The "virtual cell" is one of biology's most exciting goals — a model that predicts how a cell responds to a perturbation without running the experiment. Our Virtual Cell Modeling survey maps the field. But the most important recent development isn't a new model; it's an honesty correction, and it shapes exactly how BioMate is approaching the Arc Institute's Virtual Cell Challenge 2026.
The correction the field needed
Two 2025 benchmarks landed hard: Ahlmann-Eltze et al. showed deep perturbation predictors do not yet beat a simple additive model (combinatorial) or the training mean (unseen genes), and Kedzierska et al. showed single-cell foundation models often lose to plain baselines zero-shot. Even the AI-Virtual-Cell roadmap concedes no system today is a true virtual cell. The 2025 challenge winners were hybrids — curated data + explicit statistical priors + metric-aligned loss — not brute-force neural nets. One top-PDS entry used no neural network at all.
What BioMate does
- A validated scoring harness first. Before modeling, BioMate built a challenge-accurate PDS/DES/MAE harness with cell-mean, additive, and oracle-nearest-neighbor baselines on Perturb-seq data — so every claim is measured against a floor and a ceiling, exactly the discipline the survey argues for.
- Statistics-first transfer, reproduced. BioMate independently reproduced the winning behavior on its own data: a constant prediction scores chance (PDS 0.5), and similarity-weighted cross-line transfer with global PDS scale-calibration climbs to ~0.81 — confirming that perturbation-specific prediction and distance calibration are the high-leverage levers, and adding deep learning only where it beats the statistical floor.
- Mechanistic virtual-cell workflows, today. Four Phase-1 workflows run on AWS Batch — CiPA cardiac-safety ODE, ODE signaling, metabolic-flux FBA, and LINCS-L1000 transcriptomic mechanism-of-action — the regulatorily-accepted end of the virtual-cell spectrum, wired to existing ADMET/PBPK/RNA-seq pipelines.
Where BioMate fits in the comparison
The survey's honest state is that hybrids beat scale and no one has a true virtual cell. BioMate's differentiators match that reality: an execution engine for mechanistic + ML cell workflows, a rigorous baseline/eval harness, and a statistics-first challenge strategy aligned with what actually won in 2025 — rather than a headline model claim it can't back.
Honest boundaries
BioMate does not claim a trained perturbation model that beats SOTA, nor a whole-cell simulator; several ML perturbation workflows are staged, not shipped. The point of the survey — and this post — is to be clear about that.
Go deeper: the full survey
This showcase is one half of a pair. For the complete, verified-citation map of the field — every method, honest strengths and limits, and what is actually winning — read the Virtual Cell Modeling survey.