Every model needs a first principle. Ours is age.
The Virtual Aging Cell is Insilico's next step beyond the Precious series of Large Language of Life Models — a generative model of the cell you can perturb in silico, with one axis most virtual cells leave out: age. Set a species, a tissue, an age, a compound, and read the response across transcriptome, proteome, and methylome.
What a virtual cell is — and why its fate is the whole point.
A virtual cell is a computational model of a living cell. Instead of running an experiment at the bench, you run it in silico: set the conditions — the cell type, its environment, a drug — and the model predicts how the cell responds, right down to its molecules. Trained on vast amounts of omics data, it turns the cell from something you can only observe into something you can question.
An experiment you can re-run a thousand times
Set conditions, read the cell's molecular response, and repeat — screening compounds, probing disease mechanisms, and generating hypotheses long before a pipette is involved. The bench becomes the place you confirm, not the place you search.
A cell is never just one thing
It is always on its way to becoming another — dividing, specializing, breaking down, growing old. A single snapshot misses what matters most: where the cell is headed. That trajectory has a name — cell fate.
Modeling cell fate is modeling where biology goes wrong.
Cells move between states, and most disease is a cell taking the wrong path — a healthy cell turning cancerous, a functional cell senescing, a young cell growing old. Most therapy, in turn, is an attempt to steer it back. So the real prize isn't a photograph of a cell; it's a map of its fates — predicting which state a cell will adopt under a perturbation, and how to redirect it.
The basis of regenerative medicine and reprogramming.
The transition most drug discovery is trying to reverse.
The one fate every cell shares — and the one most models ignore.
Every fate plays out over time. That's why age doesn't sit beside a virtual cell as one more label — it belongs at the center of the model.
A virtual cell you can steer.
Set a species, a tissue, an age, and a compound, then watch the cell's synthesized multi-omics response — and its predicted biological age — shift in real time.
Transcriptome · RNA-seq
Proteome
Methylome · DNAm
Illustrative synthetic output for demonstration — not real model predictions. It mirrors the class of conditional multi-omics generation described in the published Precious papers.
A public demonstration of conditional biological generation.
One important step toward a virtual cell is generative: a model that does not only read omics data, but can synthesize molecular profiles under specified conditions. The public demonstration below shows that direction. It is evidence for the underlying model class — not a claim that a fully validated virtual-cell product already exists.
Read it as the seed of the Virtual Aging Cell: once a model can synthesize a cell's state on demand, the next questions are which cell, in which species, at which age — and that's the model taking shape now.
From an aging clock to a virtual cell.
The demonstration above didn't come from nowhere. Insilico didn't start with the cell — it started with time. Each Precious model widened what a single transformer could learn about biology, until synthesizing a cell's state became possible at all.
Precious1GPT
A transformer-based aging clock — "one clock to rule them all" — that also surfaced protein targets, shown in a case study on the Apelin receptor across age-related disease.
Read the paper →Precious2GPT
A transformer paired with a diffusion model, generating synthetic multi-omics samples across species and tissues — the first hint the cell could be simulated, not just measured.
npj Aging →Precious3GPT
A genuinely multimodal model over text, tables, and knowledge graphs — outputting genes and compounds directly. Open-sourced with Harvard's Gladyshev lab as a community resource.
bioRxiv →Virtual Aging Cell
A proposed, queryable model of cellular state with biological age as an explicit condition — a direction rooted in a decade of Insilico geroscience and the public PreciousGPT research lineage.
See what makes it different →Most virtual cells model space. Ours models time.
The virtual cell is one of biology's most active frontiers. Groups such as the Arc Institute and the Chan Zuckerberg Initiative, alongside academic projects including Geneformer and scFoundation, are developing large-scale cell models to predict how molecular state changes under genetic, chemical, or biological perturbation.
Aging is often not the organizing axis. Our proposed direction makes biological age a first-class condition, so the model could compare perturbations across stages of the lifespan rather than treating age as an incidental label.
A digital lab that speaks in genes and compounds.
The public Precious research demonstrates important components of this direction. A future AI Virtual Cell would integrate and validate them within one queryable model of cellular state.
Conditional multi-omics
Explore synthesized transcriptomic, proteomic, and methylation profiles conditioned on species, tissue, sex, age, and disease state — with validation and intended-use limits made explicit.
Compound screening in silico
Trained on large-scale compound-perturbation data across many cell lines. Paired with Nach01, any SMILES structure becomes a query — making the screening library effectively limitless.
Cross-species, cross-tissue
Proteomes, RNA-seq, and DNA methylation from human and model organisms live under one model, so a signal seen in mouse can be read in the human context.
Aging clocks, natively
The lineage began as an aging clock. Predicted biological age travels with every synthesized profile — a built-in readout of how a perturbation moves the cell through time.
Target identification
Because outputs are expressed as genes and pathways, the model doubles as a target-discovery tool — surfacing regulators behind a disease or a rejuvenation signature.
Designed for scientific workflows
A future interface could connect validated model outputs to custom research pipelines and the wider Pharma.AI ecosystem. This remains a proposed integration direction, not a live product claim.
Biology, tokenized.
The Precious models learn from the raw output of experiments — not only the distilled conclusions in papers — so the cell can be reproduced at the level of its own molecular signals.
Three signal types feed the model: tabular experimental data (the unadulterated readouts), the text of PubMed, and biomedical knowledge graphs. Genes and chemical structures are native tokens — so a question and its answer are written in the same language the cell uses.
Because aging is the point.
Insilico was founded in 2014 on a single mission: to extend healthy, productive longevity. That's why age isn't an afterthought in our virtual cell — it's the reason the model exists. Long before the field converged on "virtual cells," Insilico was building deep aging clocks, publishing across Aging, npj Aging, and Aging and Disease, and collaborating with leading longevity labs including Vadim Gladyshev's at Harvard. The Precious lineage itself began as an aging clock. An virtual aging cell is simply where that decade of work leads.
See where this thinking is headed — at ARDD 2026
Insilico's Alex Zhavoronkov has long co-organized the Aging Research & Drug Discovery (ARDD) Meeting — the world's largest conference in translational geroscience — alongside Morten Scheibye-Knudsen, Daniela Bakula, and Evelyne Bischof. In 2026 the meeting moves to Boston as part of Boston Longevity Week (with Harvard's Vadim Gladyshev and Jesse Poganik), bringing together the researchers and companies building the next generation of aging models — the same community this virtual cell work is meant to engage.
Visit agingpharma.org →& Drug Discovery
An aging clock estimates biological age. An age-aware virtual cell could help ask which interventions shift molecular state — and which hypotheses deserve experimental testing.
Everything here traces to published work.
This preview is assembled from Insilico's peer-reviewed papers, preprints, and press. The primary sources:
Precious3GPT: multimodal, multi-species, multi-omics, multi-tissue transformer
2024 npj AgingPrecious2GPT: transformer + conditional diffusion for synthetic multi-omics generation
2024 AgingPrecious1GPT: multimodal transfer learning for aging-clock development and target discovery
2023 Nature MedicineRentosertib Phase IIa trial: AI-discovered TNIK inhibitor in idiopathic pulmonary fibrosis
2025 ARDDAging Research & Drug Discovery Meeting — co-organized by Insilico since its earliest years
Annual Pharma.AIPreciousGPT — model lineup, capabilities, and open-source access
Live Hugging FaceOpen-source Precious3GPT weights and multi-modal model card
LiveThe cell is becoming something you can query.
Explore the open Precious3GPT research today and help shape the scientific questions, evaluation criteria, and evidence standards for an age-aware virtual cell.
Contact us
Questions about the Virtual Aging Cell, or interested in working with us? Leave your details and the team will get back to you.
Thank you — we’ve received your details.
Someone from the Virtual Aging Cell team will be in touch.