The open-source orchestrator for computational biology.
From preclinical research to translational decisions. Inflexa runs in your terminal, on your machine, with the model you choose, and every result carries a signed record of how it was made.
curl -fsSL https://inflexa.ai/install.sh | bashRecommended. Downloads a prebuilt binary.
Apache 2.0 · No account required · Requires Docker or Podman
Reading the cohort — 206 samples, 19,634 genes, week-4 remission label.
Therapy assignment is confounded with baseline severity — a naive contrast would credit severity effects to response. Adjusting with IPTW and keeping only what survives across methods.
9 pathways concordant across all three methods. CXCL13 is the lead non-responder driver (E-value 2.12), independently supported by PMID 36378498.
lineage deseq2.R ← counts.tsv ← GSE109142model z-ai/glm-5.2signed ed25519:9f3c…a71b
A terminal, on your machine. No browser, no account, no upload.
Analytics
An analysis agent reads the literature for your biology, picks methods that fit your cohort, and runs every step end-to-end, recording the method, the parameters, and the model behind each one.
Learn more ›Reporting
Build interactive reports and dossiers from your analysis output, in one conversation. They come out as files on your disk: portable, version-controllable, and yours.
Learn more ›Provenance
Every claim traced to source: PubMed-cited literature, a signed audit trail, and file lineage you can walk. Ask any output where it came from.
Learn more ›Inflexa does not recall biology from training data. It goes and looks live at the databases a computational biologist would open themselves.
… and roughly thirty more. A few (DrugBank, DisGeNET, EPA CompTox) need your own API key.
Four steps, and none of them involve uploading your data to anyone.
Install
One command. No account, no sign-up, no sales call. It lands on your machine and it is yours.
Ask
Describe the biology you care about, in a terminal, in plain language. Inflexa inspects your data, reads the current literature for your modality, and proposes a plan before it touches anything.
Run
Approve the plan and it executes in a sandbox on your machine. If your laptop dies mid-run, the workflow resumes where it stopped.
Read
You get scripts, figures, tables, and a report, and beside them, the record: which method, which parameters, which input file, which model. Ask any output where it came from and it will tell you.
Bring the model you trust.
Most AI tools for science are a wrapper around one company’s model, and the wrapper is the product. Your work then inherits that company’s pricing, its deprecation schedule, its data policy, and its judgment about what a good answer looks like, none of which you chose, and none of which you can change.
Inflexa is not that. The model is a component you supply, and swapping it is a line of config, not a migration. Use a frontier model for the hard reasoning and a cheap one for the routine passes. Change your mind next month.
This matters more than it sounds. If your institution will not let patient-adjacent data reach a third-party API, run a local model and it never does. If a provider raises prices or retires the model you validated against, you move. The tool does not hold your science hostage to a vendor relationship you never agreed to.
And whichever model you pick, its identity is recorded in the provenance by resolved name, so a year from now you can still say exactly which model made which call.
- AnthropicClaude, natively supported. Your key, your account, your bill.
- Any OpenAI-compatible endpointOpenAI, or anything that speaks the same API, which by now is most of them.
- A model you host yourselfPoint it at your own endpoint and no prompt ever leaves your network.
You pay your provider directly, with your own key. We never see the prompt, and we never take a margin on your tokens.
Your data never leaves your machine.
Not “encrypted in transit.” Not “stored in your region.” Not “deleted after processing.” It never goes anywhere. There is no upload step in Inflexa.
Analyses run in a container on your own hardware. Outputs land in a directory you chose. We do not have an account for you, we do not collect telemetry, and we could not read your data if we wanted to.
The one thing that leaves is the model call, and it goes where you point it, with your key, to your provider. Point it at a model you host yourself and nothing leaves at all.
- ×Your raw data
- ×Your intermediate files
- ×Your results and figures
- ×Your provenance record
- ×Telemetry, usage, analytics
- →Model calls, to the provider you chose
You do not have to believe this. The source is Apache 2.0. Read it.
Built to be read.
Science is supposed to be checkable. Somewhere along the way, the tools that produce scientific results stopped being checkable themselves, and we all agreed to pretend that was fine.
It is not fine. If a model chose your normalization method, you should be able to see why it chose that one. If a figure came from a command, you should be able to read that command. If a vendor tells you a result is traceable, you should be able to check rather than take their word for it. None of that is possible with software you are not allowed to read.
So Inflexa is Apache 2.0. The orchestrator, the analytical engine, the skill packs, and the provenance substrate. Not a trial, not a crippled edition, not a teaser for the real product. If we vanished tomorrow, you would still have all of it, and it would still run.
A general assistant writes you a script. This is the rest of it.
20+ skill packs
Bulk and single-cell RNA-seq, proteomics, cheminformatics, imaging, causal inference, survival analysis, pharmacovigilance, and more. They are guidance packs, encoded expertise about how to do an analysis properly, not certified turnkey pipelines, and we would rather say so.
30+ biological databases
Queried live, at the moment of the question. Ensembl, PubMed, ChEMBL, Open Targets, GEO, ClinVar, STRING, FAERS, PharmGKB, and the rest of the sources a computational biologist would open by hand.
Durable workflows
Long analyses checkpoint as they go. Close your laptop, lose power, or hit a crash, and the run resumes from where it stopped rather than from the beginning.
A sandbox that actually is one
Code runs in a container that is non-root, has every Linux capability dropped, mounts your inputs read-only, and cannot reach the network unless you allow it. R, Python, Node, and the bioconda stack are already inside.
Method choices that fit your cohort.
An analysis agent reads the existing literature for your biology, evaluates the shape of your dataset, and designs a multi-step plan with method choices you can inspect and override. The agent does not run a single canned pipeline. It reasons about which pipeline applies to your biology and your cohort structure.
When multiple methods are applicable, the agent runs all of them and surfaces only the findings that hold up across approaches. Multi-method consensus separates signal from statistical artifacts, especially in underpowered translational cohorts where analytical choice can change the conclusion.
Coverage spans bulk and single-cell transcriptomics, proteomics, metabolomics, and multi-omics integration. The method, the parameters, and the intermediate results are recorded mechanically as the run happens. The plan the agent proposed, and the reasoning it gave for each choice, sit alongside that record where you can read them and disagree with them.
I'll audit the cohort for that confound first, then run two independent contrasts and keep only what agrees.
Plan ready.
┌───────────────────────────────┐│ S1 RNA-seq QC + cohort audit │└───────────────────────────────┘│┌───────────────┴───────────────┐│ │┌───────────────────────────┐ ┌───────────────────────────┐│ S2 Confounder-adjusted DE │ │ S3 IPTW + causal sensiti… │└───────────────────────────┘ └───────────────────────────┘│ │┌───────────────────────────┐ ┌───────────────────────────┐│ S4 GSEA — concordant only │ │ S5 Immune deconvolution │└───────────────────────────┘ └───────────────────────────┘
Two independent contrasts sit on a shared QC foundation:
- Confounder-adjusted DE. Responder vs non-responder at baseline, adjusting for severity and therapy arm.
- IPTW + causal sensitivity. Propensity-weighted contrast with an E-value robustness ranking, so a call has to survive the adjustment.
Only pathways concordant across both models are carried forward — in an n=206 cohort, analytical choice alone can flip a result, and consensus is the guard against that.
Approve as-is, or tell me which covariates to force into the model.
Executing — run a9d491d9. QC first, then the two contrasts branch in parallel, with GSEA and immune deconvolution downstream. I'll pull the outputs and summarise once it lands.
Reports your translational team can actually use.
Tell Inflexa what story you need to tell: responder stratification, biomarker association, survival by signature, pathway breakdown. Inflexa assembles a navigable, interactive report in the same conversation as the analysis, with no exporting and no reformatting.
Reports are designed for iteration. Ask for a different visualization, pull in additional context, or tighten the interpretation. Every version is preserved with its provenance intact. Share directly with translational leadership, clinical collaborators, or external partners.
Each section of the report is linked back to the analysis steps and source data that produced it. When a reviewer asks why a finding appears, the answer is one click away.
Generated by Inflexa on February 20, 2026 at 9:00 PM
Survival & Prognosis
Total ancestry effect on survival is null
After adjusting for subtype, age, and stage, African ancestry does not confer a survival advantage or disadvantage (OS endpoint, n=957, 137 events). Breast cancer survival disparities are largely explained by clinical factors, not ancestry-specific molecular features.
SHAP Feature Importance: Ancestry-Specific Survival Models
Top 20 features by mean |SHAP| value for each ancestry-specific CoxPH model. Colors indicate feature group.
EA Model (C-index = 0.652)
Mean |SHAP|
AA Model (C-index = 0.408)
Mean |SHAP|
AA model is dominated by PI3K–Akt–mTOR (SHAP=0.510), while EA spreads importance across TF and checkpoint features. NT5E (CD73) is the top EA feature, consistent with adenosine-driven immunosuppression.
Causal Mediation: Ancestry → Survival
Indirect effect of ancestry on OS through each mediator (n=957, 137 events).
Prognostic Signature Divergence
Of 165 unique prognostic genes, only 2 (1.2%) are shared between EA and AA, the most extreme divergence at any molecular layer.
Top Prognostic Genes by Ancestry
CoxPH coefficient and hazard ratio. Direction indicates risk-up vs risk-down.
EA Signature
n=726, 101 events, C=0.652
Showing 3 of 82 genes
AA Signature
n=106, 13 events, C=0.408
Showing 3 of 85 genes
Where did this file come from?
It is a simple question, and almost no scientific tool can answer it. Inflexa records the answer as the work happens, and because the code that writes the record is open, you can check it rather than trust it.
Targeted PubMed queries against each finding's hub genes, transcription factors, and pathways. Every claim links to its supporting PMID, tiered from basic biology to preclinical to clinical.
Each step, input, method, parameters, output, is written the moment it happens into a SHA-256 hash chain signed with your Ed25519 key. Alter an entry afterwards and verification fails. Tamper-evident, which is an honest word, rather than tamper-proof, which would not be.
Files are entities keyed on path and content hash, so a file produced in one run and read by another stitches the two together on its own. Ask an output where it came from and Inflexa walks backwards to the inputs.
Walking the provenance graph back from that file. Files are keyed on path and content hash, so reads and writes across runs stitch together on their own.
out/pathways.csv← fgsea.R sha256:4c1f…9ab2← out/de_results.tsv← deseq2.R sha256:8e07…13cd← data/counts.tsv← GSE109142 (source)
model z-ai/glm-5.2signed ed25519:9f3c…a71bverify ok · 47 entries · intact
Every step above was recorded by the execution layer as it ran — method, parameters, inputs, content hash, and the model behind each choice. The record is signed, so a later edit to any entry would fail verification. A third party can check it offline, with no account and no server.
Built for scientists doing the hardest translational work.
Inflexa is not a general-purpose bioinformatics tool. It is purpose-built for translational medicine, where the biology is uncertain, the cohorts are small, and the bar for defensibility is high.
Preclinical and translational teams
Point it at a cohort, a retrospective multi-omics study, or a public dataset. Inflexa runs the analysis on your machine and writes out the figures, the tables, and the record of how each one was produced.
Learn more ›Exploratory biomarker scientists
Identify responder signatures, biomarker candidates, and confounders across omics layers, without writing a single pipeline. Get to testable hypotheses faster, with the literature context already embedded.
Learn more ›Translational safety and clinical pharmacology
Assess mechanism-of-action hypotheses, cross-species signal translation, and emerging safety signals from expression data. Every finding is literature-grounded and scored for evidence strength.
Learn more ›From Confounded Cohort to Druggable Hypotheses
Not a mock-up. A real run on the public PROTECT cohort (GSE109142, 206 patients): three mechanism-anchored hypotheses, one disqualification the directionality analysis forced, and CXCL13 as the candidate the evidence kept converging on. Produced by the tool you can install for free.
After PK/safety/readiness penalties, rosiglitazone falls despite top biology; sodium butyrate rises to #1.
The questions we would ask
Run it on your own data.
Inflexa is free and open source under Apache 2.0. Install it, point it at a dataset you care about, and read the code that produced the answer.
Apache 2.0 · No account required · Requires Docker or Podman