GEM-1

GEM-1 (Generative Expression Model) is a foundation model developed by Synthesize Bio that predicts how human tissues and cells respond to drugs, diseases, and perturbations by generating lab-quality bulk and single-cell RNA-seq data. Built on the largest consistently normalized and annotated RNA-seq corpus ever assembled, GEM-1 allows researchers to describe experiments in natural language and receive biologically grounded gene expression predictions. It is designed for drug development teams — including clinical development, translational medicine, rare disease programs, and portfolio strategy — who need to make faster, more confident decisions while reducing the cost and risk of late-stage clinical failures.

GEM-1's prediction accuracy meets or exceeds the reproducibility standards of experimental biology, correlating with real experimental outcomes at higher accuracy than the consistency of biological replicates measured across different laboratories. Synthesize Bio partners with biotech and pharmaceutical companies to integrate these generative genomics capabilities directly into clinical development programs, with proprietary data remaining private and never used to train the models.

Unlocking Actionable Clinical Predictions

• Augment early-phase trials: Small Phase 1/2 datasets often leave critical questions unanswered. GEM-1 extends cohorts with synthetic patients matched to trial profiles, enabling teams to test endpoints, refine stratification, and model dose-response before committing to pivotal studies.
• Optimize trial design and patient selection: Test trial designs in silico before protocol lock. Model patient responses across enrollment criteria and genetic backgrounds to refine inclusion/exclusion criteria, identify subgroups most likely to benefit, and explore endpoint strategies.
• Extract value from failed trials: Augment limited datasets from negative trials to identify responder subgroups, generate hypotheses for follow-up studies, or build evidence for alternative indications — turning failures into strategic pivots.

Predicting Real Human Tissue Response

• Surface safety signals early: Identify off-target effects and tissue toxicities that preclinical models miss by predicting transcriptional responses across human tissue types, enabling teams to anticipate adverse events, prioritize safety biomarkers, and design monitoring strategies before signals threaten a program.
• Validate mechanism of action in human tissues: Understand how a therapeutic behaves in human tissues before committing to expensive studies, focusing resources on paths with the strongest biological rationale and clinical potential.
• Bridge preclinical models to human biology: Map gene expression profiles across species to validate animal findings, increase translational success, and reduce the risk of preclinical candidates failing in human trials.

Transforming Small Datasets into Development Intelligence

• Inaccessible tissue profiling for rare diseases: Turn underpowered rare disease datasets into actionable intelligence by generating synthetic cohorts that expand sample sizes while preserving the biological characteristics of limited patient populations.
• Accelerate biomarker discovery: Identify and validate predictive biomarkers before large clinical studies by generating gene expression profiles across patient subpopulations and disease states, enabling stratification hypothesis testing, companion diagnostic refinement, and enrichment trial design.
• Data sharing without HIPAA friction: Collaborate across organizations using synthetic cohorts that mirror real patient data without patient-identifying information. Synthesize Bio also offers the SYNTH-cancer database — over 10,000 AI-simulated samples spanning 27 cancer types.

Who GEM-1 Is Designed For

• Clinical Development Teams: De-risk trial design, identify predictive biomarkers, stratify patients by response likelihood, and extend early-phase data to inform pivotal trial decisions.
• Translational Medicine: Validate mechanism of action in human tissues, predict safety signals before trials, and bridge preclinical findings to human biology.
• Rare Disease Programs: Overcome small sample size limitations by augmenting limited patient datasets with biologically grounded synthetic cohorts.
• Portfolio Strategy Teams: Extract value from failed trials, evaluate alternative indications, and build evidence for label expansions without waiting years for real-world data.

Partnership and Deployment Model

• Partnerships are tailored to specific needs, ranging from focused projects addressing a particular development challenge to ongoing collaboration across a full pipeline.
• Partnership capabilities include custom model development, workflow integration, and dedicated scientific support.
• Proprietary partner data is never used to train Synthesize Bio's models.
• A public version of the GEM model is available for teams to explore simulated experiments based on natural-language descriptions; clinical applications and proprietary dataset integrations are available through partnership.
• GEM-1 insights are designed to inform internal development decisions and strengthen regulatory packages by optimizing trial design and patient selection; results currently cannot serve as primary evidence in regulatory submissions.
• Synthesize Bio maintains a Trust Center and SOC 2® report, reflecting a commitment to data security and compliance standards.