary contributions.Source: §1 · p. 5

2. Background

The background establishes why BaseCamp uses language models to interpret sequencing artifacts, tool-using agents to coordinate decisions, and oversight mechanisms to keep those decisions accountable and reproducible.

2.1. Large Language Models

  • LLMs provide broad language and reasoning capabilities, while BaseCamp applies them to pipeline artifacts rather than nucleotide sequences or sequence-level analysis.The LLMs interpret reports, statistics, records, tables, and logs; established bioinformatics tools perform the computation.Source: §2.1 · pp. 6–7

2.2. Reasoning LLMs

  • The central reasoning LLM compares and synthesizes independent outputs from specialized models into a final decision or recommendation.This ensemble approach is intended to reduce model-specific bias or error propagation by validating decisions across multiple perspectives.Source: §2.2 · p. 7

2.3. LLM Fine-Tuning

  • Fine-tuning on curated sequencing operations data adapts a general model to domain vocabulary and reasoning patterns; LoRA and 4-bit quantisation reduce training resource requirements.The models are served locally, so sequencing data need not leave the operating environment.Source: §2.3 · pp. 7–8, Figure 1 · p. 8

2.4. AI Agents and Agentic AI

  • Agents manage prompts, model calls, tool invocation, response interpretation, and subsequent actions, while established tools compute and agents decide what follows.This division grounds agent reasoning in tool outputs and preserves existing tooling’s reproducibility guarantees.Source: §2.4 · p. 9, Figure 2 · p. 10

2.5. Model Context Protocol

  • MCP provides structured connections between agents and external systems; BaseCamp exposes workflows through separate servers that can be updated or rerun independently.A disputed stage can be rerun without recomputing the entire pipeline.Source: §2.5 · pp. 10–11

2.6. Responsible and Explainable AI

  • Responsible and explainable AI matter because sequencing decisions can silently remove calls and inform consequential downstream decisions.BaseCamp uses reasoning traces, human approval gates, and recorded configurations to support transparency, accountability, and reproducibility.Source: §2.6 · pp. 11–12

3. Related Works

Prior work addresses reproducible tool execution, genomic and biomedical analysis, and broader agentic scientific work, but leaves gaps in automating and overseeing sequencing-pipeline decisions. BaseCamp positions itself as a coordinated, human-supervised decision layer that works with established tools rather than replacing them.

3.1. Bioinformatics Workflow Management and Pipeline Automation

  • Workflow managers and curated pipelines reproducibly execute specified tools and parameters, but do not choose sample-appropriate thresholds, interpret reports, or diagnose and respond to failures.BaseCamp is presented as complementary: its agents supply reasoning above established tooling and, where appropriate, workflow engines.Source: §3.1 · pp. 12–13

3.2. LLM-Based Agents for Genomics and Biomedical Analysis

  • Genomic language models operate on sequence, while bioinformatics agents such as AutoBA and CellAgent focus on downstream analyses; BaseCamp targets upstream production decisions and adds local fine-tuned models and human oversight.The cited gap also includes limited structured approval gates, decision traces, and configuration recording in prior systems.Source: §3.2 · pp. 13–14

3.3. Agentic AI for Scientific Discovery and Autonomous Experimentation

  • BaseCamp narrows scientific-agent autonomy to automating decisions in an established production pipeline under explicit human approval gates, not generating hypotheses or conclusions.The stated aim is consistency and throughput in a known process, rather than autonomous discovery.Source: §3.3 · pp. 14–15

3.4. Agentic AI and Multi-Agent Workflow Systems

  • General agent frameworks identify coordination and orchestration challenges [13] [14] [36], while BaseCamp confines agents to decisions around established tools, preserving reproducibility and keeping LLM reasoning within its stated reliable scope.This differs from operational agent systems in which agents both reason and perform the work product.Source: §3.4 · pp. 15–16

3.5. Positioning

  • BaseCamp’s distinguishing feature is the conjunction of end-to-end sequencing coverage, established tooling, locally served fine-tuned models, human approval gates, and recorded configurations for reproducible runs.Table 1 is presented as a comparison of prior works with BaseCamp.Source: §3.5 · p. 16, Table 1 · p. 17

4. Manual DNA Sequencing Pipeline Workflows

Sequencing tools and workflow managers automate execution, but consequential decisions remain manual, context-dependent, and often undocumented. The section traces those decisions from preprocessing and alignment through filtering, annotation, and exception handling.

4. Manual DNA Sequencing Pipeline Workflows

  • Tool execution is largely automated, but sample-specific decisions remain manual, operator-dependent, and often undocumented across sequential pipeline stages.Informal handoffs and reactive anomaly detection can leave issues undiscovered until they have propagated downstream.Source: §4 · pp. 17–18, Figure 3 · p. 19

4.1. Raw Read Quality Control and Preprocessing

  • QC interpretation and trimming depend on library type and protocol; over-trimming discards signal, while under-trimming can impair alignment statistics.Batch-level parameter choices may not suit outliers, whose need for individual review is judged manually.Source: §4.1 · pp. 19–20

4.2. Reference Alignment and Parameter Selection

  • Alignment review requires joint interpretation of mapping rate, insert size, coverage, and other statistics to distinguish possible library, contamination, or reference problems.Reference choice can also materially affect mapping and variant discovery, especially for non-model organisms.Source: §4.2 · p. 20

4.3. Variant Calling and Filtering

  • Variant filtering balances sensitivity against specificity according to the downstream use, but thresholds are often chosen by convention and removed calls leave no trace unless recorded.Manual rationales may be unrecoverable when results are questioned later.Source: §4.3 · pp. 20–21

4.4. Annotation and Functional Interpretation

  • Analysts manually prioritize annotated variants using question-specific filters, and trait-mapping may require joining variant and phenotype data from separate systems.Filtering logic is rarely captured with results, making prioritized lists difficult to reproduce.Source: §4.4 · pp. 21–22

4.5. Pipeline Monitoring and Exception Handling

  • Execution monitoring detects failed jobs but misses analytical anomalies when jobs complete; manual detection and diagnosis are reactive and depend on records often not retained.Issues may be found only after affected results have been used, limiting intervention before compromised results propagate.Source: §4.5 · p. 22

5. Delegating Manual Sequencing Processes to AI Agents

BaseCamp delegates pipeline judgment to coordinated agents while leaving sequence analysis to established tools. Its agents pass structured context between stages, with human review for consequential decisions and ambiguous cases.

5. Delegating Manual Sequencing Processes to AI Agents

  • Established tools perform alignment, calling, and annotation; agents select and configure tools, interpret outputs, and determine next steps, keeping LLM reasoning in the judgment layer.Human adjudication is specified for samples flagged as failing or ambiguous.Source: §5 · p. 23, §5.1 · pp. 24–25
  • BaseCamp decomposes manual decisions into specialized agents that exchange structured outputs, so downstream stages can use upstream context rather than isolated handoffs.Six agents cover intake and QC, alignment, variant calling, annotation and interpretation, pipeline monitoring and exceptions, and reporting.Source: §5 · pp. 22–23, Figure 4 · p. 24

5.1. Sample Intake and QC Agent

  • The Sample Intake and QC Agent interprets quality metrics in light of library type, platform, and batch position to recommend justified trimming parameters and pass, review, or fail disposition.It identifies samples warranting individual attention and passes its outputs to the Alignment Agent.Source: §5.1 · pp. 23–25

5.2. Alignment Agent

The Alignment Agent selects reference and aligner configurations, then assesses alignment results. It reasons jointly over metrics to distinguish potential failure signatures.

5.2. Alignment Agent

  • The agent selects a reference and aligner configuration, invokes alignment and post-processing, and assesses the resulting statistics.It connects to BWA, Bowtie2, and SAMtools, and may use sample provenance and reference repositories.Source: §5.2 · p. 25
  • For non-model organisms, it uses sample provenance, available assemblies, and mapping behaviour to recommend a reference with explicit justification.Source: §5.2 · p. 25
  • It interprets mapping, pairing, insert-size, coverage, and duplication metrics jointly to distinguish signatures such as library problems, contamination, or reference divergence.The assessment passes to the Variant Calling Agent.Source: §5.2 · p. 25

5.3. Variant Calling Agent

The Variant Calling Agent selects a caller and filtering strategy using sample and study context, then makes filtering decisions inspectable and routes borderline calls for human review.

5.3. Variant Calling Agent

  • The agent invokes the selected GATK, DeepVariant, or bcftools caller and receives aligned data and upstream quality context.Source: §5.3 · p. 25
  • Caller selection and filtering thresholds are reasoned from depth, platform, ploidy, variant type, and the downstream application.The agent recommends thresholds for depth, quality by depth, strand bias, and mapping quality, with explicit justification.Source: §5.3 · p. 25
  • A filtering ledger records each threshold, removed-call count, and rationale, and travels downstream with the call set.This keeps exclusions inspectable rather than merely absent.Source: §5.3 · pp. 25–26
  • Calls near threshold boundaries are flagged individually for human adjudication rather than silently resolved.Source: §5.3 · p. 26

5.4. Annotation and Interpretation Agent

The Annotation and Interpretation Agent prioritizes enriched variants into a reproducible shortlist, while the Pipeline Monitoring and Exception Agent detects cross-stage anomalies and presents evidence-backed alerts for human judgment.

5.4. Annotation and Interpretation Agent

  • The agent enriches filtered calls with functional consequences, population frequencies, and database evidence using VEP, SnpEff, ANNOVAR, gnomAD, and dbSNP.Source: §5.4 · p. 26
  • Reasoning over variant consequence, frequency, region, inheritance, and study objective reduces thousands of calls to a reviewable, reproducible shortlist.The agent records the filter logic; trait-mapping and breeding contexts also use phenotypic records retrieved from separate systems for genotype–phenotype integration.Source: §5.4 · p. 26

5.5. Pipeline Monitoring and Exception Agent

  • Monitoring covers coverage shortfalls, contamination or index hopping, unusual variant-call rates, batch effects, and stalled or failed stages.Source: §5.5 · p. 26
  • Cross-stage comparisons against run-specific and historical expectations detect analytical inconsistencies that individual stages may not flag.Alerts prioritize urgency, cite triggering metrics and comparisons, and propose evidence-supported diagnoses to reduce backward tracing.Source: §5.5 · p. 27
  • Unlike execution monitoring, the agent can surface analytical anomalies even when every job completes successfully.Workflow monitoring reliably surfaces crashes, but successful execution does not rule out wrong results.Source: §5.5 · pp. 26–27

5.6. Reporting Agent

The Reporting Agent assembles the prioritized variant set, reasoning traces, filtering ledger, pending adjudications, and run configuration into audience-targeted, traceable reports.

5.6. Reporting Agent

  • The final artifact combines prioritized variants, per-stage reasoning traces, the filtering ledger, outstanding flagged items, and complete run configuration.Configuration includes tool parameters and versions, model versions, and adapter revisions.Source: §5.6 · p. 27
  • Technical and result-focused reports tailor findings to bioinformatics teams and requesting scientists while tracing each finding to its producing stage and decision.Recorded settings make it possible to examine whether another defensible configuration would have produced a different result.Source: §5.6 · p. 27

5.7. Agent Decomposition and Workflow Coordination

BaseCamp decomposes pipeline work into independently addressable workflows that a bioinformatician can supervise at consequential checkpoints. Shared structured outputs preserve upstream reasoning for downstream agents, while disputed stages can be rerun independently.

5.7. Agent Decomposition and Workflow Coordination

  • Six independent workflows let one bioinformatician supervise and intervene across pipeline functions through a unified natural-language interface.Human checkpoints include QC dispositions, alignment configurations, filtering thresholds, flagged variants, and escalated exceptions.Source: §5.7 · pp. 27–28, Figure 5 · p. 29
  • A disputed stage can be interrogated or rerun alone, avoiding recomputation of the full pipeline.Agents can also be refined, retrained, or extended independently.Source: §5.7 · pp. 28–29
  • Structured intermediate outputs preserve upstream context, so downstream agents receive the reasoning behind prior results rather than results alone.For example, QC assessment informs alignment interpretation, which in turn informs confidence in marginal calls.Source: §5.7 · p. 28

5.8. Responsible and Explainable AI Agents

BaseCamp combines domain-specialized LLM recommendations through a central reasoning LLM, escalating unresolved disagreement to human judgment. Reasoning traces, approval gates, and recorded configurations support auditability and reproducibility.

5.8. Responsible and Explainable AI Agents

  • Independent domain-specialized models make recommendations that the reasoning LLM synthesizes; unresolved divergence is recorded and escalated to a human rather than averaged.Each recommendation cites the pipeline metric it relied upon, and the consolidated decision includes an explicit rationale.Source: §5.8 · pp. 29–30, Figure 6 · p. 30
  • Every agent output records its reasoning, the metrics and thresholds used, and the criteria and call counts associated with filtering decisions.The trace makes the basis and effects of decisions inspectable.Source: §5.8 · p. 30
  • Human approval gates protect consequential transitions, while complete run configurations are recorded so results can be reproduced or recomputed with different settings.Filtered call sets require validation before proceeding downstream, and reports require approval before release.Source: §5.8 · p. 30

6. Implementation and Evaluation

BaseCamp’s prototype uses specialized agents to guide established bioinformatics tools, with evaluation focused on buildability, inspectable decisions, and anomaly detection—not proof of superior variant accuracy. On held-out material, call sets were highly concordant with expert configurations, while monitoring surfaced anomalies earlier than historical downstream discovery.

6.1. Implementation Overview

  • Agents invoke established bioinformatics tools through separate MCP servers, while the complete deployment remains within institutional compute and sends no sequencing data externally.Bioinformaticians use a unified interface to review, approve, or override decisions.Source: §6.1 · pp. 31–32, Figure 7 · p. 32

6.2. Fine-Tuning Corpus and Configuration

  • Fine-tuning records pair historical operational artifacts with experienced operators’ decisions and rationales; held-out runs are disjoint from training runs.Function-specific adapters are hot-swapped while one base model remains resident in memory.Source: §6.2 · pp. 32–33, Table 2 · p. 33
  • Inter-operator agreement provides a reference ceiling for agent agreement because some decision classes allow multiple defensible answers.The figure describes exception diagnosis as the weakest class, but its plotted values are placeholders pending measurement.Source: §6.2 · p. 35, Figure 8 · p. 35

6.3. Evaluation of the Variant Calling Agent

  • Variant configurations received a 4.6/5 evaluator rating, and the filtering ledger made discarded information inspectable while documenting departures from laboratory defaults.The agent selects a caller and configures parameters from upstream context, sample metadata, and the stated downstream application.Source: §6.3 · pp. 35–38

6.4. Evaluation of the Pipeline Monitoring and Exception Agent

  • Cross-stage monitoring surfaced anomalies at a median of under one hour after batch completion, compared with 6 days historically; alerts were rated 4.5/5 for interpretability and actionability.Batch effects were the weakest detection class because protocol differences can require context the agent lacks.Source: §6.4 · pp. 38–39

6.5. Discussion

  • 99.1% call-set concordance with expert configuration; anomalies were surfaced at a median of under one hour, versus a historical median of 6 days.The evaluation also reports that 4.3% of decisions were escalated for insufficient context rather than silently resolved.Source: §6.4 · p. 39, §6.5 · p. 40
  • Cross-stage inconsistency detection addresses a gap in execution monitoring, which does not identify samples that pass individual checks but are jointly implausible.The reported reduction in detection latency is described as the practical value of catching anomalies before results propagate downstream.Source: §6.5 · p. 40
  • The evaluation establishes concordance with the practices of the contributing laboratory, not superior accuracy or generalization across laboratories, sample types, organisms, or study designs.Whether agentic configuration produces more accurate calls than expert manual configuration remains unestablished.Source: §6 · p. 31, §6.5 · p. 40

7. Conclusion and Future Works

BaseCamp delegates sequencing-pipeline decisions to agents while established tools retain sequence analysis, and its proof of concept demonstrates expert-concordant configurations and new visibility into filtering and cross-stage anomalies. Evaluation remains bounded by one laboratory’s conventions, with cross-laboratory transfer and accuracy against benchmark truth sets left for future work.

7. Conclusion and Future Works

  • Although workflow systems automate tool execution, sample-specific thresholds, borderline calls, anomaly diagnosis, and expert-review decisions remain manual.These decisions are repetitive, inconsistently exercised, and often undocumented.Source: §7 · p. 41
  • Six agents handle the decision layer while established tools perform sequence analysis, preserving existing-tool reproducibility under human oversight.Agents configure tools, interpret their outputs, and decide what follows.Source: §7 · p. 41
  • Agent configurations were concordant with expert practice on held-out runs, while the filtering ledger exposed removed calls and monitoring detected cross-stage anomalies execution monitoring missed.The framework ran locally without data egress, and decision latency was negligible relative to tool execution time.Source: §7 · p. 41
  • The evaluation establishes concordance, not superiority, and measures it against the conventions of the laboratory whose records trained the adapters.Source: §7 · p. 41
  • Cross-laboratory validation is needed to test whether adapters encoding one laboratory’s conventions transfer, while examining expert departures from defaults could test contextual reasoning.Prospective benchmark truth sets would permit direct sensitivity and precision measurement, which the present evaluation cannot provide.Source: §7 · p. 41
  • Testing long-read platforms and non-model organisms would assess whether the agent decomposition holds when tools change and reference selection becomes consequential.The paper also identifies downstream genotype–phenotype analysis as a possible extension.Source: §7 · p. 42
  • Recording decision rationales alongside configurations could preserve reasoning that would otherwise be discarded and unavailable when results are later questioned.Source: §7 · p. 42
Loading 2609.28557v1…