Valinor Discovery Launches to Simulate Drug Efficacy in Virtual Patients
Valinor Discovery, a San Francisco-based startup, has launched from stealth to build generative machine learning models that simulate how patients may respond to therapies—prior to any clinical testing. These models are trained on matched multi-omics and clinical assay data derived from individual patients, with initial applications in oncology aimed at predicting treatment response to cancer therapeutics.
CEO Joshua Pacini told BiopharmaTrend that Valinor plans to release its first oncology-focused model, trained on matched primary cell and clinical samples, within six months. He described it as "the first to accurately simulate the holistic physiological impact of a chemical on cancer progression", framing the effort as a shift from virtual cell models to virtual patient models.
Valinor Discovery co-founders Joshua Pacini (CEO) and Zhanel Nugmanova (CSO)
To support this, Valinor is generating proprietary longitudinal datasets that integrate transcriptomic, proteomic, and methylation profiles with clinical assay results and primary cell measurements from the same patients. The simulations are intended to predict molecular-level changes (transcriptomic, protein abundance, methylation) as well as clinical assay outcomes, enabling compound prioritization, biomarker identification, and patient stratification.
Pacini noted that "the scarcity of high-quality clinical data significantly limits the development of translatable biological models", and stated that Valinor is addressing this by constructing matched datasets—from cellular samples to biopsies—explicitly designed to enable patient-level simulation of drug performance.
See also: The Pivotal Role of AI in Clinical Trials: From Digital Twins to Synthetic Control Arms
Valinor is also working with academic groups to evaluate its models and help define standards for perturbation-based predictions across therapeutic areas.
Valinor is collaborating with Prof. Fabian Theis and the Computational Health Center at Helmholtz Munich to develop benchmarking standards for perturbation models. The benchmarks are intended to support evaluation across therapeutic areas and drug types and will be shared via the OpenProblems platform. Theis, a physicist and computer scientist, leads the CHC and directs the Institute for Computational Biology at Helmholtz Munich, where his team applies machine learning to single-cell sequencing data to model biological systems.
A separate partnership with the Montgomery Lab at Stanford Medicine will focus on applying Valinor’s models to the evaluation of compounds targeting Alzheimer’s disease. The lab, led by Prof. Stephen Montgomery, has contributed to large-scale transcriptomic studies, including the Genotype-Tissue Expression (GTEx) project—a public resource that analyzed tissue samples from nearly 1,000 people to understand how genetic differences affect gene activity across organs.
According to Pacini, there are currently no industry-standard benchmarks for assessing the clinical translatability of perturbation model predictions—a gap Valinor is addressing through its collaboration with the Theis Lab at Helmholtz Munich. The company plans to release these benchmarks publicly once complete. In parallel, it is developing case studies based on both active and historical clinical trials to evaluate its models’ ability to simulate perturbations with what Pacini describes as “a very high level of accuracy compared to what actually happens at the bench or in the clinic".
Platform Access and Integration
Valinor is offering its platform through a hosted interface developed with Latch Bio, providing no-code access to generative models and workflows for biopharmaceutical developers. The platform supports use cases such as hit-to-lead analysis, biomarker panel selection, patient stratification, adverse effect modeling, clinical endpoint simulation, dose-level modeling, disease association analysis, and drug repurposing. It also enables compound evaluation against reference therapeutics to support competitive analysis and lead candidate selection.
Ready-to-use workflows are available out of the box, and Valinor offers bespoke retraining to align models with a sponsor’s asset, indication, or trial protocol. The company does not develop its own therapeutics and focuses solely on supporting external R&D teams.
The company’s advisors include individuals from both academic and industry backgrounds:
- Stephen Montgomery, Stanford University
- Chase Neumann, Recursion
- Bryan Norman, formerly Eli Lilly and Enveda Biosciences
- Tim Sullivan, Infinimmune
- Matt Donne, formerly Spring Science
Valinor is also collaborating with -omics sequencing companies to develop custom models trained on data from biopharma-led clinical trials, both ongoing and historical.
Topics: Startups & Deals
