MORGAN

Deciding which gene or compound to pursue as a drug target ultimately comes down to a single question: what happens to a human cell when you perturb it? Answering that empirically is slow and expensive, and the public perturbation data that machine learning models train on is fragmented across assays, cell lines, and laboratory protocols. MORGAN — Multi-Omic Regulatory Genomics using Artificial Neural Networks — is Relation Therapeutics' flagship platform of cellular foundation models, built to predict how human cells respond to genetic and pharmacological interventions across disease contexts. It was unveiled on 30 July 2026.

MORGAN is explicitly predictive rather than generative: it forecasts the response of a cell to an intervention rather than proposing new molecules. The platform is organized as a general-purpose model applicable across cell types and disease areas, alongside tissue-specific MORGAN models that each concentrate on a therapeutically important cell type central to a particular disease biology. Relation frames the model and its data supply as one program — the company runs automated laboratories engineered specifically for high-throughput cellular perturbation experiments, and states that these will generate petascale multi-omic perturbation datasets to train MORGAN.

That coupling of model and data factory is what a simultaneous expansion of Relation's collaboration with GSK buys. Under the agreement, Relation may receive up to $110 million in upfront and success-based milestone payments to generate large-scale human cellular perturbation data and deploy it into models including MORGAN; the deal pays for data generation and joint target work rather than licensing the model itself. MORGAN is Relation's second entry in this catalog after PatchDNA, a DNA language model from the same company, and it enters a field that already includes perturbation-response models such as STATE, SCALE, and Tahoe-x1.

#Key Features

• Perturbation response prediction: Forecasts how human cells respond to genetic and pharmacological interventions, with the stated aim of revealing disease mechanisms and nominating therapeutic targets before wet-lab work begins.
• Predictive, not generative: The platform models cellular response to a given intervention rather than designing new molecules, distinguishing it from generative design models in drug discovery.
• General-purpose plus tissue-specific models: A base model spanning cell types and disease areas is complemented by tissue-specific variants, each focused on a cell type central to a particular disease.
• Purpose-built data generation: Training data comes from Relation's own automated laboratories for high-throughput cellular perturbation, producing time-resolved multi-omic readouts at a consistency and scale the company describes as beyond conventional laboratory approaches.
• Proprietary and closed: No weights, code, hosted API, or license terms have been released, and the model is available only through Relation's own programs and partnerships.

#Technical Details

MORGAN was announced through a press release rather than a preprint or technical report, and Relation has not disclosed its architecture, parameter count, context length, input modalities in detail, or any benchmark results. What the company has stated is the training regime: large-scale, high-resolution multi-omic perturbation data generated at petascale by automated laboratories designed for high-throughput cellular perturbation experiments, capturing time-resolved responses to genetic and pharmacological interventions through integrated automation. The platform is described as combining frontier-scale computation with this experimental data generation, and as being applied from a trained state across cell types and disease contexts rather than refit per dataset. No independent evaluation of MORGAN's predictive accuracy has been published.

#Applications

MORGAN is aimed at target discovery and mechanistic disease biology inside Relation's own pipeline, which spans immunology, metabolic disease, and bone disease, and at partner-facing work such as the GSK collaboration, where predictions over large perturbation datasets are used to investigate biological pathways and nominate therapeutic opportunities. The practical value proposition is triage: using predicted cellular responses to narrow the space of genetic and pharmacological interventions worth testing experimentally, then feeding the resulting measurements back into the model. Because there is no public release, researchers outside Relation and its partners cannot run the model.

#Impact

MORGAN's significance lies less in a disclosed methodological advance than in the bet it represents: that the binding constraint on cellular foundation models is not architecture but the availability of consistent, large-scale perturbation data, and that a drug discovery company should therefore build the laboratory that produces it. A major pharmaceutical partner committing up to $110 million specifically for that data generation is a meaningful validation of the thesis. The claims themselves remain unproven: without a technical report, public weights, or independent benchmarks, MORGAN's predictive performance cannot be assessed externally, and no cellular foundation model has yet produced an approved medicine.