Proteomics: From DIA-MS Quality Control to Biomarker Discovery

Mass spectrometry proteomics generates massive datasets with systematic technical variation, missing values, and statistical challenges that make analysis time-consuming even for experienced bioinformaticians. This case study demonstrates three progressively complex proteomics workflows using Lobster AI's proteomics agents: from cerebrospinal fluid quality control through colorectal cancer differential expression to ovarian cancer chemoresistance biomarker discovery with nested cross-validation. These workflows show how Lobster coordinates multiple specialized agents to compress workflows that typically span days or weeks into minutes of conversational queries.

Agents and Data Sources

This analysis uses the lobster-proteomics package, which provides three agents working in a parent-child hierarchy:

All datasets in this case study are synthetic proteomics data (OC_XXXX protein IDs, patient identifiers, and clinical metadata) designed for controlled demonstration of analytical workflows. External APIs queried include Enrichr (GO/Reactome pathway enrichment) and STRING (protein-protein interaction networks). Local computation uses scipy (Welch t-test), scikit-learn (LASSO, cross-validation), and lifelines (survival analysis).

Simple: Alzheimer's CSF Proteomics QC Pipeline

The first workflow demonstrates core proteomics QC and preprocessing: import, quality assessment, filtering, normalization, and unsupervised clustering in two conversational turns.

The Research Question

How do we quality-assess and preprocess a cerebrospinal fluid proteomics dataset for downstream Alzheimer's disease biomarker analysis?

CSF proteomics in Alzheimer's disease is a high-impact clinical use case. CSF biomarkers like A-beta, tau, and neurofilament light are central to AD diagnosis, and untargeted proteomics is expanding the biomarker landscape. This 12-sample pilot study represents a realistic entry point for validating a new cohort before scaling to larger studies.

Turn 1: Import and Quality Assessment

The first query imports the DIA-MS dataset and performs comprehensive quality assessment.

lobster query --session-id alzheimers_csf \
  "Import the DIA-NN proteomics dataset from alzheimer_csf_proteomics.tsv \
   with the sample metadata from alzheimer_csf_metadata.csv. This is a \
   cerebrospinal fluid (CSF) proteomics study comparing Alzheimer's disease \
   patients versus healthy controls. After importing, assess the data quality \
   and give me a comprehensive summary of the dataset including missing value \
   patterns, intensity distributions, and sample metadata."

The proteomics_expert imported 800 proteins across 12 CSF samples and immediately characterized the dataset.

The agent flagged the small sample size as a limitation while noting the excellent data completeness (88.1%) and rich clinical metadata including APOE genotype — the strongest genetic risk factor for Alzheimer's disease.

Turn 2: Preprocessing and Pattern Analysis

The second query runs the complete preprocessing pipeline with dimensionality reduction and clustering.

lobster query --session-id alzheimers_csf \
  "Now preprocess the Alzheimer CSF proteomics data: (1) filter out proteins \
   with more than 50% missing values and samples with more than 40% missing, \
   (2) normalize using median normalization with log2 transformation, (3) run \
   PCA and Leiden clustering to identify sample grouping patterns. Report on \
   whether the biological groups (Alzheimer vs Control) separate in the PCA \
   and if batch effects are visible."

The agent retained all 12 samples and 732 of 800 proteins after quality filtering.

Median normalization with log2 transformation was applied — the standard approach for mass spectrometry data that preserves the MNAR missing value structure. PCA with Leiden clustering identified two clusters, which the agent noted should be visualized to determine whether they correspond to biological groups (AD vs Control) or technical batches.

Medium: Colorectal Cancer Tumor Proteomics DE Analysis

The second workflow exercises the full parent-child agent coordination: proteomics_expert handles preprocessing, then hands off to proteomics_de_analysis_expert for differential expression, pathway enrichment, and STRING network analysis.

The Research Question

What are the proteomic signatures of colorectal cancer tumor tissue compared to matched normal tissue, and what protein interaction networks drive tumor biology?

Colorectal cancer is the third most common cancer worldwide, and tumor-vs-normal tissue proteomics is a standard study design for biomarker discovery and pathway characterization. This dataset includes KRAS mutation stratification — a key oncogenic driver in CRC — making it relevant for precision medicine. A 54-sample paired cohort provides sufficient statistical power for robust differential expression testing.

Turn 1: Import and Quality Assessment

The first query imports the tumor-normal paired dataset with clinical metadata.

lobster query --session-id crc_proteomics \
  "Import the proteomics data from crc_proteomics_maxquant.tsv with sample \
   metadata from crc_proteomics_metadata.csv. This is a colorectal cancer \
   tumor vs matched normal tissue proteomics study. After importing, assess \
   the data quality comprehensively and give me a summary."

The proteomics_expert imported 4,500 proteins across 54 samples (27 tumor + 27 matched normal).

The agent identified the MNAR missing value pattern, verified metadata integration including KRAS mutation status and TNM/AJCC staging, and recommended a standard preprocessing pipeline before differential expression testing.

Turn 2: Differential Expression and Pathway Enrichment

The second query runs preprocessing, differential expression, and pathway enrichment in a single turn.

lobster query --session-id crc_proteomics \
  "Preprocess the CRC proteomics data: (1) filter proteins with more than 50% \
   missing values and remove any low-quality samples, (2) apply median \
   normalization with log2 transformation, (3) run PCA with Leiden clustering \
   to assess tumor/normal separation and batch effects. Then run differential \
   expression analysis comparing Tumor vs Normal using the condition column, \
   and perform GO and Reactome pathway enrichment on the significant DE proteins."

The proteomics_expert preprocessed the data, then delegated to proteomics_de_analysis_expert for differential expression and enrichment.

Differential expression testing identified 271 significantly dysregulated proteins (6.6% of the proteome) at FDR < 0.05. Pathway enrichment returned 303 associations but none survived FDR correction — a common pattern in heterogeneous tumor biology where DE proteins span diverse biological processes rather than concentrating in specific pathways. The agent noted this and recommended directional pathway analysis (separate up/down-regulated sets) as a next step.

The 27 tumor and 27 matched normal samples represent paired tissue from the same patients. Welch's t-test treats these as independent groups; a paired t-test or limma with a blocking factor would better exploit the paired structure and improve statistical power in a production analysis.

Turn 3: STRING Network Analysis and Age Correlation

The third query runs protein-protein interaction network analysis to identify functional modules.

lobster query --session-id crc_proteomics \
  "Run STRING protein-protein interaction network analysis on the \
   differentially expressed proteins from the CRC tumor vs normal comparison. \
   Use the functional network type with a confidence score threshold of 700 \
   (high confidence). Identify the top hub proteins and report the network \
   topology. Also, run a correlation analysis between the DE protein levels \
   and patient age to identify age-associated proteomic changes."

STRING network analysis at high confidence (score > 700) revealed three functional modules among the 271 DE proteins.

The PKM-PGK1 glycolytic axis confirms the Warburg effect in CRC — a metabolic reprogramming where cancer cells favor glycolysis even in the presence of oxygen. The RUNX3-CBFB tumor suppressor complex is known to be disrupted in colorectal cancer, and its downregulation in tumors aligns with published literature.

The CRC synthetic dataset used protein identifiers that mapped to real UniProt accessions, enabling genuine STRING interaction queries. However, the starting DE protein list was generated from synthetic expression data, so the network modules should be interpreted as illustrative of the type of output STRING analysis produces rather than novel biological findings.

Hard: Ovarian Cancer Chemoresistance Biomarker Discovery

The third workflow exercises all three proteomics agents: proteomics_expert for preprocessing, proteomics_de_analysis_expert for differential expression, and biomarker_discovery_expert for LASSO/stability-based panel selection with rigorous nested cross-validation.

The Research Question

Can we identify a proteomics biomarker panel that predicts platinum chemoresistance in high-grade serous ovarian cancer before treatment initiation?

High-grade serous ovarian cancer has the highest mortality of gynecological cancers, largely because most patients develop platinum chemoresistance. Predicting resistance before treatment would transform patient care — allowing resistant patients to receive alternative regimens upfront instead of failing first-line therapy. This 80-patient clinical cohort with PFS (progression-free survival) endpoints, BRCA mutation status, and FIGO staging represents a realistic translational proteomics study.

Turn 1: Import, QC, and Preprocessing

The first query imports the clinical dataset and runs the complete preprocessing pipeline.

lobster query --session-id hgsoc_biomarker \
  "Import clinical proteomics data from a high-grade serous ovarian cancer \
   (HGSOC) chemoresistance study. The data file is at hgsoc_clinical_proteomics.tsv \
   with clinical metadata at hgsoc_clinical_metadata.csv. This is an 80-patient \
   cohort with 40 chemo-sensitive and 40 chemo-resistant HGSOC patients profiled \
   by mass spectrometry. Metadata includes PFS (progression-free survival) in days, \
   event status, FIGO stage, residual disease, and BRCA mutation status. After \
   importing: (1) assess data quality, (2) filter proteins with >60% missing and \
   samples with >50% missing, (3) normalize with median normalization and log2 \
   transform. Report the dataset overview and any quality concerns."

The proteomics_expert imported 3,200 proteins across 80 patients (40 chemo-sensitive, 40 chemo-resistant).

The agent identified excellent data quality (19.3% missing, well below the typical 30-70% range for MS data) and a perfectly balanced cohort design (40 vs 40). All samples and proteins passed QC thresholds.

Turn 2: Differential Expression Analysis

The second query runs differential expression and pathway enrichment comparing resistant vs sensitive patients.

lobster query --session-id hgsoc_biomarker \
  "Run differential expression analysis comparing Chemo_Resistant vs \
   Chemo_Sensitive groups using the condition column. Use Welch's t-test with \
   FDR correction. Then run GO and Reactome pathway enrichment on the \
   significant DE proteins. Also run STRING network analysis on the DE proteins \
   to identify protein interaction hubs and functional modules."

The proteomics_de_analysis_expert identified 217 significantly dysregulated proteins (7.0% of the proteome) at FDR < 0.05.

Turn 3: Biomarker Panel Selection and Nested Cross-Validation

The third query exercises the biomarker_discovery_expert for feature selection and rigorous validation.

lobster query --session-id hgsoc_biomarker \
  "Select a chemoresistance biomarker panel using LASSO and stability selection \
   methods with the condition column as target, n_features=20, n_iterations=100. \
   Then evaluate that panel using nested cross-validation with 5 outer folds \
   and 3 inner folds using logistic regression."

The biomarker_discovery_expert selected a 20-protein panel using LASSO regularization with bootstrap stability selection (100 iterations, >50% appearance threshold).

The panel achieved perfect classification performance (AUC 1.000) under rigorous nested cross-validation — a methodology that prevents data leakage by fitting scalers and hyperparameters only on training data within each fold. The agent identified 14 upregulated and 6 downregulated proteins as the resistance signature.

What This Demonstrates

Multi-Agent Coordination

No single agent could produce these analyses. The proteomics_expert handled data import, quality assessment, filtering, and normalization across all three workflows. The proteomics_de_analysis_expert ran differential expression testing with Welch t-test and FDR correction, queried Enrichr for pathway enrichment, and interfaced with the STRING REST API for protein interaction networks. The biomarker_discovery_expert implemented LASSO regularization with bootstrap stability selection and nested cross-validation with proper scaling and hyperparameter tuning within folds. The supervisor routed each sub-question to the appropriate specialist and synthesized results across all turns.

Database Integration

The agents queried Enrichr (GO/Reactome pathways) and STRING (protein-protein interactions) programmatically through validated API tools — not through LLM approximation. Statistical testing (Welch t-test, FDR correction) and machine learning (LASSO, logistic regression, nested CV) ran locally via scipy and scikit-learn with exact reproducibility.

Rigorous Methodology

The hard workflow demonstrates publication-quality methodology:

• Bootstrap stability selection (100 iterations) ensures robust feature selection not driven by random split artifacts
• Nested cross-validation (5 outer folds x 3 inner folds) prevents data leakage by tuning hyperparameters only on training data
• KNN imputation (k=5) preserves biological structure in missing values
• L2-regularized logistic regression prevents overfitting on high-dimensional data

These methodological details are often mishandled in manual analyses, leading to inflated performance estimates that fail external validation.

KNN imputation assumes missing-at-random patterns. After abundance-based filtering removed the lowest-intensity proteins (which are predominantly MNAR), the remaining missing values in the filtered dataset are more compatible with MAR assumptions, making KNN appropriate for this reduced feature set.

Human vs Raw LLM vs Lobster AI

Estimates based on these three case study sessions. Human researcher timing assumes manual workflows without automation.

Raw LLMs cannot load files, compute statistics, query APIs, or execute machine learning. They may suggest methods or approximate results, but cannot produce validated outputs. Lobster AI compresses weeks of manual proteomics work into minutes of conversational queries with full provenance tracking.

Limitations

• Synthetic data. All three datasets use synthetic protein expression matrices. While protein identifiers map to real UniProt accessions (enabling genuine pathway and network queries), the expression values and group separations are engineered. Results demonstrate methodology, not biological discovery.
• AUC 1.000 is a synthetic data artifact. Perfect classification does not occur in real clinical proteomics. Published biomarker panels for ovarian cancer achieve discovery AUCs of 0.72-0.88.
• Paired design analyzed with unpaired test. The CRC dataset contains matched tumor-normal pairs, but Welch's t-test was applied without accounting for the pairing structure. A paired test would be more appropriate.
• Missing value strategy. MNAR was correctly identified as the dominant missing mechanism, but KNN imputation (an MAR method) was applied to the filtered dataset. This is acceptable after abundance-based filtering but represents a methodological simplification.
• Stability scores at 100%. The top 8 biomarker candidates all achieved 100% stability across bootstrap rounds, reflecting strong synthetic signal. Real clinical data typically produces stability scores of 40-70% for the top features.
• No visual outputs. Volcano plots, heatmaps, and PCA score plots would be generated in a full analysis session but are not shown here.

Reproducibility

To reproduce these analyses, install the proteomics package and run the queries sequentially:

pip install 'lobster-ai[full]==1.0.12'

Simple Workflow (CSF QC)

lobster query --session-id alzheimers_csf \
  "Import the DIA-NN proteomics dataset from alzheimer_csf_proteomics.tsv \
   with the sample metadata from alzheimer_csf_metadata.csv. This is a \
   cerebrospinal fluid (CSF) proteomics study comparing Alzheimer's disease \
   patients versus healthy controls. After importing, assess the data quality \
   and give me a comprehensive summary."

lobster query --session-id alzheimers_csf \
  "Preprocess the data: filter proteins with >50% missing and samples with \
   >40% missing, normalize with median + log2, run PCA and Leiden clustering."

Medium Workflow (CRC DE Analysis)

lobster query --session-id crc_proteomics \
  "Import crc_proteomics_maxquant.tsv with metadata from \
   crc_proteomics_metadata.csv. Assess data quality comprehensively."

lobster query --session-id crc_proteomics \
  "Preprocess, then run differential expression comparing Tumor vs Normal, \
   and perform GO and Reactome pathway enrichment."

lobster query --session-id crc_proteomics \
  "Run STRING network analysis on DE proteins (confidence > 700) and age \
   correlation analysis."

Hard Workflow (HGSOC Biomarker Discovery)

lobster query --session-id hgsoc_biomarker \
  "Import hgsoc_clinical_proteomics.tsv with metadata from \
   hgsoc_clinical_metadata.csv. Filter proteins >60% missing, samples >50% \
   missing, normalize with median + log2."

lobster query --session-id hgsoc_biomarker \
  "Run differential expression comparing Chemo_Resistant vs Chemo_Sensitive. \
   Run GO/Reactome pathway enrichment and STRING network analysis."

lobster query --session-id hgsoc_biomarker \
  "Select chemoresistance biomarker panel using LASSO + stability selection \
   (n_features=20, n_iterations=100). Evaluate with nested cross-validation \
   (5 outer folds, 3 inner folds) using logistic regression."

Session continuity via --session-id ensures each turn builds on prior context. Results are stored in the .lobster_workspace/ directory and can be exported with /pipeline export.