Bioinformatics analysis service
High-dimensional omics, imaging, or clinical tables feed nested cross-validated models with holdout checks when you have them, interpretability outputs, full code, and a Methods draft.
| Category | Details |
|---|---|
| Data types supported | RNA-seq, proteomics, metabolomics, and methylation feature matrices; clinical covariates; single-cell embeddings; multi-omics matrices merged by sample ID; vendor-normalized outputs (e.g., Olink NPX, DESeq2 vst) |
| Reference builds or standards used | DOME reporting checklist [4]; TRIPOD+AI for clinical prediction models when scoped [5]; patient-level nested cross-validation; independent holdout or external cohort validation when available; GENCODE v44 / Ensembl 110 gene IDs and UniProt Swiss-Prot accessions for feature harmonization |
| Primary tools (with versions) | scikit-learn 1.6.1; XGBoost 2.1.3; scikit-survival 0.24.1; glmnet 4.1-8 (R); SHAP 0.46.0; Boruta 11.0.0; PyTorch 2.5.1 + scGPT 0.2.4 for transformer fine-tuning when scoped |
| Typical turnaround range | 3-5 weeks (single cohort, one outcome, ≤500 features, internal validation); 5-8 weeks (survival models, multi-cohort external validation, or scGPT fine-tuning). Confirmed at kickoff |
| Deliverable formats | Feature matrices; split manifests; model_performance_summary.csv; SHAP plots; nested-CV results; serialized models; PDF/SVG figures; DOME summary table; commented R/Python scripts; Methods draft |
| Regulatory/reproducibility standards followed | DOME checklist in every deliverable; TRIPOD+AI checklist when clinical prediction reporting is scoped; version-pinned conda or renv environments; logged random seeds; documented leakage audit per Bernett et al. [3]; private Git or Zenodo archival on request |
| Custom / bespoke analysis | Non-standard outcomes, custom feature engineering, domain-specific models, multi-omics fusion beyond standard signatures, client-specified validation schemes, or foundation-model adaptation (e.g., scGPT fine-tuning for perturbation or cell-state tasks) scoped at kickoff |
Machine learning (ML) learns patterns from data without hand-coded rules. Supervised learning trains on labeled outcomes (e.g., responder vs non-responder). Cross-validation repeatedly splits data to estimate generalization. Data leakage shares information between training and test sets, inflating performance [3]. AUROC (area under the receiver operating characteristic curve) measures classifier discrimination. C-index (concordance index) measures survival-model discrimination. A foundation model is pretrained on large datasets and fine-tuned for downstream tasks. A transformer is an attention-based neural architecture used in models such as scGPT [1].
Machine learning learns predictive or clustering patterns from high-dimensional biological data, often when features (genes, proteins, variants) outnumber samples. Supervised algorithms adjust parameters from labeled outcomes; unsupervised methods find structure without labels. Typical questions include which features or combinations predict an outcome, treatment response, or patient subgroup. scGPT was pretrained on more than 33 million cells and fine-tuned for annotation, integration, and perturbation prediction [1]. A 2023 systematic review screened 682 PubMed articles (2017-2023) and critically reviewed 30 cancer biomarker ML studies; the review noted reporting heterogeneity and data-leakage risks [2].
Published examples of questions machine learning can address:
Feng et al. [11] combined scRNA-seq and bulk RNA-seq to derive an 11-gene NK-cell-related signature from 77 ML algorithms, validated across TCGA, GEO, and ICGC cohorts with concordance-index benchmarking.
Mirza et al. [12] applied seven ML methods to 701 samples from 11 GEO datasets, identifying a nine-gene diagnostic signature and an eight-gene prognostic signature validated by qRT-PCR.
Liu et al. [13] profiled 47 biopsies from 36 patients before and after PD-1 therapy, linking precursor exhausted T cell expansion to response. Those states are candidate features for immunotherapy stratification models.
Cui et al. [1] showed that scGPT, a generative pretrained transformer on single-cell data, predicts perturbation responses after fine-tuning, which allows transfer learning when labeled perturbation data are limited.
This category includes biomarker discovery and predictive modeling. Each spoke page lists inputs, validation design, tools, and deliverables.
| Service | Description | Primary tools |
|---|---|---|
| Biomarker discovery | Feature selection and compact signature construction from omics or clinical matrices with stability analysis and cross-validation | glmnet 4.1-8, Boruta 11.0.0, XGBoost 2.1.3, SHAP 0.46.0 |
| Predictive modeling | Supervised classifiers and survival models with nested CV, calibration assessment, and TRIPOD+AI-aligned reporting when scoped | scikit-learn 1.6.1, scikit-survival 0.24.1, XGBoost 2.1.3 |
Each project includes validated model outputs, interpretability artifacts, version-pinned code, and a Methods draft.
Non-standard outcomes, multi-omics fusion, or foundation-model fine-tuning are scoped at kickoff.
Confirm the prediction target, inclusion criteria, endpoint type (binary, multiclass, survival), and whether an independent external cohort is available [4, 5].
Output: signed scope with primary and secondary metrics
Match sample IDs; audit missingness, duplicate patients, and batch structure.
Output: sample_manifest.csv and feature_qc_summary.csv
Apply transforms and scaling; use batch correction only when batch is not confounded with the outcome [6].
Tools used: scikit-learn ColumnTransformer; ComBat or limma when scoped
Output: analysis-ready feature matrix
Apply the seven guiding questions from Bernett et al. [3] before splitting.
Output: leakage_audit.md
Use patient-level nested cross-validation to separate tuning from performance estimation [4].
Tools used: scikit-learn CV splitters; scikit-survival for censored outcomes
Output: fold assignment table
Compare regularized linear models, tree ensembles, and survival forests with simpler baselines [7]. Fine-tune scGPT when single-cell transfer learning is scoped [1].
Output: candidate model list
Report discrimination and calibration on held-out or external data; avoid claiming clinical utility without independent validation [6].
Output: model_performance_summary.csv
Annotate selected features with pathways and cell types; document where interpretability limits causal claims [1].
Output: enrichment tables and interpretation memo
Assemble figures, models, DOME table, scripts, environment locks, README, and Methods draft.
Output: final deliverable bundle
Versions are pinned at kickoff and cited in the Methods draft. Representative tools across biomarker and prediction projects:
| Tool | Version | Role | Primary citation |
|---|---|---|---|
| scikit-learn | 1.6.1 | Classification, regression, nested CV, metrics | [8] — https://jmlr.org/papers/v12/pedregosa11a.html |
| XGBoost | 2.1.3 | Gradient-boosted classifiers and regressors | https://doi.org/10.1145/2939672.2939785 |
| scikit-survival | 0.24.1 | Random survival forests, gradient boosting survival, C-index | [9] — https://www.jmlr.org/papers/v21/20-729.html |
| glmnet | 4.1-8 | LASSO and elastic-net feature selection (R) | https://doi.org/10.18637/jss.v033.i01 |
| SHAP | 0.46.0 | Post hoc feature attribution and interpretability | [10] |
| Boruta | 11.0.0 | All-relevant feature selection against shadow features | https://doi.org/10.18637/jss.v036.i11 |
| scGPT | 0.2.4 | Transformer foundation model for single-cell fine-tuning | https://doi.org/10.1038/s41592-024-02201-0 |
| PyTorch | 2.5.1 | Deep learning backend for foundation-model adaptation | https://doi.org/10.48550/arXiv.1912.01703 |
| clusterProfiler | 4.20.0 | GO/KEGG enrichment of selected feature lists | https://doi.org/10.1089/omi.2011.0118 |
Validation reporting follows DOME [4] and TRIPOD+AI when clinical prediction reporting is scoped [5]. Feature annotation uses GENCODE v44 (human) unless a project requires a custom reference. Studies with p >> n typically require regularization and nested cross-validation rather than single train/test splits [6].
Common issues include data leakage, overfitting, incomplete reporting, unstable feature selection, and models that are hard to interpret. Pepkio documents validation design and interpretability in the deliverables.
Illicit information sharing between training and test sets is common in high-dimensional biological data and can make models fail in real application [3]. Pepkio completes a leakage audit before splitting and uses patient-level partitions throughout.
When features outnumber samples, complex models memorize noise [6]. Pepkio applies regularization and nested cross-validation, and reports simpler baselines alongside complex models so reviewers can assess incremental value.
A 2023 systematic review of 30 cancer biomarker ML studies listed data leakage and inconsistent evaluation among common limitations [2]. Pepkio includes a DOME summary table in every deliverable and a TRIPOD+AI checklist when clinical prediction reporting is scoped.
Features selected in one split may not replicate in another, which weakens biomarker claims. Pepkio reports selection frequency across cross-validation folds and prioritizes stable signatures over single-split lists.
Deep models can discriminate well but obscure biological mechanism [1]. Pepkio delivers SHAP plots and coefficient tables and documents where interpretability does not support causal inference.
Provide a sample-level feature matrix, a metadata table with the outcome and covariates, and a data dictionary. For survival endpoints, include event indicators and follow-up times. DOME recommends reporting sample-size rationale [4]. Custom formats are accepted when scoped in advance.
Standard projects typically complete in 3-5 weeks. Survival models, external validation, or scGPT fine-tuning may take 5-8 weeks. Exact timelines are confirmed at kickoff.
You receive performance tables, figures, SHAP plots, feature lists, serialized models, a DOME summary table, scripts, a README, and a Methods draft. Optional GitHub or Zenodo archival is available on request.
When scoped at kickoff, Pepkio accepts matrices from RNA-seq, proteomics, metabolomics, methylation, and clinical-laboratory tables, including vendor-normalized outputs with documented upstream processing.
Small cohorts need regularized models and nested cross-validation. Pepkio avoids claiming external validity without a holdout cohort [6]. QC flags high missingness or batch confounding before modeling.
You retain full ownership. Pepkio delivers commented R or Python scripts with conda or renv lock files and a README.
Checkpoint reviews occur after QC, after the leakage audit, and before delivery. A PhD-level bioinformatician is your primary contact.
Methods clarification and minor revisions within agreed scope (typically ≤20% of deliverables) are included. Substantial new analyses are scoped as separate milestones.
When scoped, Pepkio fine-tunes scGPT, a transformer pretrained on more than 33 million cells, for single-cell perturbation and integration tasks [1]. Regularized linear models and tree ensembles remain the default for small tabular cohorts [7].
Bespoke workflows, including custom outcomes, multi-omics fusion, domain-specific loss functions, client-specified validation schemes, or non-standard deliverables, are scoped at kickoff after a feasibility review.
Deep-dive pages for specific machine learning methods and workflows.
Tell us:
We'll tell you:
Contact us to start with a free consultation. Need everyday bench calculators? Try our free lab tools.
