OWKIN

Owkin is an artificial intelligence-based biotechnology company developed to analyze biological knowledge and optimize medical decision-making processes. The company’s focus is on understanding disease biology at scale, selecting therapeutic targets, guiding clinical development processes, and supporting diagnostic decisions. Within this framework, Owkin is developing a biological artificial superintelligence approach, termed Biological Artificial Superintelligence (BASI), which integrates agentic AI systems, multimodal patient data, and experimental biology infrastructure into a single closed-loop system. The aim of this approach is to systematically model and engineer biology to generate novel solutions in drug discovery, clinical development, diagnostics, and aging biology.

Owkin defines itself as an “AI biotech” focused on deciphering biological complexity at the patient level. The company’s core assertion is that biology cannot be understood through a single data type or computational tool; therefore, it is necessary to integrate genomic, transcriptomic, spatial omic profiles, histopathology images, clinical outcome data, and longitudinal patient records. In parallel, Owkin is developing autonomous decision systems capable of identifying causal relationships in biology, generating hypotheses, and experimentally testing and feeding back these hypotheses. This approach is described as a transition from traditional narrow AI tools to general-purpose biological reasoning systems.

Biological Artificial Superintelligence (BASI)

Biological Artificial Superintelligence (BASI), as defined by Owkin, is a multi-tiered reasoning infrastructure with high capacity to understand, intervene in, and reprogram biological systems. The BASI concept refers to a closed-loop intelligence architecture that goes beyond models that merely automate single tasks, enabling inference about biological phenomena, decision-making, and updating decisions with new experimental evidence. Components of this architecture include multimodal biomedical data layers, domain-specific biological reasoning models, agentic AI workflows, and automated experimental validation processes. The goal is the systematic generation of breakthrough biological insights for therapies, diagnostic tools, and regenerative medicine approaches. BASI forms the foundation of Owkin’s long-term vision of a general-purpose biological intelligence, described as the operating system of biology.

Owkin K and K Pro: Agentic AI Platform

Owkin K is the company’s agentic AI platform. The platform aims to establish causal relationships on multimodal patient data, propose biological hypotheses, and generate decision support. Owkin K is used for tasks such as data exploration, prioritization of biomarker candidates, classification of patient subgroups, support for clinical trial design, and evaluation of molecular targets.

Owkin K Pro is the productized version of this infrastructure for the biopharmaceutical industry. K Pro is described as a decision support co-pilot that operates with agentic decision logic. The term “agentic” indicates that the software is not merely a reactive tool but an autonomous agent capable of planning its own actions to achieve goals, collecting and analyzing data, invoking appropriate tools, evaluating outcomes, and refining its approach. The system can sustain a cycle of planning, execution, and re-evaluation without continuous human guidance. This architecture is applied in complex stages such as research hypothesis formulation, biomarker discovery, patient stratification strategies, and clinical study optimization.

The company positions K Pro as a system that generates action-class responses in biopharma—not merely data presentation, but evidence-based recommendations for specific biological or clinical questions.

OwkinZero: Biological Reasoning Model

OwkinZero is defined by Owkin as a reasoning model specifically trained for biological discovery. The model’s training is based on over 300,000 validated question-answer pairs spanning eight core biomedical domains. These question-answer sets cover critical decision-making questions in drug discovery, such as drug target druggability, selection of appropriate therapeutic modalities, drug perturbation effects, gene expression changes, and cellular contexts.

The training data for OwkinZero is not merely literature-based general knowledge. Patient data, spatial and bulk transcriptomic profiles, and Owkin’s internal expert knowledge are converted into natural language form and fed into the model. The model is then further refined using reinforcement learning—for example, via post-training with GRPO on Qwen3-based models. Thus, the model aims to move beyond merely repeating existing knowledge and instead achieve reliability at biological decision points.

OwkinZero is embedded within K Pro as a reasoning engine. Complex biological questions are directed to this engine, which attempts to integrate multimodal omic data, literature knowledge, and clinical outcome data into a single coherent response.

Owkin defines agentic AI as a four-stage cycle: perceive, reason, act, learn. The system first collects and interprets data from its environment. It then derives goals and an action plan from this information. Subsequently, it executes operations through connected tools—such as querying databases, generating hypotheses, triggering experimental validation steps, or reporting results. In the final stage, the system evaluates outputs, records strengths and weaknesses, and readjusts its strategy. This structure implies a decision architecture that operates without constant human oversight, adapts to context, and responds to new conditions.

In healthcare, such agents can be categorized into sub-roles: data agents connected to electronic health records, document agents extracting structured information from clinical notes, compliance agents performing regulatory checks, research agents scanning literature, and discovery agents predicting drug-response relationships in drug discovery. Owkin asserts that integrating these capabilities into a single platform will enhance efficiency in drug research and development processes.

Data Network and Multimodal Biological Data

Owkin positions its access to high-resolution, multimodal patient data as a strategic defensive moat. Through collaborations with academic research hospitals and leading clinical centers, the company collects and standardizes histopathology images, spatial omic data, single-cell and bulk transcriptomic profiles, clinical outcome measures, and longitudinal follow-up data. This structure aims to represent heterogeneous patient populations and preserve biological diversity.

Through initiatives such as MOSAIC, the company focuses on molecular mapping of biological tissues at single-cell and spatial resolution. The goal is to precisely define cell type composition, cellular interaction zones, tumor microenvironments, and immune infiltration areas with high spatial accuracy. Such data are subsequently used as training inputs and validation sources for models like Owkin K and OwkinZero.

This data strategy is described as a deepening data moat that is difficult to replicate externally. Through access agreements, longitudinal clinical follow-up data, and automated laboratory infrastructure, the system becomes a self-sustaining feedback loop. This mechanism ensures that results from each new patient cohort and each new experimental perturbation flow back into the model.

Clinical Applications and Product Lines

Owkin’s technology is positioned in three primary application areas: drug discovery and development, optimization of clinical trials, and diagnostic decision support.

For drug discovery, the company employs agentic AI workflows in decision stages such as target prioritization, modality selection, drug response prediction, and biomarker extraction. One of the first clinical candidates emerging from this pipeline is OKN4395, a selective triple inhibitor program developed under the Epkin umbrella. This molecule targets EP2, EP4, and DP1 receptors and has entered the INVOKE Phase 1 study. This program is presented as an example of an AI-supported pipeline from target selection to clinical development.

In clinical development, Owkin works on methods influencing trial design, including patient stratification strategies, external control arms, and multivariate covariate adjustments. Projects published by the company—such as RlapsRisk BC, MSIntuit CRC, MesoNet, PACpAInt, and TLS Detect—aim to extract patient prognosis, treatment response, or risk classification from histopathology images or multimodal omic data. The goal is to make trial arms more homogeneous, define biomarkers, and clinically meaningful patient subgroups.

In diagnostics, Owkin is developing AI-powered digital pathology solutions under the brand Owkin Dx. These solutions are designed to support pathologists, perform biomarker screening, and predict patient outcomes. The company also states that it is developing AI-assisted diagnostic products compliant with regulatory requirements for medical device classification.

Owkin reports that alongside its closed commercial product line, it maintains open science activities as a separate axis. The company publishes code, model weights, and datasets via GitHub and Hugging Face. Components of this approach include PyDESeq2 (a Python adaptation of the DESeq2 method for RNA-seq differential expression analysis), FedECA (federated methods for causal inference and external control arms in time-dependent outcomes), and self-supervised visual transformer models for histopathology (such as Phikon and Phikon v2).

Published academic works cover topics such as training models across multiple centers without centralized data sharing via federated learning, interpreting spatial transcriptomic data to derive patient-level predictions through multi-sample learning, predicting metastatic recurrence risk from histological images, and forecasting gene essentiality and drug response. The company also issues public invitations to collaborative open projects aimed at accelerating target discovery in challenging tumors such as glioblastoma. The open science approach is offered both to provide tools to the research community and to encourage independent validation of models.

AI Safety and Ethics

Owkin emphasizes that due to the high sensitivity of health data, it treats security, privacy, and bias management as institutional priorities. Within this framework, privacy-enhancing techniques such as federated learning—which enable model training without centralizing patient data—are prioritized, alongside clinical validation and audit processes, mitigation of model bias, and compliance with regulatory requirements.

The company states that when integrating AI into clinical decision-making, it does not rely solely on automation; instead, it adopts a validation framework that engages clinicians and patients and incorporates human oversight. The rationale is that ethical risks in healthcare extend beyond technical accuracy to include dimensions of justice, accessibility, and accountability. For this purpose, the company underscores principles of regular re-evaluation, preparedness for unforeseen risks, and continuous education.

Organizational Structure

Owkin’s governance structure is defined around the executive committee, senior leadership team, scientific advisory board, and board of directors. Executive responsibilities include general management, operations, finance, technology development, medical affairs, commercial activities, partnerships, legal, marketing, and human resources.

The company has a multi-site geographic structure with offices and laboratory infrastructure in Paris (Boulevard Poissonnière and BioLab Campus Broussais), London (1 Pancras Square), New York (110 E 25th Street), Geneva (Rue François-Bellot), Nantes (4 rue Voltaire), and Boston (Cambridge, Massachusetts). This distribution is designed to enable simultaneous execution of clinical partnerships, biological sample flows, high-volume computation, and regulatory activities across different regions.

Strategic Positioning

Owkin’s long-term positioning rests on three foundational pillars. The first pillar is privileged access to high-quality, multimodal, patient-level, and longitudinally tracked biomedical data. The second pillar consists of AI models that transform this data into biologically meaningful representations capable of interpretable and causal inference. The third pillar is the integration of these models into agentic workflows that establish closed-loop learning from hypothesis generation to clinical validation. The company describes this structure as the operating system of biology and asserts that this operating system will be scalable across drug development, diagnostics, and regenerative biology.