Skip to content
Back to Liquid Agent DOCUMENTATION

Liquid-Biopsy Capability Matrix

This page is the single overview of the current user-facing Liquid Agent capability surface. It maps callable models, tools, methods, and workflow blocks to the liquid-biopsy data types they handle.

Status Semantics

Status Meaning
Internal executable Liquid Agent can run this through the Python kernel, scripts, shell, Web API, or autopilot when required inputs exist.
Internal model / encoder The model or deterministic encoder is registered in the blood-signal encoding layer.
External runtime wrapper Liquid Agent can check, install/prepare where possible, smoke-test, and call the runtime through liquid-agent tools. Real analysis may still need model files, references, manifests, or compatible input files.
External method guidance The method advisor can recommend it, show requirements, and choose internal fallbacks, but no stable local wrapper is registered yet.
Reimplementation candidate The original method is useful but the public runtime is too old or unstable for ordinary users; prefer a future small internal proxy or a maintained alternative.
LLM engine Used for conversation, planning support, documentation context, result explanation, and recovery suggestions. It is not a biomedical signal model.

Recently Added Kernel Features

Feature Status Data or question type User-facing entrypoints
Frozen prediction studies Internal executable Explicit binary train/validation contracts; logistic, SVM and forest candidates; training-only preprocessing and threshold selection; ROC/PR/calibration and frozen probabilities Natural-language Web/CLI tools; see prediction studies. Not a separate Plan-button entry or automatic clinical cohort design.
Paired prediction comparisons Internal executable Two to six completed feature-set studies with matching cohorts and unchanged inputs/probabilities; paired AUROC differences without parent refitting Natural-language Web/CLI tools and owned Results reports; exploratory intervals, not multiplicity-adjusted or clinical validation
Liquid-biopsy method advisor registry Internal executable Method/tool questions across fragmentomics, methylation, copy-number, variants, cfRNA, small RNA, CTC tables, and proteomics /methods, liquid-agent methods, scripts/run_liquid_biopsy_method_advisor.py, Web Method advice button, Python API
Named-tool ranking Internal executable Questions that explicitly name tools such as cfDNAPro, WisecondorX, FinaleMe, or cfTools Natural language, /methods, CLI, Web API
Dependency and manual-resource status Internal executable External methods that require executables, R packages, Python modules, Java, reference panels, model files, or code bundles Method-advice JSON/Markdown reports
External runtime manager Internal executable Selected non-kernel Python, R/Bioconductor, source, Snakemake, and system-binary tools liquid-agent tools status/install/smoke/run, external-tool status reports
Supplied signal-matrix analysis Internal executable Processed CNV, methylation, EPIC-like methylated/unmethylated, or other liquid-biopsy numeric matrices scripts/run_cfdna_analysis_suite.py, scripts/run_cfdna_plot_suite.py, /plan, /autopilot, Web results
Metadata profile scanner Internal executable CSV, TSV, Excel, Parquet, JSON, JSONL, and NDJSON metadata or label tables /metadata, Web Metadata card, Web scan/plan responses, planner inputs
Label-aware planning Internal executable Datasets with candidate sample labels, groups, response/status columns, or user-selected labels /plan, /metadata use, Web Change/Ignore controls; chooses unsupervised, grouped, or exploratory supervised mode
Exploratory supervised modeling Internal executable when backend is available Labeled feature stores, embeddings, or processed signal matrices with adequate matched labels Standard cfDNA analysis task; sklearn_logistic_regression by default, optional pytorch_linear_probe for larger labeled cohorts
Markdown-first result reports Internal executable Completed runs, generated tables, static figures, method-advice reports, and result evaluations Web Results panel, liquid-agent results, autopilot reports; hides backend JSON/TXT/HTML clutter from the default user view
Agent planning integration Internal executable Dataset folders with blood raw candidates, fragmentomics outputs, CNV inputs, coverage tracks, variant inputs, region-signal tables, or browser-track style inputs /plan, /autopilot, Web scan/plan
LiquidBiopsyDataState summary Internal executable Any scanned liquid-biopsy source or generated result set Backend planner, result evaluator, plan ledger, autopilot report; records signal families, input/output counts, metadata coverage, minimum-output contract gaps, blockers, and safe next actions without adding user-facing controls
Liquid-biopsy FeatureBook contracts Internal executable Fragmentomics, methylation, copy-number, variant, supplied matrix, archive, and metadata/grouped-comparison contexts Planner, result evaluator, plan ledger, autopilot report, Python API; tracks expected artifacts, QC checks, interpretation limits, and satisfied/pending contract status
ToolCard output-kind verification Internal executable Completed backend tasks and external-wrapper task summaries Executor, plan ledger, autopilot report; verifies minimum output kinds such as summary, table, figure, method-advice report, review, or feature store; a bare output directory is not treated as completion, but real result files inside a reported output/run directory are sampled and counted
Content-derived result signals Internal executable Ranked effect tables, grouped summaries, label-metric summaries, sample outlier tables, and summary JSONs from liquid-biopsy runs Result evaluator, planner notes, plan ledger, autopilot report; can promote ready visualization, raw-signal follow-up, signal-aware method advice, or review tasks without adding user-facing controls
Result-driven analysis concepts Internal executable Content-derived liquid-biopsy result signals plus FeatureBook contracts Result evaluator, planner notes, plan ledger, autopilot report; converts real evidence into prioritized auditable follow-up questions, suggested task families, QC checks, interpretation limits, and stable novelty keys without adding user-facing controls
Method-advice internal fallback routing Internal executable Method-advice JSON with top methods, resource limits, and internal fallback routes Planner and plan ledger; can map fallback evidence back to ready standard cfDNA analysis, standard visualization, raw-signal analysis, or raw-signal visualization without adding user-facing controls
Source-aware Web method advice Internal executable Single-source or multi-source Web sessions POST /api/session/{session_id}/methods/advice
Planner robustness for incomplete candidates Internal executable Folders with incomplete candidate files where encoding cannot be safely planned yet Planning continues with feasible tasks and records a note instead of crashing
Documentation and prompt examples for method advice Documentation Natural-language examples and reproducible command forms README, docs site, script cookbook

Direct Internal Workflow Blocks

Explicit Assay Contracts

These operations use a validated *.assay.json declaration or the controller's configure_assay tool. They are not inferred from a filename or started by loading a skill. See schemas and limits.

Operation Implemented engine Boundary
Digital-PCR accepted partitions Occupancy concentration and transformed exact-binomial 95% intervals No raw droplet gating, duplex fitting or clinical detection threshold
CTC accepted counts/volume Cells per mL and exact Poisson 95% intervals No image-based identity assignment or enrichment correction
RNA, protein, metabolite and EV processed matrices Missingness, distributions, sample/feature QC, bounded PCA and heatmaps Not raw instrument identification or biological source validation
Methylation beta matrices Range validation, missingness and sample/probe displays Not IDAT normalization, conversion QC or DMR fitting
Independent-group RNA raw counts PyDESeq2, Wald tests and adjusted p-values Explicit separate request; no paired/batch/time design; experimental replication required

Every completed operation produces a Markdown report, actual PNG/CSV outputs and provenance. Optional dependencies are checked when used.

Existing Workflow Blocks

Workflow block Status Main data types Main entrypoints Typical outputs
Dataset/source scanning Internal executable Folders, files, multiple attached liquid-biopsy sources /use, /sources, Web source manager, scan_project_profile(...) Detected signal candidates, labels, outputs, source inventory
Blood-signal preprocessing Internal executable BED/BED.GZ, peak files, BAM/CRAM, VCF/MAF, CSV/TSV/parquet variant tables /preprocess-style natural language, autopilot, scripts/preprocess_*_signal.py, Python API Cleaned intervals, region-signal tables, bin counts, segments, arm burden, normalized variant tables, summaries
Blood-signal encoding Internal executable cfChIP/cfMeDIP/MeDIP intervals or alignments, LPWGS/ULPWGS intervals or CNV tables, VCF/MAF variant tables, continuous tracks /blood, natural language, scripts/encode_*_features.py, Python API Feature stores, encoder summaries, reusable per-sample vectors
Standard cfDNA analysis Internal executable Feature stores, fragmentomics summaries, methylation-proxy tables, CNV directories, region-signal tables, supplied CNV/methylation/signal matrices, labels scripts/run_cfdna_analysis_suite.py, autopilot, Python API Distance/correlation/outlier summaries, grouped metrics, supplied-matrix summaries, module summaries
Standard cfDNA visualization Internal executable Feature stores, metadata/labels, fragmentomics summaries, methylation-proxy tables, CNV directories, supplied CNV/methylation/signal matrices scripts/run_cfdna_plot_suite.py, autopilot, Python API UMAP/t-SNE/PCA scatter, heatmaps, grouped plots, supplied-matrix PNG outputs, static figures for markdown reports, optional Plotly HTML files from Python
Raw-signal visualization Internal executable Fragment directories, CNV/bin directories, coverage tracks, region-signal tables, variant inputs, end-motif tables, loci tables scripts/run_cfdna_raw_signal_suite.py, autopilot, Python API Fragment-length plots, genomewide profiles, sample/bin heatmaps, region metaprofiles, VAF plots, motif plots, browser-track inventories, optional Plotly HTML files from Python
Raw-signal numeric analysis Internal executable Fragment directories, CNV/bin directories, coverage tracks, region-signal tables, variant inputs, end-motif tables, sample-time tables, browser-track files scripts/run_cfdna_raw_signal_analysis_suite.py, autopilot, Python API Fragmentomics metrics, CNV summaries, arm burden, VAF summaries, longitudinal summaries, browser-track summaries, numeric report JSON
Method/tool advice Internal executable Dataset folders and natural-language method questions /methods, CLI, Web API, Python API liquid_biopsy_method_advice.json, liquid_biopsy_method_advice.md
Result review and report summary Internal executable Existing Liquid Agent outputs and reports /review, autopilot final pass, Python API Consolidated review JSON/TXT, next actions
Closed-loop agent ledger and result evaluation Internal executable Dataset scans, generated summaries, content-derived result signals, result-driven analysis concepts, pending/actioned/blocked concept lifecycle records, method-advice content, tables, figures, reports, failures, and run records /plan, /autopilot, Web plan/run/results, PlanLedger, PlanLedger.concept_memory(), PlanLedger.write_concept_book(), evaluate_project_results(...) assistant/ledger/plan_*.json, run_*.json, result_evaluation_*.json, analysis_concept_book.json, planner evidence notes, plan-level concept-memory snapshots and deltas, concept verification and blocker status, concept-memory-grounded next actions, audit-style autopilot reports
FeatureBook-based signal verification Internal executable Common liquid-biopsy signal families and generated artifacts Internal planner/evaluator, list_feature_specs(...), compile_user_analysis_idea(...) Feature contracts, compiled user idea, plan novelty evidence, report-level satisfied/pending status
Professional skill memory and workflow contracts Internal executable Markdown, text, PDFs where supported by runtime, URLs, directories, expert notes, workflow.yaml playbooks /skills, /skills explain-plan, Web skills endpoints, Python API Skill documents, skill cache, workflow triggers, required outputs, quality checks, context snippets for future planning

Preprocessing Profiles

Signal family Supported inputs Default profile Other callable profiles Main data products
cfchip_seq BED/BED.GZ, narrowPeak/broadPeak/gappedPeak, BAM/CRAM cfchip_interval_cleanup cfchip_panel_summary, cfchip_background_aware Cleaned intervals, optional region-panel summaries, optional background-normalized summaries
cfmedip_seq BED/BED.GZ, narrowPeak/broadPeak/gappedPeak, BAM/CRAM cfmedip_interval_cleanup cfmedip_panel_summary, cfmedip_scale_normalized_panel Cleaned intervals, methylation-panel summaries, optional scale-normalized summaries
medip_seq BED/BED.GZ, narrowPeak/broadPeak/gappedPeak, BAM/CRAM medip_interval_cleanup medip_panel_summary, medip_scale_normalized_panel Same route as cfMeDIP-style methylation enrichment
lpwgs BED/BED.GZ, BAM/CRAM lpwgs_interval_cleanup lpwgs_cleanup_only, lpwgs_gc_corrected Cleaned intervals, genome bins, corrected bins when annotations exist, segments, arm burden
ulpwgs BED/BED.GZ, BAM/CRAM ulpwgs_interval_cleanup ulpwgs_cleanup_only, ulpwgs_gc_corrected Same route as LPWGS, tuned for ultra-low-pass data
ctdna_variant VCF/VCF.GZ, MAF/MAF.GZ, MAF TSV, CSV/TSV/parquet tables variant_table_qc variant_strict_somatic, variant_matched_normal Normalized variant tables, conservative filters, matched-normal overlap filtering when supplied
variant VCF/VCF.GZ, MAF/MAF.GZ, MAF TSV, CSV/TSV/parquet tables variant_table_qc variant_strict_somatic, variant_matched_normal Same route as ctDNA variant preprocessing

Internal Encoders And Models

Signal family Data types Default encoder/model Optional encoders/models Status
cfchip_seq BED/BED.GZ, peak files, BAM/CRAM, bedGraph/WIG/bigWig ntv2 dnabert2, hyenadna, caduceus, epibert, epcot, enformer, coverage_profile Internal model / encoder
cfmedip_seq BED/BED.GZ, peak files, BAM/CRAM, bedGraph/WIG/bigWig epibert ntv2, dnabert2, hyenadna, caduceus, epcot, enformer, coverage_profile Internal model / encoder
medip_seq BED/BED.GZ, peak files, BAM/CRAM, bedGraph/WIG/bigWig epibert ntv2, dnabert2, hyenadna, caduceus, epcot, enformer, coverage_profile Internal model / encoder
lpwgs BED/BED.GZ, BAM/CRAM, cnv_parquet, bedGraph/WIG/bigWig lpwgs_cnv_profile coverage_profile, ntv2, dnabert2, hyenadna, caduceus, epibert, epcot, enformer Internal model / encoder
ulpwgs BED/BED.GZ, BAM/CRAM, cnv_parquet, bedGraph/WIG/bigWig lpwgs_cnv_profile coverage_profile, DNA foundation encoders listed above Internal model / encoder
ctdna_variant VCF/VCF.GZ, MAF/MAF.GZ, MAF TSV vcf_signature variant_effect_profile when CADD, SpliceAI, DeepSEA, or similar effect scores exist Internal model / encoder
variant VCF/VCF.GZ, MAF/MAF.GZ, MAF TSV vcf_signature variant_effect_profile Internal model / encoder

External Method Advisor Registry

Method/tool Status Primary data types Best used for
FinaleToolkit External runtime wrapper Paired-end cfDNA WGS BAM/CRAM or indexed fragment files Fragment length, coverage, WPS, DELFI-style features, end motifs, cleavage profiles
DELFI-style features Advisory/internal proxy Low-pass paired-end cfDNA WGS fragments Genomewide fragmentome features and cancer-monitoring style feature matrices
cfDNAPro Advisory external Paired-end cfDNA WGS BAM in R/Bioconductor workflows Robust fragment curation, fragment-length metrics, motif-oriented summaries
Griffin Advisory external cfDNA WGS around predefined loci Nucleosome profiling, tissue-of-origin and accessibility-style follow-up
LIQUORICE Advisory external cfDNA WGS BAM plus BED regions Bias-corrected region-centered coverage changes
LBFextract Advisory external cfDNA WGS BAM plus region/BED sets Regulatory-region fragmentomics and coverage/fragment-length feature extraction
cfDNApipe Advisory external cfDNA WGS/WGBS FASTQ or BAM Broad external WGS/WGBS QC, CNV, DMR, and fragment-size workflows
cfDNA UniFlow External runtime wrapper cfDNA WGS FASTQ/BAM plus workflow configuration Standardized WGS preprocessing, QC, GC-bias correction, copy-number state estimation, and region signal extraction
cfDNAFE External runtime wrapper cfDNA WGS/WGBS intermediate files or fragment files Multi-signal feature extraction across fragmentation, WPS, OCF, CNV, and methylation-derived features
cfDNAanalyzer External method guidance cfDNA sequencing BAM manifests CNA, end-motif, footprinting, nucleosome, WPS, OCF, and promoter-fragmentation-entropy feature matrices
EMIT External runtime wrapper / research cfDNA end-motif tables Transformer-based end-motif representation learning and linear-probe cancer-detection experiments
DeepFRAG External runtime wrapper / watchlist cfDNA fragment-size distributions with labels Deep fragment-size probability modeling for supervised cancer-detection experiments
ichorCNA Advisory external ULP-WGS/LPWGS cfDNA WIG/read-count bins Tumor fraction and broad copy-number alteration inference
QDNAseq Advisory external Shallow WGS / LPWGS BAM files Binning, correction, segmentation, and copy-number calling
WisecondorX Advisory external Shallow WGS / cfDNA low-pass WGS BAM/CRAM Reference-based shallow-WGS CNV detection
HMMcopy Advisory external Windowed WGS readcounts with GC/mappability tracks Readcount correction before CNV workflows
CNVkit External runtime wrapper Targeted DNA, WES, WGS BAM files Read-depth CNV detection and visualization
Control-FREEC Advisory external WGS/WES/targeted read depth plus BAF Copy-number and LOH calling in higher-coverage designs
CopywriteR Reimplementation candidate / legacy with internal copywriter-proxy Targeted or exome off-target reads, interval/bin-count tables First-pass off-target/bin-count CNV screening when the original archived R/Bioconductor stack is not suitable
FACETS / facetsSuite External runtime wrapper, core FACETS installed separately from pileup resources Paired tumor-normal WGS/WES/targeted SNP pileups Allele-specific copy number, purity, ploidy, and LOH
PureCN External runtime wrapper Targeted short-read DNA or WES BAM/coverage plus optional VCF Targeted-panel copy number, purity/ploidy, LOH, and SNV classification support
BayesCNV External runtime wrapper / watchlist Targeted cfDNA panel coverage/features Bayesian somatic amplification detection in low-tumor-content targeted cfDNA settings
QSEA Advisory external MeDIP/cfMeDIP enrichment BAM and windows Methylation-enrichment modeling using CpG density and calibration assumptions
MEDIPS Advisory external MeDIP/cfMeDIP BAM and reference CpG annotations Methylation-enrichment QC, saturation, CpG coverage, enrichment summaries
Bismark Advisory external WGBS/RRBS/bisulfite FASTQ Bisulfite alignment and methylation calling
MethylDackel Advisory external Aligned bisulfite/EM-seq BAM/CRAM CpG methylation extraction from alignments
nf-core/methylseq Advisory external Bisulfite/EM-seq FASTQ samplesheets Production-style FASTQ-to-report methylation workflow
FinaleMe Advisory external cfDNA WGS fragmentation-derived features plus FinaleMe resources Exploratory methylation prediction from cfDNA fragmentation
cfTools / cfSort Advisory external WGBS/cfMethyl-Seq methylation calls and marker references Tissue-of-origin, tumor-burden, CancerDetector, cfDeconvolve, cfSort analyses
MethylBERT External runtime wrapper / research Read-level WGBS or Dorado-called methylation patterns Transformer-based read classification and methylation deconvolution
cfDecon External runtime wrapper / research source checkout cfDNA methylation reads or method-specific feature tables Deep autoencoder-based cell-type deconvolution
CelFiE-ISH External method guidance / research Single-molecule/read-haplotype methylation data Haplotype-aware multi-cell-type deconvolution and rare-cell-type detection
CelFEER External runtime wrapper / research source checkout Read-level cfDNA WGBS methylation Benchmark-supported read-level methylation deconvolution
UXM External runtime wrapper with binary/resource gaps Fragment-level methylation states plus atlas resources Unmethylated-fragment deconvolution of cfDNA tissue fractions
MethAtlas External runtime wrapper / research source checkout Methylation ratio matrices plus atlas resources Fast interpretable atlas-based tissue deconvolution
cfNOMe External runtime wrapper / research source checkout NOMe/cfNOMe-compatible methylation outputs Tissue-of-origin deconvolution plus nucleosome occupancy summaries
MetDecode External runtime wrapper / research Marker-region methylated/total CpG counts Methylation-based cfDNA deconvolution for multi-cancer typing
CpGPT / MethylGPT / MethFormer Mixed: CpGPT/MethylGPT runtime wrappers; MethFormer model-resource guidance Methylation matrices or regional methylation tensors Foundation-model embeddings, imputation, and transfer-learning experiments
cfMethylPre Advisory literature watchlist cfDNA methylation profiles plus sequence embeddings Transfer-learning cancer-detection model concept
Dorado + modkit External runtime wrapper for modkit; Dorado remains explicit system/model resource Nanopore cfDNA POD5/FAST5 or modified-base BAM Modified-base calling and methylation coordinate extraction
fgbio Advisory external UMI-tagged targeted sequencing BAM/FASTQ UMI consensus preprocessing before low-VAF variant calling
Mutect2 / LoFreq / VarDict Advisory external Targeted, WES, or cfDNA BAM/CRAM Low-VAF ctDNA SNV/indel calling
Salmon / STAR / featureCounts Advisory external cfRNA FASTQ or count matrices cfRNA quantification and expression matrix generation
sRNAbench / miRge-style routes Advisory external Small-RNA FASTQ or miRNA count matrices EV-miRNA or plasma small-RNA profiling
Scanpy / Seurat / CellTypist-style downstream routes Advisory external CTC count tables, marker matrices, h5ad/RDS-style outputs CTC table and single-cell expression interpretation
DIA-NN / MaxQuant / OpenMS-style upstream routes Advisory external mzML/vendor raw files or protein/peptide abundance matrices Plasma or extracellular-vesicle proteomics preprocessing and matrix-level review

LLM Engines For Agent Interaction

These engines are used for natural-language understanding and user assistance. They do not replace the deterministic legality checks, method registry, or local analysis code.

Backend Selection Use
OpenAI GPT auto (currently gpt-6-luna) Default economical tier
OpenAI GPT gpt-6-sol Explicit advanced tier
OpenAI GPT gpt-6-astra Explicit flagship tier
Offline diagnostics /llm off No model calls; not conversational intelligence

The same adapter handles user messages, domain context, task and result summaries. See OpenAI configuration for the current policy.

Practical Reading Order

Current Web interaction guide

For the current interaction model and screenshots, use the workspace walkthrough, image/result controls and reviewed skills. They distinguish inspection, proposed follow-up, authorized execution and reviewed preferences.