NovaFlow
NovaFlow is an AI data analyst for biology labs. Scientists upload raw instrument data, ask a question in plain English, and get statistics, publication quality figures and cited reports. The workflow starts by consolidating inputs, harmonizing sample IDs, running quality control, normalizing and correcting batch effects, then proposes an analysis plan that the scientist approves before anything runs. Pipelines cover single cell and single nucleus RNA, ATAC and multiome data, bulk RNA, epigenetic assays such as ChIP and CUT&RUN, spatial platforms including Xenium, MERFISH, CosMx and Visium HD, CRISPR screens and joint multi omics, with differential expression, trajectory analysis and machine learning models among the outputs. It can reanalyze public data from GEO, CELLxGENE and ARCHS4 and draft methods sections and supplementary tables.
Analysis Portals let colleagues explore and discuss the same results, and the Standalone Platform runs the engine on an institution's own computing cluster or cloud, with data staying there. A scientific team helps with harder studies.
Amulya Balakrishnan, a former Zoom engineer, founded the San Francisco company in 2025 with a computational biologist cofounder. It went through Y Combinator's Summer 2025 batch and is backed by BioRender and FCVC, and its YC profile names researchers at UCSF, Mount Sinai, UC Berkeley and Harvard as users. The hosted workspace is free to start, with Pro at $20 a month, Max 20x at $200 and custom Enterprise plans.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
The analysis layer is the product. A plain language question becomes an analysis plan, and the system then validates the data, aligns it to relevant literature and known signatures, runs the pipeline and writes a notebook and cited report. Every hosted plan sells usage of that layer, and the Standalone Platform sells the same engine for an institution's own infrastructure. Portals and the scientific team sit around it, sharing results and helping with hard studies.
Without the model there is no way to turn a question into a pipeline across single cell, spatial, epigenetic and bulk data, which is the whole pitch: analysis in minutes rather than the months a lab can spend waiting on bioinformatics help.
A scientist approves the proposed plan before any analysis runs, and the workflow includes three auditors that check the work before the cited report and notebooks come back for review. The company's scientific team can step in for preparation and interpretation. Usage limits can pause an analysis and resume it when the allowance renews, which is a billing control rather than oversight. The structure has a clear gate at the start and inspectable output at the end.
What it leaves out is when the system stops on its own: no rule routes a failed or low quality analysis to a person, and nothing defines how uncertain results are flagged before they reach a figure in a manuscript.
The services side describes a spatial foundation model for comparing patterns across experimental measurements, and the hosted workflow pairs the plain language interface with analysis notebooks and a separate literature review. Machine learning models built for a customer get automated tuning. Each run leaves a notebook and report, so a researcher can see the code that ran.
What is not identified is the model that turns a question into a plan and the model behind the auditors, their training, or a release history. The demo reports 77.0% final accuracy on a 3,696 cell, 27,998 gene dataset processed in 18.5 seconds, without saying accuracy of what or against which reference.
The Standalone Platform runs on infrastructure the institution operates, local servers, high performance computing clusters or its own cloud, and the company states that data stays within that environment. That delivery option has a short chain by design. The hosted workspace is a different matter.
It has its own signup through Google or email and its own usage plans, and nothing names the hosting provider, the model provider that plans and runs analyses, whether outside inference services receive uploaded data, or what retention and training terms apply. No privacy policy or terms appear on the company's site. The control described for privately run infrastructure does not show who processes data in the hosted service.
The company claims 12 or more hours saved per week, 40% faster analysis and more than 2.4 million data points analyzed, with no study period, baseline or named lab behind any of them. Its Y Combinator profile names researchers at UCSF, Mount Sinai, UC Berkeley and Harvard using the software in their labs, which is adoption rather than results. The demo run, 3,696 cells and 27,998 genes in 18.5 seconds at 77.0% final accuracy, shows speed on a sample dataset. There is no reproducibility study across sites and no comparison with a reference analysis by an outside bioinformatics team.
On the Standalone Platform, data stays in the institution's environment. The hosted product lets researchers upload their own datasets and keeps results available when an analysis pauses at a usage limit, which describes continuity of work.
Beyond that, nothing covers what happens to uploaded data: no retention period after a study ends, no deidentification process for datasets with human samples, no statement on whether hosted data can train models, and no privacy policy on the company's site. The plan approval and auditors help catch analytical errors, but there is no account of how unsupported scientific conclusions are mitigated or how a safety incident would be handled.
The workflow is built for experimental data from academic and industry labs, much of it from human tissue and single cell studies. There is no statement on whether a research group may upload identifiable patient data, whether the company acts as a business associate for either the hosted or standalone service, or what limits apply to hosted uploads.
No privacy notice, terms of service or business associate agreement appears on the site, and Enterprise's security review is the only route to discuss it. The Standalone Platform keeps data on institutional infrastructure under the institution's own controls, but that is a deployment choice rather than a stated position.
Enterprise includes a security review, custom usage limits and a dedicated contact, and the standalone option lets an institution set its own infrastructure and access controls. No SOC 2 report, ISO 27001 certificate or HITRUST certification is claimed for the product or the company, and there is no trust center or stated testing cadence. The security review is a procurement step whose outcome and scope stay private. The company listed a team of four on Y Combinator, and the hosted service's controls are not described anywhere a buyer can read before signing up.
The product is research software for academic and industry laboratories. It produces plots, statistics, reports and manuscript material for investigation and publication, starting from experimental data and a scientist's question, and its outputs are research figures such as UMAPs, volcano plots and heatmaps rather than results about an individual patient. It does not diagnose patients, recommend treatments or interpret a test to direct care.
The company presents that research setting in its product material and in its BIO International Convention presentation, and the product on offer matches it. No FDA clearance, diagnostic indication or formal exclusion is claimed for any version of the platform.
A scientist can approve an analysis plan, read the notebook and review a report with citations, so each research run can be inspected. That is review at the level of a single study. There is no organization wide AI governance process, no check on whether a model update changes analytical results, no measured differences across cell types, tissue sources, platforms or institutions, and no framework for when a scientist can rely on the platform's interpretation without redoing it. For software that drafts methods sections for papers, reproducibility across versions is the question left open.
The plan approval gives the researcher a chance to review an analysis before it runs, and the notebook and cited report give the team a record to inspect a result, discuss it or recreate it through its own process. Support is by email on paid plans, with a dedicated contact on Enterprise.
No terms of service appear on the site, so there is no warranty, no obligation to correct a wrong result and no assurance covering research conclusions, and a scientist whose figure or statistic turns out wrong has no route against the company. The review features leave control, and responsibility, with the researchers using them.
The software works with research datasets: instrument outputs across single cell, spatial, epigenetic and bulk assays, plus public data pulled from GEO, CELLxGENE and ARCHS4. Teams share results through Analysis Portals and notebooks, institutions can run the engine inside their own research computing, and manuscript material such as methods sections and supplementary tables leaves the platform as files for a journal submission. Those are research data flows.
There is no connector to an electronic health record, no access to a patient chart, no link to a laboratory information system and no route for writing a result into a clinical workflow, which fits a product built for discovery labs rather than care.
Two delivery models are offered. The hosted workspace is a web app with Google or email signup, and the Standalone Platform runs the engine on infrastructure the institution operates, whether local servers, high performance computing clusters or a customer cloud, with data staying in that environment and the institution controlling where it runs. That second option settles residency for groups that need it.
For the hosted service, no region, hosting provider or tenant isolation is described, and nothing says whether processing crosses a border, which matters for labs bound by an institutional data use agreement.
Pricing is public and self service. Free costs nothing, Pro is $20 a month and Max 20x is $200 a month with 20 times the Pro allowance, and all three run the same research workflow with different usage. The allowance is measured in usage rather than a fixed number of analyses, because cost depends on dataset size and complexity. The start link opens sign in with Google or email, and when an analysis hits the plan limit it pauses and resumes when usage renews. Enterprise is quoted, with custom limits, a dedicated contact and a security review, and the Standalone Platform and scientific services are scoped with the company.
Academic research labs and biotechnology and pharmaceutical teams are the intended users, working on single cell, spatial, epigenetic, bulk and multi omics data, from Xenium and MERFISH slides to CRISPR screens. Institutions can run the standalone engine on research infrastructure, and project teams share results through portals. The users named are researchers at academic medical centers: UCSF, Mount Sinai, UC Berkeley and Harvard. No study measures accuracy or reliability by assay, tissue or research field, so the breadth of supported data types describes coverage rather than performance validated in each.
Alternatives to NovaFlow
Other products that do the same kind of work as NovaFlow, and what each one does.
Precision oncology company in Waltham, Massachusetts, combining sequencing in its own CAP accredited, CLIA certified laboratory with models of tumor and immune biology.
Mirvie predicts complications of pregnancy from a blood sample by sequencing cell free RNA from the mother, placenta and fetus.
DNAnexus runs a cloud platform where diagnostic laboratories, biopharma companies, health systems and biobanks store, process and analyze genomic and clinical data.
Healthcare information technology company (NASDAQ: HSCS) selling ECG management software to hospitals and practices.
Foundation Medicine, based in Boston and a member of the Roche Group, runs comprehensive genomic profiling for cancer in its own laboratories.
Invitae is a genetic testing brand owned by Labcorp, which acquired the business in 2024 after Invitae's bankruptcy. It keeps its own site and test menu and now presents itself as Invitae and Labcorp together, with some billing handled by Labcorp Genetics.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Free; Pro $20/month
$20 baseline
|
Monthly hosted plans priced by usage allowance: Free, Pro at $20 and Max 20x at $200 with 20 times the Pro allowance. Enterprise is quoted. | No business associate agreement or HIPAA position is offered on any plan. | Hosted plans start at signup with no setup charge. Standalone deployment and scientific team support are scoped with the company. | Vendor Published |
Free at standard usage, Pro at $20 per month and Max 20x at $200 per month, with Max allowing 20 times Pro. Analyses pause at the plan limit and resume when usage renews. Enterprise has custom limits and a security review.