Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one

Home›Services›Data, analytics & AI›Bioinformatics Service

Omics and NGS Data Analysis

Bioinformatics Services

Pubrica’s computational biologists, statisticians and domain specialists provide bioinformatics and omics data analysis for academic researchers, pharmaceutical and biotechnology companies, clinical and diagnostic teams, and public research institutes. Each project delivers a written analysis plan, the analysed results, publication-ready figures, the code and environment, and a methods section, with the reference build, annotation release and unit of replication recorded so the analysis can be checked at review.

  • A written analysis plan naming the reference build, the annotation version and the unit of replication
  • End-to-end omics analysis: bulk and single-cell RNA-seq, WGS and WES, methylation, proteomics and metabolomics
  • Multi-omics integration and systems biology, with batch effects modelled in the design
  • Machine learning with independent validation, using held-out data and reporting only the performance the validation supports
  • Code, environment and the methods section, so the analysis can be run again
Sample analysis summary · bulk RNA-seq 23 of 24 samples · DESeq2

Reference build and annotation release recordedGRCh38 · v44

Exclusions decided on QC, before groups were looked at1 sample

Batch found in the PCA and put in the design2 runs

Enrichment against the genes actually tested14,382

Code, environment and raw data depositedAccession issued

Reads → QC → quantify → model → interpret → deposit

Included in every analysis
  • Versions and parameters recorded
  • QC before groups are seen
  • Code and environment shipped
  • Methods section written
  • Data deposit-ready
1,200+Omics and computational biology projects
Since 2009Supporting researchers and industry worldwide
FAIR · MINSEQEGATK Best Practices, HIPAA, GDPR, GCP
ISO 9001:2015Quality management certified · cert. H3CR9TCU20251212INDQM10N

Service Scope

Six Bioinformatics Service Areas, Available Individually

Our bioinformatics services are offered individually because projects reach us at different stages, from a study design question to a completed analysis that a reviewer has queried.

Design and analysis planning We define what the experiment needs to answer the question: the number of replicates, sequencing depth, batching and the covariates to record. We then write the analysis plan before the data exists, specifying the pipeline, tools, thresholds, exclusion rules and models. This is the least expensive stage and the only one at which the design can still be changed.
End-to-end omics data analysis The full route from raw files to interpreted results for genomic, transcriptomic, single-cell, proteomic, metabolomic, epigenomic and microbiome data — next-generation sequencing analysis including whole-genome, whole-exome and targeted sequencing, variant calling and annotation, differential expression, and the statistical models that go with each.
Multi-omics integration and systems biology Genomics, transcriptomics, proteomics, metabolomics and epigenomics data integrated on the same samples, with the integration method chosen for what the datasets can support. Includes network and pathway analysis, biomarker discovery and a clear statement of which data layer carries the signal.
Machine learning and predictive modelling Classification, predictive modelling and pattern recognition in high-dimensional omics data, built with the split made before anything is looked at and validated on data the model has not seen. Reported to TRIPOD+AI, with calibration shown as well as discrimination — because an AUC on its own does not tell a clinician whether a probability means anything.
Structural bioinformatics and drug discovery support Protein structure prediction, molecular docking and molecular dynamics for structure-function relationships, protein-ligand interaction and rational drug design, alongside target identification and validation. Predicted structures are used with their confidence scores attached, because a low-confidence region is not a structural finding.
Pipelines, reanalysis and rescue Reproducible workflows built in Nextflow or Snakemake with containers, deployed locally or in the cloud, for groups who will run the analysis repeatedly. We also reconstruct what a previous analysis did, assess whether a reviewer’s objection is valid, and rerun the analysis correctly where it is.

Visualisation and publication support

Publication-ready figures are part of every project rather than an addition to it: heatmaps, volcano plots, principal component plots, phylogenetic trees, gene expression profiles, pathway enrichment charts, survival curves and interactive dashboards where a reader benefits from one. Each comes with the code that produced it, in vector format at the resolution the journal asks for, with colour chosen so the figure survives greyscale printing and common forms of colour blindness. Where the journal comes back asking for a panel to be rebuilt, rebuilding it is part of the work.

Types of Bioinformatics Analysis

Fifteen Bioinformatics Analyses and Their Applications

Analyses across pre-clinical, translational and clinical research, each with its own pipeline, statistical methods and common sources of error.

Genomic data analysis and NGSWhole-genome, whole-exome and targeted sequencing for genetic variation, disease predisposition and pharmacogenomic response — with the reference build and annotation release recorded, because coordinates from different builds silently disagree.
Variant calling and interpretationLarge-scale discovery from SNP arrays, WGS and WES: joint calling across the cohort, filtering, annotation against versioned databases, and prioritisation under an explicit framework rather than by plausibility.
Transcriptomic analysisRNA-seq for expression differences between conditions, for which isoform is being made, and for non-coding species, modelled on counts with the design in the formula — batch, sex and site included rather than corrected for afterwards.
Single-cell analysisSingle-cell RNA sequencing for cell type, state and heterogeneity, with doublets and ambient RNA removed, integration checked for over-correction, and condition testing done on pseudobulk at the level the study randomised.
Spatial transcriptomicsExpression with its location kept, for tissue architecture and for the neighbourhoods cells sit in — with spot deconvolution handled explicitly, since a spot is usually several cells rather than one.
Proteomic data analysisProtein identification, quantification and post-translational modification from data-dependent and data-independent acquisition, with false discovery controlled at the protein level rather than the spectrum level and missing values treated as the scientific choice they are.
MetabolomicsTargeted and untargeted metabolite profiling: feature detection, annotation with confidence levels stated, and pathway interpretation that connects genetic, environmental and microbiome influences to biomarkers and response.
Microbiome and metagenomics16S amplicon and shotgun sequencing for microbial diversity and host-microbiota interaction, analysed compositionally, with negative controls sequenced and reported because low-biomass studies pick up their reagents.
EpigenomicsDNA methylation profiling by array or bisulfite sequencing, histone modification and ChIP-seq, and chromatin accessibility by ATAC-seq — with blacklisted regions removed, replicates assessed before pooling, and cell composition accounted for in methylation work.
Integrative multi-omicsGenomics, transcriptomics, proteomics, metabolomics and epigenomics combined on shared samples to find mechanism and biomarkers, with the contribution of each layer reported rather than absorbed into one figure.
Structural bioinformaticsProtein structure prediction, molecular docking and molecular dynamics for structure-function relationships, protein-ligand interaction and rational drug design, with prediction confidence carried through to the conclusion.
Machine learning and AIBiomarker classification, predictive modelling and pattern recognition in complex omics data, with the data split before inspection, validation on held-out or external data, and reporting to TRIPOD+AI.
Clinical and translational bioinformaticsDiagnostics, pharmacogenomics, patient stratification and clinical decision support, run under the governance clinical data requires and reported so a clinical team can act on it.
Population and comparative genomicsPopulation structure, phylogenetics, selection and association studies at cohort scale, with ancestry handled explicitly rather than assumed, which also prevents a frequency database from misclassifying a variant as rare.
Pipeline and software developmentReproducible, containerised workflows in Nextflow or Snakemake, deployed locally or in the cloud, with the documentation and tests your own team needs to run and maintain them after the project ends.

Research Areas We Cover

Analyses are carried out by specialists who know the relevant biology, across oncology and tumour genomics, immunology and immuno-oncology, infectious disease and vaccines, the gastrointestinal tract and the microbiome, cardiovascular and metabolic disease, neurology and neurodegeneration, rare and inherited disease, reproductive biology, pharmacology and toxicology, and preclinical and translational drug discovery. Outside human medicine, the same methods are applied in plant and crop genomics, animal and veterinary science, agricultural and soil microbiome research, food, nutrition and nutrigenomics, environmental and forensic science, and industrial biotechnology.

Who we serve

Researchers and Organisations We Support with Bioinformatics

We support six types of client across academia, industry, healthcare, agriculture and the public sector, each needing an analysis that will withstand independent checking.

Academic researchers and universitiesGenomics, transcriptomics, proteomics and metabolomics analysis for theses, dissertations and publications — with the code and the deposit package that a journal and an examiner will both ask for.
Pharmaceutical and biotechnology companiesFinding and validating targets, identifying biomarkers, repurposing existing compounds, analysing trial data and supporting precision medicine programmes — run under the governance and documentation a regulated programme needs.
Healthcare, hospitals and clinical research organisationsPatient stratification, genetic diagnostics, biomarker validation and clinical sequencing interpretation — with variant classification under an explicit framework rather than by plausibility, because a clinical review board will ask for the evidence codes.
Diagnostics and medical device companiesAssay development, biomarker validation, companion diagnostics and NGS-driven diagnostic pipelines, with the performance evidence written so it can go into a clinical evaluation or a submission without being rebuilt.
Government and nonprofit research institutesGenomic projects at national scale, the genetics of whole populations, epidemiology, public health genomics and disease surveillance, at a scale where the pipeline has to be maintainable by your own team afterwards.
Agriculture, food and life science startupsPlant and crop genomics, trait and transcriptomics analysis, soil and gut microbiome work, nutrigenomics and probiotic development — and for startups, end-to-end computational support without the cost of building a team first.

When to Involve a Bioinformatician

The most useful point is before the samples are sequenced, while replicate number, depth, batching and recorded covariates can still be decided; this consultation is short and is the only stage at which the design can change. The second is when the data arrives and the analysis must be run, which is when most projects begin. The third is after a reviewer has queried the analysis, when the objection can often be answered with a sensitivity analysis rather than a full rerun, and we can usually establish which within a week. A design in which batch is fully confounded with the comparison cannot be corrected afterwards, because no statistical method can separate effects that the experiment never separated.

Compliance and Standards

Bioinformatics Reporting and Data Deposition Standards

Pubrica works in compliance with this guidance. In computational work, the standards mainly concern disclosure: what was recorded, what was deposited and whether another analyst could rerun the analysis and obtain the same results.

FAIRfindable · reusable

FAIR data principles (findable, accessible, interoperable, reusable)

Data that can be found, reached, read by other software and used again — applied to the raw datasets, to the processed outputs and to the code. In practice this means persistent identifiers, machine-readable metadata, a licence, and file formats somebody else’s software can open. Funders increasingly require a data management plan written to these principles at application, so the decision about what can be shared is made before the data exists rather than at submission.

MINSEQEMIAME · MIxS

Minimum information reporting standards (MINSEQE, MIAME, MIxS, MIAPE)

MIAME for microarray experiments and MINSEQE for high-throughput sequencing set out what has to accompany deposited data: the experimental design, the sample annotation, the raw and processed values and the protocols. MIxS does the same for sequences from environmental and microbiome work, and MIAPE for proteomics. Repositories enforce these at deposit, and a study that has not collected the metadata as it went has to reconstruct it afterwards.

GATKbest practices

Variant discovery best practice (GATK Best Practices and equivalents)

The documented workflows for germline and somatic short variant discovery, from duplicate marking and base quality recalibration through joint calling to filtering. Following them is not a formality: joint calling across a cohort is what makes a reference call distinguishable from a no-call, and skipping it turns poorly covered regions into apparent shared variants. Equivalent documented practice applies where another caller is used.

TRIPOD+AI · STARD

Reporting guidelines for models and markers (TRIPOD+AI, STARD, REMARK)

Where the deliverable is a predictive model rather than a gene list, TRIPOD+AI governs how it is reported — the data source, the predictors, how missing data was handled, how the model was validated and on what. STARD applies where the claim is diagnostic accuracy, and REMARK to prognostic biomarker studies. These are the guidelines journals apply to machine learning in biomedicine, and they are what a methods reviewer will have open.

GDPRHIPAA · ICMJE

Privacy, governance and publication ethics (GDPR, HIPAA, NIH GDS, ICMJE, COPE)

Human genomic data is identifying and cannot be fully anonymised, which is why it goes to controlled-access archives under a governance framework rather than into a supplementary file. GDPR and HIPAA govern the handling, the NIH Genomic Data Sharing Policy governs deposit for work it funds, and ICMJE and COPE govern authorship and the data availability statement — which has to describe what the consent actually permits.

Reproducibility and Version Records

Every analysis is delivered with the version of every tool, the reference build and annotation release, the exact parameters, the random seeds and the environment definition, supplied as a container or lock file rather than a list of package names. Thresholds and exclusion rules are written down before the data is grouped, and every exclusion is reported with its reason. This allows a reviewer, a collaborator or your own team to rerun the analysis later and obtain the same numbers. Where work arrives from a previous provider without these records, reconstructing what was run is usually the first deliverable.

Responsible Use of AI

Reproducible Pipelines with Recorded Versions, Parameters and Reference Builds

In bioinformatics, automated pipelines are the core of the work. The relevant question is therefore whether the versions, parameters and reference builds were recorded well enough for another analyst to obtain the same result.

What we use it forDrafting and scaffolding analysis code, read before it runs. Parameter sweeps so the sensitivity of a result is visible rather than assumed. Literature and annotation lookups a person then verifies.
What we do not use it forBiological interpretation; choosing the reference genome, the annotation release or the significance threshold; accepting enrichment results without a human reading them against the biology; and anything in a submission without a named analyst and a versioned, re-runnable pipeline behind it.
What goes in your declarationThe tool, the stage it was used at, and the human oversight applied — because journals, funders, regulators and the ICMJE January 2026 update all now expect it, and an undeclared tool found later is treated as an integrity matter rather than a style preference.

Why We Do Not Publish AI Accuracy Figures

Accuracy figures for AI tools are usually reported by the companies that sell them.

Pubrica therefore quotes no self-generated accuracy figure, because it would carry the same conflict of interest. We will tell you which tool was used on your project, at which stage, and what a specialist changed afterwards.

Methodology

Bioinformatics Analysis Methodology

The analysis runs in eight stages, each led by the appropriate Pubrica specialist: a computational biologist, a statistician, a domain specialist, and a project coordinator responsible for the deadline.

Stages one and two

Research Question and Analysis Plan

A computational biologist and a statistician. We establish what the analysis must answer and what the design can and cannot support before any data is processed. We then write the plan, specifying the pipeline, tools, thresholds, exclusion rules and models, while the groups are still blinded, so that the result is pre-specified.

Stages three and four

Data collection, annotation and QC

Computational biologists. Datasets assembled, structured and annotated so they are FAIR-compliant from the start, with sample metadata reconciled against the sample sheet rather than trusted. Then quality control per sample against the thresholds already agreed, with identity, sex and relatedness checked wherever the assay allows it, because sample swaps invalidate a study rather than a sample.

Stage five

Pipeline construction and custom development

Pipeline engineers. An existing workflow is adapted, or a new one written, in Nextflow or Snakemake with containers and a pinned environment rather than a folder of scripts. Established community pipelines are used and cited where they fit, and custom code is written only for the parts of the question that require it.

Stage six

Analysis, modelling and validation

Statisticians and computational biologists. The analysis is run as planned, with the statistical model matching the design and the unit of replication matching how the experiment was randomised. Models are tested for stability and, where the deliverable is predictive, validated on data not used in training.

Stage seven

Interpretation, visualisation and the methods section

Domain specialists. Results placed in the biology, enrichment run against the correct background, and figures built to journal specification — heatmaps, volcano plots, PCA, phylogenies, pathway and expression plots, each with the code that made it. The methods section is written here rather than months later, while the parameters are still known.

Stage eight

Deposit, handover and post-analysis support

The project coordinator. Data prepared for deposit at GEO, SRA, ENA, PRIDE, EGA or dbGaP with the metadata each requires, and the accession issued before submission. Then the code, the environment and the documentation handed over, and continuing support as reviewers come back, as new samples arrive, or as a tool that produced your figures is superseded.

Request a Free Feasibility Review of Your Study Design

Before quoting, a computational biologist reviews your design and data and tells you what the analysis can support, what it cannot, and which pipeline decision matters most for your question. The review takes two business days and is free. Occasionally it shows that the design cannot answer the question with the available samples, which is better established before further costs are incurred.

What you provide
The questionWhat the analysis has to establish
The assayPlatform, chemistry, read length, depth
The designGroups, replicates, batches, covariates
The dataRaw files, or where they sit
The metadataSample sheet, with everything recorded
Any prior workEarlier analyses, code, or a reviewer’s comments

The sample sheet is essential. Records of which samples were extracted, run and processed together are what distinguish a batch effect from a finding. Please send the sample sheet, even if it exists only as a photograph of a lab notebook.

Bioinformatics Pipelines by Assay

Bioinformatics Pipelines by Assay Type and Common Pitfalls

Each stage of the analysis, with the tools used, the output file, the quality gate it must pass and the decision that most often invalidates the result. In omics, the most damaging errors are usually reasonable-looking choices made in the wrong order rather than software bugs.

Choose an assay on the left. Its pipeline is then displayed. 6 stages · 27 tools named

Choose the assay

Methylation arrays and bisulfite sequencing, shotgun metagenomics, metabolomics, spatial transcriptomics, long-read assembly and multi-omic integration follow the same shape and are quoted the same way.

Analysis Pipeline for This Assay

Bulk RNA-seq, differential expression

Which genes differ between conditions, and by how much — with the variation between people separated from the variation between groups.

1

Raw reads and quality control

FASTQ

Per-base quality, adapter content, duplication and over-represented sequences assessed for every sample, then adapters and low-quality tails trimmed. Library size, RNA integrity and 3′ bias are recorded per sample rather than averaged.

FastQCMultiQCfastpTrim Galore
GateEvery sample passes the same QC thresholds, and the thresholds are written down before the data arrives.
Where it goes wrongDropping a sample after seeing which group it belongs to. Exclusions are decided on QC metrics alone, and every exclusion is reported with its reason.
2

Alignment or pseudoalignment

BAM / abundance

Reads aligned to the reference genome with a splice-aware aligner, or quantified directly against the transcriptome. Which route is right depends on whether novel isoforms matter: pseudoalignment is far faster and cannot find what is not in the index.

STARHISAT2SalmonKallistoTopHatCufflinks
GateAlignment rate, duplication and gene body coverage reported per sample, with the reference genome build and annotation version stated.
Where it goes wrongNot stating the genome build and annotation release. Results computed against GENCODE v38 and v44 differ, and a reviewer who cannot reproduce the gene count is likely to question the analysis.
3

Counting and filtering

count matrix

Reads are summarised to genes and low-count genes are removed, because testing 60,000 features when only 15,000 are measurable reduces power at the multiple-testing correction.

featureCountsHTSeqtximportedgeR::filterByExpr
GateThe filter is defined by expression and sample count, applied before any group is looked at, and reported.
Where it goes wrongFiltering on a statistic that depends on the outcome. Removing genes by variance across all samples, or by significance, biases everything downstream and is invisible in the final table.
4

Normalisation and the model

results table

Counts normalised for library composition, then modelled on the negative binomial with the experimental design in the formula. Batch, sex, site or extraction date belong in the design rather than in a separate correction, wherever the design allows it.

DESeq2edgeRlimma-voomsva/ComBat
GatePCA and sample-to-sample distances are inspected before modelling so that any batch effect is identified before review.
Where it goes wrongUsing FPKM or TPM for differential testing. They are for comparing genes within a sample, not samples within a gene, and a test built on them has no valid dispersion estimate.
5

Multiple testing and effect size

DE gene list

P-values adjusted across all tested genes, and the gene list defined by an adjusted threshold together with a fold-change threshold that means something biologically. Shrunken fold changes are used for ranking, because raw ratios at low counts are mostly noise.

Benjamini-HochbergIHWapeglmashr
GateThe adjusted threshold, the fold-change threshold and the number tested are all stated together.
Where it goes wrongReporting nominal p-values as though they were adjusted, or picking the threshold after seeing how many genes it returns.
6

Interpretation and deposit

GEO / ArrayExpress

Enrichment run against a background of the genes actually tested rather than the whole genome, results visualised, and raw data deposited with the metadata a reader needs to reuse it. The accession goes in the manuscript before it is submitted.

clusterProfilerfgseaGSEAEnrichrGEO
GateDeposited to MINSEQE-level metadata, with the processed matrix alongside the raw reads.
Where it goes wrongEnrichment against the whole genome as background. It makes almost any list look enriched for whatever is highly expressed in that tissue, and it is the single most common analysis error in published transcriptomics.

Deliverables

What you receive. The count matrix and the normalised matrix, the full results table for every gene tested rather than the significant ones only, the code and the environment that produced them, per-sample QC, publication-ready figures, and a methods section written with the versions, the parameters and the thresholds already in it.

Struck-through tools have been declared superseded by their developers. They are shown because they remain in use and still appear in submitted methods sections. Where one of them produced your existing results, we will tell you whether rerunning changes the conclusion.

These pipelines reflect how analyses are run in practice, and the listed errors are ones we find in submitted analyses. The right tools for your project depend on the data, the question and the expectations of your field. Where more than one approach is defensible, the choice is made with you and recorded. Tool versions change quickly, and we tell you when a pipeline is no longer current.

Common Errors in Omics Analysis

The three most consequential errors in published omics are common ones. The first is running enrichment against the whole genome rather than the genes tested, which makes almost any list appear significant for whatever is abundant in the tissue. The second is testing single-cell differential expression across cells as though each were a replicate when the experiment randomised people; with forty thousand cells from six donors, nearly every gene reaches significance. The third is testing microbiome proportions with methods that ignore the compositional nature of the data, so that one taxon rising forces the others down. All three regularly pass peer review, and all three can be avoided only when the analysis is designed.

Our experts

Your Bioinformatics Team

A named team of subject specialists stays with your project from scoping to delivery, and your project coordinator introduces each member before work begins.

Subject matter expertClinical or scientific expert matched to your topic

Reviews the scientific content for accuracy and confirms that each conclusion is supported by the evidence. Matched to your field rather than allocated from a rota.

OwnsScientific accuracy and interpretation
Reads againstThe study protocol and current literature
FieldMatched to your topic
BiostatisticianStatistical input where the work reports analysed data

Checks that statistical methods, tables and conclusions are consistent with the analysis plan and that the analysis can be reproduced.

OwnsStatistical accuracy
Reads againstICH E9 and the statistical analysis plan
ToolsR, SAS, Stata, Python
Data specialistData handling, cleaning and documentation

Prepares, cleans and documents the data so that every step from raw data to final output can be traced and reproduced.

OwnsData handling and documentation
Reads againstThe data management plan and applicable data standards
OutputDocumented, reproducible datasets
Second reviewerIndependent check of code and outputs

Independently checks code, tables and outputs against the analysis plan. Discrepancies are logged and resolved before delivery.

OwnsIndependent quality control
Reads againstThe analysis plan and outputs
Signed off byProject lead

Sample work

Sample Bioinformatics Analysis Report

Every Pubrica analysis report includes the decisions as well as the results: what was fixed before the groups were unblinded, what was excluded and why, and how the results change under alternative choices. Use the tabs to view each part.

Bulk RNA-seq analysis report

24 sequenced, 23 analysed · GRCh38, GENCODE v44 · DESeq2

Quality control and the one exclusion · the batch effect · results · sensitivity analyses

Design
12 treated against 12 control, one tissue, one timepoint
Reference
GRCh38 primary assembly, GENCODE release 44 — 62,700 annotated gene features
Alignment
STAR 2.7.11a, two-pass
Quantification
featureCounts, gene level, reverse-stranded
Model
DESeq2 1.42.0, design ~ batch + condition
Thresholds
Adjusted p < 0.05 and |log₂FC| > 1, fixed before unblinding
Environment
Nextflow, Singularity containers, seeds fixed, lock file shipped
Why the build is stated
The same reads against GENCODE v38 and v44 do not give the same counts

Redacted extract · client and study details changedPage 1 of 4

The report is the result and the decisions that produced it. Use the tabs to see each part.

The batch tab shows the most important decision. Leaving the batch term out of the design would have returned 158 genes rather than 699, a smaller result that appears publishable but is wrong. The finished figure would not have shown this, which is why the principal component plot is inspected before any model is fitted.

The exclusion is reported both ways. One sample was excluded on quality metrics recorded before the group labels were viewed, and the report states how the result changes if it is kept. Reviewers often ask this, and answering it in the report avoids addressing it in a response letter.

Thresholds fixed before unblindingIncluding the exclusion rule, with the date recorded.
Every version and parameter recordedReference build, annotation release, tool versions, seeds.
Sensitivity analyses pre-specifiedAll of them are reported, not only the reassuring ones.
Rerunnable pipeline suppliedPipeline, containers and lock file ship with the report.

If you request sample work, we will send a report based on an analysis similar to yours, with the same assay, organism and type of design.

Bioinformatics · sample analysis report

Bulk RNA-seq analysis report

24 samples sequenced, 23 analysed · GRCh38, GENCODE v44 · DESeq2 with batch in the design · quality control, model, results, sensitivity analysis and the methods section. Client, organism and study details have been removed or changed. Ask us and we will send one built around an analysis like yours.

1Analysis Settings and References

Everything below was fixed in the analysis plan before the sample groups were unblinded. The plan, this report and the code produce the same numbers on a rerun, which is why they are recorded in full detail.

ElementValue
Design12 treated against 12 control, single tissue, one timepoint
ReferenceGRCh38 primary assembly, GENCODE release 44 — 62,700 annotated gene features
AlignmentSTAR 2.7.11a, two-pass, default splice handling
QuantificationfeatureCounts, gene level, reverse-stranded
ModelDESeq2 1.42.0, design ~ batch + condition
ThresholdsAdjusted p < 0.05 (Benjamini–Hochberg) and |log₂ fold change| > 1
EnvironmentNextflow pipeline, Singularity containers, seeds fixed, lock file shipped

The reference build and annotation release are stated because they affect the results. The same reads quantified against GENCODE v38 and v44 produce different gene counts, and a reviewer who cannot reproduce a count from the methods is likely to question the analysis.

2Quality Control and Sample Exclusion

MetricResultNote
Read pairs per samplemean 38.2M
range 31.4–44.7M
All above the 25M planned minimum.
Retained after trimming97.1%Adapter content negligible across the run.
Uniquely mappedmean 91.6%
range 87.3–94.1%
Excluding the sample below, which mapped at 74.8%.
Duplicationmean 22.4%Consistent with library complexity at this input.
Genes passing the filter14,382 of 62,700Filter applied before groups were compared.
Excluded1 sampleRIN 5.2 against a cohort median of 8.4, with marked 3′ bias and 41% exonic reads against a cohort mean of 78%. Excluded on these metrics, recorded before the group label was looked at.

The exclusion rule was written down before the data was examined. Excluding a sample after seeing which arm it belongs to is not a quality control decision, and a reviewer cannot detect the difference from the finished table. The rule, the metrics and the date are therefore all included in the report. The sample was in the control arm. Section 5 reports the result if it is kept.

Analysed set: 23 samples, 12 treated and 11 control.

3Batch Effect and How It Was Handled

Principal component analysis was inspected before any model was fitted. The first component did not separate the conditions.

ObservationValue
PC124.1% of variance — corresponds to library preparation batch, not condition
PC211.3% of variance — corresponds to condition
Batch structureTwo preparation batches, 12 and 11 samples, partially crossed with condition
HandlingBatch included as a term in the design, rather than corrected before modelling
Genes called with batch in the design699
Genes called without it158

This is the single most consequential decision in the analysis. Omitting the batch term would have produced 158 genes and a publishable-looking result. Including it produces 699 and a correct one. Because the batches are only partially crossed with condition, some of the condition effect is not separable from batch; that limitation is stated in the manuscript rather than left implicit, and the design that would have avoided it is noted for the next study.

4Results

ResultValue
Genes tested14,382
Differentially expressed699 (adjusted p < 0.05 and |log₂FC| > 1)
Higher in treated412
Lower in treated287
Fold changes reportedShrunken (apeglm) for ranking; unshrunken values in the full table
Enrichment backgroundThe 14,382 genes tested — not the annotated genome

The choice of background has a large effect. Run against all 62,700 annotated features, this gene list returns many enriched terms, most describing what is highly expressed in the tissue rather than what changed. Against the tested background the list is shorter and correct. This is the most common analysis error in published transcriptomics.

5Sensitivity analyses

Each analysis below was specified in the plan. All are reported, including the least favourable.

  • Keeping the excluded sample. 512 genes are called, of which 451 are in the main set of 699. The direction of effect is unchanged for all 451, so the exclusion increases power but does not create the finding.
  • An alternative quantifier. Salmon in place of STAR and featureCounts gives 681 genes, with 94% agreement on the called set. The conclusion does not depend on the quantification route.
  • An alternative filter. Raising the expression filter reduces the tested set to 12,906 and the called set to 654, with no change in the enriched pathways.
  • Leave-one-out. No single sample accounts for more than 4% of the called genes, so the result is not carried by one outlier.

6Files Delivered

  • The raw and normalised count matrices, and the full results table for all 14,382 genes tested rather than the 699 that were called.
  • The Nextflow pipeline, the container definitions, the lock file and the fixed seeds — enough to reproduce every number in this report.
  • Per-sample quality control, including the metrics behind the exclusion and the date they were recorded.
  • Publication-ready figures in vector format: PCA, volcano, heatmap and enrichment, each with the script that produced it, in a palette that survives greyscale and common colour blindness.
  • A methods section written to journal length, with versions, parameters, thresholds and the accession already in it.
  • The submission package for GEO, with MINSEQE-level metadata and the processed matrix alongside the raw reads.

Pubrica runs the analysis; the biological interpretation and the authorship remain yours. Where a bioinformatician’s contribution meets ICMJE authorship criteria we say so and let you decide. Any AI assistance in code or drafting is declared.

Prepared by Pubrica. Pubrica is a trading name of Guires®, registered in England, Wales and India.

Download this as a PDF
Before and after · the analysis methods paragraph, redactedBulk RNA-seq · 24 samples

As supplied

Differentially expressed genes were identified using DESeq2, and pathway enrichment analysis was performed in DAVID.

  • No reference build or annotation release, so the gene identifiers cannot be reconciled with other datasets.
  • No design formula, so a reader cannot tell whether batch was accounted for; in this dataset, batch was correlated with the groups.
  • No threshold and no multiple-testing correction named.
  • Enrichment run against the whole genome rather than the genes actually tested, which inflates every p-value in the table.

As delivered

Reads were quantified against GRCh38 with GENCODE v44. Differential expression was tested in DESeq2 1.44.0 using the model ~ batch + group, batch having been identified in the PCA before any group comparison; genes with an adjusted p-value below 0.05 after Benjamini-Hochberg correction and an absolute log₂ fold change above 1 were called. Over-representation was tested against the 14,382 genes that passed the expression filter, not against the genome. Code, package versions and the container image are deposited with the accession.

  • Build and annotation release are recorded, so the identifiers mean something.
  • The batch term is in the model, and the paragraph says when batch was discovered — before the comparison, not after.
  • The threshold and the correction are both named.
  • The enrichment background is the tested set, which is the single most common error in this kind of analysis.

With the batch term the analysis returns 699 differentially expressed genes. Without it, 158. Omitting the batch term produces a result that does not replicate. The paragraph on the right also serves as the methods section, written while the decisions were being made rather than reconstructed afterwards.

Benefits

Benefits of a Pubrica Bioinformatics Analysis

Each analysis produces a result that withstands independent checking, together with the files needed to check it.

01An analysis that reruns and gives the same numbersThe pipeline, containers, lock file, seeds, reference build and annotation release are all delivered, so a reviewer, a collaborator or your own team can reproduce each figure later. A result that can be reproduced can be defended without repeating the analysis from the raw data.
02Analysis Decisions Made in the Correct OrderThresholds and exclusion rules are written down before the groups are unblinded, the principal component plot is inspected before a model is fitted, and the enrichment background is set to the genes actually tested. None of these is visible in a finished figure, and each of them affects the result.
03Methods Section IncludedWritten during the analysis, while the versions and parameters are known, rather than reconstructed from memory months later. Reviewers in computational fields often read the methods first, and an incomplete methods section commonly leads to a request for major revision.
04Data Prepared for Repository DepositPrepared for GEO, SRA, ENA, PRIDE, EGA or dbGaP with the metadata each requires, and the accession issued before submission rather than requested during it. Where the consent does not permit open deposit, that is established at design rather than discovered at acceptance.

Packages

Bioinformatics Service Packages

Every package is run by a computational biologist with a statistician and a domain specialist drawn in as the analysis needs them, and a dedicated project coordinator.

Rush

1 week

Available where capacity allows. Rush delivery is quoted on capacity rather than urgency, so it does not carry a deadline surcharge.

Standard

2–3 weeks

The standard option, which leaves time for a revision round before your own deadline.

Extended

4–5 weeks

For larger scopes, multi-database searches and projects that need a second reviewer throughout.

Basic

Exploration and primary analysis

Researchers who need the data processed, described and looked at properly

From $600up to 12 samples · one assay

Delivery 2–3 weeks

  • ✓Quality control and preprocessing, reported per sample
  • ✓Primary genomic, transcriptomic or proteomic analysis
  • ✓Summary report with figures, and the code that made them
  • ✓Versions, parameters and environment recorded

Add later: statistical validation; reference alignment; figure preparation.

View a sample Select Basic
Advanced

Full analysis and integration

Academic and industry projects needing depth and an integrative answer

From $1,450up to 24 samples · design in the model

Delivery 4–6 weeks

  • ✓Everything in Basic
  • ✓Differential analysis with the design in the model
  • ✓Multi-omics integration where the data supports it
  • ✓Pathway and enrichment against the correct background
  • ✓Structural bioinformatics and molecular modelling
  • ✓Pre-specified sensitivity analyses, all reported

Add later: machine learning; clinical data integration; journal formatting.

View a sample Select Advanced
Premium

Pipeline, publication and translation

High-impact publication, grant submission and translational programmes

From $3,400one pipeline · containerised · deposit package

Delivery 8–10 weeks

  • ✓Everything in Advanced
  • ✓Custom pipeline development, containerised and documented
  • ✓Multi-layered omics across genomics to epigenomics
  • ✓Machine learning for biomarker discovery, reported to TRIPOD+AI
  • ✓Manuscript support with supplementary materials
  • ✓Deposit package prepared and the accession issued

Add later: regulatory check; clinical trial datasets; interactive dashboards.

View a sample Select Premium
Review

Reanalysis and rescue

For analyses run elsewhere that a reviewer has queried

From $750one analysis reconstructed · one reviewer’s objection

Delivery 1–3 weeks

  • ✓Reconstructing what the previous analysis actually did
  • ✓An independent assessment of whether the reviewer’s objection is valid
  • ✓Rerunning the analysis correctly where needed, and explaining where it is not
  • ✓A written response to the reviewer, supported by the results

Often the least expensive option on this page and the most useful at the revision stage.

View a sample Select Review

Per sample

Standard assays

From $100per sample

Bulk RNA-seq, WES, ATAC-seq, ChIP-seq

Raw reads to an interpreted result, with per-sample QC, a model matched to the design, and the code and containers delivered. Higher volumes reduce the per-sample price, and the quote states by how much.

Per sample

Deep and single-cell assays

From $200per sample

Single-cell RNA-seq, WGS, multi-omic

The same route with the extra stages these assays need: cell calling, doublet and ambient RNA removal, integration checked for over-correction, and condition testing at the level the study randomised.

Free

Analysis plan review

Freein 2 business days

Before you sequence anything

What your design can support, how many replicates at what depth, and how to batch the sequencing so batch is not confounded with the question. The only stage at which the answer can still be changed.

Per-sample pricing covers the analysis, not the sequencing. What moves the figure is rarely the compute: it is metadata that has to be reconstructed, access approvals for controlled data, or a quality problem that needs a decision from you. All three are flagged at scoping rather than added to an invoice.

An analysis plan can be bought on its own, before any data exists. What the design will support, how many replicates at what depth, how to batch the sequencing so that batch is not confounded with the question, and which covariates to record while it is still possible to record them. It costs less than any of the packages above and it is the only stage at which the answer can still be changed.

Quality Assurance

Our Service Commitments and Limitations

An analysis reports what the data shows, which may differ from what you expected. Our terms are set out below before you commit.

The Pubrica guarantee

Revision Support Until Agreed Requirements Are Met

We revise the analysis until it meets the plan and the standard it is written to, with no limit on rounds within the agreed scope.

01

Unlimited revisions within the agreed scope

There is no cap on rounds or time within the agreed scope. If reviewers request a different model, a collaborator needs another comparison or a new batch arrives, we rerun the analysis.

02

Corrections at Our Cost

If an error originates with us, such as a wrong parameter, a mismatched build or a misspecified background, we correct it and rerun all downstream steps at no charge within the agreed scope, and we notify you first.

03

Reproducible Results

Every analysis is delivered with the code, containers, versions and seeds. If it does not rerun to the same numbers on your system, we will correct it within the agreed scope.

04

Your Interpretation and Authorship

We run the analysis, state what it supports and tell you plainly where we disagree with a reading of it. What goes in your name is settled by you.

Limitations of the service. We cannot guarantee a significant result and will not produce one. If the analysis returns a null result, that is the deliverable, and we will help you report it accurately, including whether the study was underpowered. We will not run comparisons until one reaches significance and report it as planned, set a threshold after seeing how many genes it returns, or change an exclusion rule once the group labels are visible. If asked to do any of these, we will decline and explain why.

Free Review by a Computational Biologist

Before quoting, a Pubrica computational biologist reviews your design and data and tells you what the analysis can and cannot support. Occasionally the review shows that the available samples cannot answer the question, which is better known before further costs are incurred.

Every version recorded

Tool versions, reference build, annotation release, parameters, seeds and the environment definition, shipped as a container or a lock file rather than a list of package names. Where a previous analysis arrived without them, reconstructing what was actually run is usually the first thing we do and occasionally the only thing needed.

GDPR, HIPAA and human genomic data

Human genomic data is identifying and cannot be fully anonymised, so it is handled as identifying throughout. Work is done in access-controlled environments under a quality management system certified to ISO 9001:2015, with a non-disclosure agreement signed before any material moves, and data kept only as long as the project needs it. Pubrica works with groups in India, across the United Kingdom and Europe, in North America, throughout Asia-Pacific and the Middle East, and in Africa and Latin America; nothing is used to train anything and nothing is shared outside your project team.

Authorship and AI Declaration

We run the analysis and write the methods; the biological interpretation and accountability remain yours. Where a bioinformatician’s contribution to design, analysis or interpretation meets ICMJE authorship criteria, as it often does when we designed the analysis, we tell you and let you decide rather than relying on an acknowledgement. Any use of AI in code or drafting is declared.

Client Testimonials

Client Feedback on Bioinformatics Projects

Testimonials are rephrased from client feedback.

“They found a batch effect in our sequencing that we had not seen, and showed us what the result looked like with and without it in the model. That comparison is what we ended up putting in the paper.”
Olivia Bennett — Genomics researcher · UK Published inScientific Reports4.9
“The RNA-seq analysis came back with the code and the containers, so when a reviewer asked us to rerun it with an extra covariate we could do it in an afternoon rather than starting again.”
Mason Reed — Molecular biologist · USA Published inJAMA Network Open11.7
“What we valued was being told which of our pathway results were an artefact of the background we had used. It was not the answer we wanted and it was the right one.”
Henry Walker — Research scientist · United Kingdom Published inNature56.1

Bioinformatics FAQs

Bioinformatics FAQs

If your question about study design or a reviewer’s objection is not answered here, Ask an expert — no quote is required.

Open-source tools for almost everything, with licensed platforms where a project genuinely needs one. In practice that means STAR, HISAT2, Salmon, DESeq2, edgeR and limma for transcriptomics; Seurat, Scanpy and scVI for single cell; BWA-MEM2, GATK and DeepVariant for variant calling, with VEP or ANNOVAR and gnomAD and ClinVar for annotation; DADA2 and QIIME 2 against SILVA or GTDB for microbiome; MACS3 and the ENCODE quality metrics for chromatin; and FragPipe, MaxQuant or DIA-NN for proteomics, deposited through PRIDE. Pipelines are written in Nextflow or Snakemake with Singularity or Docker containers. Which tool is right depends on the data and the question, and in several places there is more than one defensible answer — we make the choice with you and record it rather than making it silently. We also say when a tool you are already using has been superseded: TopHat and Cufflinks are both in low-maintenance status and their own authors point users at HISAT2 and StringTie.

Send your design and a computational biologist will advise you at no cost within two business days. The choice depends less on the amount of data than on the scope of the analysis: processing and describing a dataset, modelling group differences with the design handled correctly, and building a pipeline your team will run repeatedly are different types of work. It also depends on the destination. For a thesis chapter, the primary analysis is usually sufficient. For a journal with a data availability policy, you also need the deposit package and the code. For a regulatory submission or a clinical decision, documentation is most of the work and should be scoped from the start.

By working to the standard that actually governs the deliverable, which differs by assay. Data handling and deposit follow the FAIR principles, with MIAME for microarray work, MINSEQE for sequencing, MIxS for environmental and microbiome data and MIAPE for proteomics — these are what repositories enforce at deposit. Variant discovery follows GATK Best Practices or the documented equivalent for the caller used. Where the deliverable is a predictive model, reporting follows TRIPOD+AI; where it is a diagnostic accuracy claim, STARD; where it is a prognostic biomarker, REMARK. Clinical and laboratory work sits under GCP and GLP where applicable, and human data under GDPR, HIPAA and, for NIH-funded work, the Genomic Data Sharing Policy. Where the analysis supports a regulatory submission, the expectations of the FDA, the EMA or the equivalent authority govern the documentation. Publication ethics and authorship follow ICMJE and COPE. One point worth making plainly: CONSORT is a reporting guideline for randomised trials and does not govern a bioinformatics workflow, so we do not claim it does.

Yes, although integration is more difficult than it appears and is often less informative than a well-run single-layer analysis. Combining genomics, transcriptomics, proteomics, metabolomics and epigenomics on shared samples can reveal mechanisms that no single layer shows, but only where the samples are genuinely shared, the sample sizes are adequate and the layers were not processed in different batches. Where these conditions hold, we use factor-based and network methods and report which layer carries the signal. Where they do not, we will tell you. Integrating clinical or phenotype data adds a further issue: clinical variables are usually incomplete and rarely missing at random, so the handling of missing data becomes part of the result.

A small dataset with a standard design usually takes one to two weeks, a medium project with modelling and interpretation three to four weeks, and a large or multi-omics analysis five to eight weeks, with priority delivery available. Timelines are more often affected by incomplete sample metadata that must be reconstructed, access approvals for controlled data, or quality problems that need your decision at the first pass than by computation. We flag all three at scoping, and if a date is at risk we tell you early and explain the options.

A non-disclosure agreement is signed before any material moves — yours if you have one, ours if you prefer — and writers, analysts and project staff are bound by internal confidentiality agreements as well. Data is held in encrypted, access-controlled environments under a quality management system certified to ISO 9001:2015, processed under GDPR and HIPAA-compliant handling, kept only as long as the project needs it and then deleted on your instruction. Nothing is used to train any model and nothing is shared outside your project team. The specific point about genomic data is worth stating: a human genome is identifying and cannot be fully anonymised, so it is treated as identifying throughout rather than as de-identified once the name is removed, and controlled-access deposit is planned at design rather than at submission.

Yes. Reviewers may request an extra covariate, a different model or a sensitivity analysis, and because the pipeline, containers and seeds are delivered with the report, these can usually be run quickly. New samples can be added to the existing analysis with batch handled correctly, figures can be rebuilt to another journal’s specification, and where a tool that produced your results has been superseded, we will tell you whether rerunning changes the conclusion.

Replicates almost always matter more than depth, and buying depth instead of replicates is a common way to waste a sequencing budget. For bulk differential expression, most designs are better served by more biological replicates at moderate depth than by fewer at high depth, because the main source of variance is between individuals rather than between reads. For single-cell work, the number of samples matters more than the number of cells: cells within a sample are not independent replicates, so a study with forty thousand cells from three people has three replicates. For variant calling, depth does govern sensitivity, particularly for low-frequency variants. The right design depends on the effect size of interest and the variability in your system, so we estimate it from pilot data or published estimates and tell you what the design can detect, including when it cannot detect the effect you are looking for.

Yes, and this is a separate service from rerunning the analysis. The first step is reconstructing what the original analysis did, which is straightforward if code and versions exist and is most of the work if they do not. The objection is then assessed on its merits. Objections generally fall into three groups: those that are correct and require a rerun; those that are correct in principle but can be answered with a sensitivity analysis; and those based on a misreading of the methods, which are usually resolved by revising the text. Where the objection is correct, we say so, including when the original conclusion does not hold.

Then that is the result, and we will help you report it accurately rather than searching for a comparison that reaches significance. A study that found no effect differs from a study that could not have found one, and establishing which applies is the useful next step. If the design was underpowered for the relevant effect size, that is worth knowing before planning the next experiment and is a legitimate, publishable finding. We will not run comparisons until one is significant and present it as planned, or move a threshold after seeing the results, because such results do not replicate.

Either. Working from raw data such as FASTQ, BCL or raw mass spectra lets us control and record every decision from quality control onwards, which makes the analysis reproducible and the methods section complete. Working from a processed count matrix or peak table is faster and less expensive but inherits earlier, often undocumented choices about trimming, alignment, normalisation and filtering. In that case the report states which decisions were made before the data reached us and which we could not verify, which protects you at review.

Yes, and it is best to start early. Deposit is a metadata exercise governed by community standards, such as MINSEQE for sequencing, MIAME-derived expectations for arrays and the PRIDE requirements for proteomics, and the time-consuming part is assembling consistent sample-level metadata. Journals with a data availability policy hold acceptance until an accession exists, including accessions issued under embargo. We prepare the submission alongside the analysis so the accession is available when the manuscript is submitted.

Pipelines are written in Nextflow or Snakemake with Singularity or Docker containers, which means they are portable by construction: the same workflow runs on our infrastructure, on your institutional cluster, or in your cloud account. Running in your environment is the right choice when the data cannot leave it — patient-identifiable sequence data under HIPAA or GDPR, or data governed by a controlled-access agreement — and we work that way regularly. You receive the workflow, the container image and the parameter files whichever route is taken, so the analysis does not depend on us being available to re-run it.

Often, and the attempt is informative in itself. Where the code, reference build, annotation release and parameters were recorded, reproduction is straightforward and usually confirms the result. Where they were not, we report a range: the result under the most plausible reconstruction and how much it changes under alternatives. That range is the appropriate answer to a reviewer who asks whether the result is robust. We have also found cases where the original result depended on an unrecorded decision, which is better identified before publication.

We do not treat cells as replicates, which is the most consequential error in the field. Thousands of cells from three mice provide precision about those three mice, not about the population; the experimental unit is the donor, and a differential expression test run across cells returns many significant genes that do not replicate. The appropriate approach is pseudobulk aggregation to the sample level followed by a standard differential expression framework, or a mixed model with a random effect for donor. We state which was used and why, and where a published analysis treated cells as independent, we explain how that affects the claim.

Request a Free Design Review by a Computational Biologist

A computational biologist will assess what the analysis can support, what it cannot, and which pipeline decision matters most for your question, within two business days, at no cost and with no obligation. Some design decisions can only be changed before sequencing.

Guides and articles in this section

Show all 1
Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.