Partner
With Us

Anubio builds digital twins of the human immune system to simulate how patients, diseases, and treatments interact before a single experiment is run. Our platform, TRAILBLAZER, predicts multicellular phenotypic responses to perturbations. This allows partners to discover novel therapeutics, identify responder populations, and de-risk clinical decisions using virtual patient cohorts.

TRAILBLAZER is the world's first multicellular phenotypic perturbation model. It takes single-cell transcriptomics, phenotype labels, and treatment information as inputs, and outputs predicted multicellular responses, including counterfactual single-cell distributions, cell-importance maps, and phenotypic classifications. It can operate in three directions: if you define a phenotype, it finds a treatment; if you define a treatment, it predicts the phenotype; if you define both, it finds the target patient population.

Most single-cell AI models treat cells as independent observations and ignore the multicellular context that governs tissue behavior. They also scale quadratically, which caps them at single-cell resolution. TRAILBLAZER is natively multicellular, uses patent-pending latent-space shaping for robust zero-shot generalization to unseen treatments and patients, and achieves quadratic-grade performance with linear-scaling compute. In practice, that is roughly a 1,000× efficiency improvement over comparable architectures.

The virtual cell field spans a range of approaches, each with different strengths. Recursion, Insitro, and similar platforms focus on industrialized phenomics at massive scale, generating proprietary image and transcriptomic datasets. Turbine builds mechanistic simulations of individual cell lines. Academic efforts such as GEARS and Arc Institute's State model focus on single-cell generalization across perturbations. Anubio is distinct in three ways: TRAILBLAZER is natively multicellular rather than single-cell, which matters for immune-mediated diseases where the response is a coordinated system behavior; our architecture is the most data-efficient in its class, so partners do not need to generate petabyte-scale datasets to see value; and we focus specifically on phenotypic patient-level outcomes rather than cell-line response, which shortens the path from prediction to clinical decision.

TRAILBLAZER is designed from the ground up to work with low-data sources, making it the most data-efficient model of its kind. As few as 5 patients can be sufficient for certain tasks. Requirements vary by use case, so we recommend reaching out to discuss your specific project.

TRAILBLAZER is trained on a harmonized atlas of more than 60 million single cells spanning healthy and diseased states. For partner engagements, the platform ingests single-cell RNA-seq, phenotype annotations, and treatment metadata. It can also incorporate bulk transcriptomics, clinical covariates, and multiomic layers where available.

TRAILBLAZER's foundational atlas is built on human data, which is critical for producing clinically relevant predictions. The platform can also ingest animal-model data for specific applications, and one of its core capabilities is translating findings from animal models into predicted human outcomes. This cross-species translation is supported by our latent-space shaping approach, which aligns representations across biological contexts.

Yes. The core model is built on single-cell transcriptomics, and the architecture is designed to integrate additional modalities such as proteomics, epigenomics, and clinical phenotypic data to enrich predictions and patient stratification.

Our primary focus is immune-mediated disease. Active programs include immuno-oncology (checkpoint resistance), multiple sclerosis, systemic lupus erythematosus, rheumatoid arthritis, and trained innate immunity for respiratory viral resilience. The underlying architecture is disease-agnostic and can be extended wherever multicellular immune dynamics drive outcomes.

TRAILBLAZER is target-agnostic and therapeutic-modality-agnostic. It can evaluate small molecules, biologics, cell therapies, and combination regimens. Because it operates on phenotypes rather than molecular targets, it is particularly well-suited to discovering novel mechanisms of action that target-based methods miss.

Yes. TRAILBLAZER predicts individual patient responses at both molecular and phenotypic levels. In a published case study on α-PD1 checkpoint therapy, the model correctly predicted responsiveness in 10 previously unseen cancer patients and rediscovered known combination therapies zero-shot. It also nominated 10 novel first-in-class combinations that are currently in validation.

Yes. In silico clinical trial simulation is a core use case. TRAILBLAZER can project early-phase results onto larger or different populations, translate animal-model findings into human outcomes, translate in vitro results into human outcomes, and stress-test therapies against virtual patient cohorts. This approach achieves a predictive positive value of roughly 75% or higher, compared with 8 to 15% for traditional Phase III transitions.

Yes. Given a defined treatment and desired phenotype, TRAILBLAZER identifies the target patient population most likely to respond. This enables enrichment strategies, responder signature discovery, and rational inclusion and exclusion criteria for clinical trials.

No, and we do not claim it does. TRAILBLAZER is designed to guide and de-risk experimental work. Partners use the platform to prioritize which experiments to run, which combinations to test, and which patient populations to enroll, reducing wasted cycles on candidates unlikely to translate.

We believe stating scope clearly saves everyone time. TRAILBLAZER is not currently used for pure pharmacokinetic or pharmacodynamic modeling, structural drug design, or wet-lab toxicology. The platform supports GLP-tox and regulatory safety studies as a de-risking tool but does not replace them. If your question sits at the edge of our current scope, reach out and we will tell you plainly whether we can help.

TRAILBLAZER has been validated on held-out patient cohorts, across multiple cancer types, and against published combination therapies that the model rediscovered zero-shot. Our methodology is described in our preprint on bioRxiv and is currently under review in a peer-reviewed journal. We achieve roughly 85% prediction accuracy across multiple validation cohorts and are actively validating novel candidates (ANU-01 through ANU-10) experimentally.

Yes. Mechanistic interpretability is built into the architecture rather than added on top. For every prediction, TRAILBLAZER outputs cell-importance maps that identify which cell types and states drive the predicted response, counterfactual single-cell distributions that show what changes under a given perturbation, and latent-space trajectories that make vector arithmetic between healthy, diseased, and treated states biologically meaningful. Partner scientists can trace a prediction back to the cellular signals responsible for it, which supports hypothesis generation and scientific review.

Every engagement operates as a closed loop. Experimental results from partner wet labs feed back into model refinement, either through targeted fine-tuning or through updates to our phenotypic classification. When a prediction disagrees with experimental data, we treat that as informative rather than a failure. Disagreements typically fall into three categories: data quality issues at either end, model-scope limitations we then flag openly, or genuinely novel biology that updates the model. We report all three back to partners transparently.

Every prediction is delivered with calibrated confidence metrics derived from the model's latent geometry. Partners receive ranked candidate lists with associated confidence scores. This allows teams to focus experimental resources on high-confidence predictions while treating lower-confidence outputs as hypotheses for broader screening.

TRAILBLAZER is versioned, and every prediction delivered to a partner is traceable to a specific model version, training dataset snapshot, and configuration. Model updates are communicated to active partners in advance, and prior versions remain available for reproducibility. This matters for programs that span months or years and for any results that may eventually support regulatory filings.

We are transparent about scope. The platform is strongest in immune-mediated contexts where single-cell atlases are well-developed and phenotypes are measurable at the cellular level. It is less suited to diseases driven primarily by non-cellular mechanisms (for example, pure pharmacokinetic questions), rare conditions where training data is sparse, and mechanisms outside the immune system that we have not yet incorporated.

Our 60-million-cell atlas is harmonized across donors, laboratories, and protocols using proprietary batch-effect correction. We actively monitor coverage across ancestry, sex, age, and disease stage, and flag predictions where training data is thin. Partner-specific data can be used to fine-tune the model for underrepresented populations of interest.

TRAILBLAZER is a research and decision-support platform, so it does not itself require regulatory approval. Its outputs can support regulatory submissions as part of a broader evidence package, and the platform is aligned with the direction regulators are moving. The FDA's 2025 announcement on phasing out animal-testing requirements for monoclonal antibodies and other biologics signals growing acceptance of validated in silico approaches. We work with partners on a case-by-case basis to ensure that TRAILBLAZER outputs intended for regulatory use are generated under appropriate controls, with full versioning and audit trails.

All partner data is processed in isolated, access-controlled environments with encryption at rest and in transit. Data residency can be configured to meet partner and regulatory requirements. For engagements involving identifiable patient data, we operate under HIPAA-aligned controls in the United States and GDPR-aligned controls for EU data. We are prepared to complete partner security reviews and are happy to discuss specific compliance requirements, including SOC 2 and GxP alignment, during scoping.

We work with partners in three primary modes: drug discovery and co-development partnerships, typically structured as milestones plus royalties on delivered assets; drug-rescue collaborations that revive stalled assets through AI-driven patient matching; and platform access or subscription for pharma R&D teams who want TRAILBLAZER as an internal decision-support tool across their pipeline.

Most engagements begin with a scoped pilot built around a defined disease, phenotype, or asset, and designed to deliver actionable predictions within a few months. Successful pilots typically expand into broader co-development or platform access agreements. We work with you to define success criteria and deliverables up front.

For most engagements, we ask for a clearly defined phenotype or treatment of interest, along with any proprietary single-cell, bulk transcriptomic, or clinical data you would like incorporated. We can also work entirely from public and in-house data where partner data is not available or not yet ready to share.

Deliverables are tailored to each engagement and typically include ranked candidate treatments or targets, predicted responder populations, cell-importance maps showing which cell types drive the response, counterfactual single-cell distributions for mechanistic interpretation, and written reports with confidence-scored recommendations.

Yes. Many engagements involve partner-proprietary single-cell, clinical, or experimental data. All partner data is handled under strict confidentiality and processed in isolated environments. Unless explicitly agreed otherwise, it is never used to train our general platform or inform work with any other partner.

IP terms are negotiated per engagement. As a default, partners retain full ownership of discoveries made using their proprietary data and compounds. Anubio retains ownership of the underlying platform, model architecture, and any improvements not specific to partner data. We aim for clean, unambiguous IP structures that make scaling the relationship straightforward.

Anubio was founded by Adrian Grzybowski, PhD (Founder and CSO) and Naveen Chandramohan, MS (Founder and CEO). Our team spans computational biology, AI, immunology, and protein biochemistry. Our scientific advisory board includes Dr. Russell Schwartz (Head of Computational Biology, Carnegie Mellon University), Dr. Amit Awasthi (Senior Professor, THSTI), and Dr. Sriram Venkatraman (General Partner, Foundery Innovations).

Technologies at every layer of our stack are developed in-house, and we have full freedom to operate. Core components are covered by pending patents, and other layers are protected as trade secrets with planned conversion to patents. Our platform is a research and decision-support tool and does not itself require regulatory approval. Downstream therapeutic applications follow standard regulatory pathways.

Reach out via our Partner with Us page to start a conversation. We typically respond within one business day.