About the role

About CellType

CellType is building foundation models and agent systems for biology.

We train models to reason over cells, tissues, and patients: predicting how a perturbation, a drug, or a disease state changes biology, and how findings in cell lines and animals translate to humans. We work with pharma, biotech, and diagnostics partners on live problems such as preclinical-to-clinical translation, cross-species toxicology, perturbation and response prediction, cell-type deconvolution from clinical assays, and spatial biology.

We are building the core intelligence layer for biology.

About the role

We are hiring a Founding Research Scientist to own the modeling in our biology programs end to end.

This is the role that turns a partner's scientific question into a model task, a dataset, an evaluation, and a result a pharma team will act on. You will work directly with the founders on the projects that define the company over the next year, and you will have a lot of room to decide how we model biology, not just implement someone else's plan.

We care about two things in equal measure: you have trained models at scale before, and you understand biology well enough to know when a result is meaningful versus an artifact. If you come from a different modeling domain (vision, multimodal, speech, protein) and have since worked seriously with biological data, that is exactly the profile we are looking for.

What you'll do

• Own modeling for one or more live pharma and biotech programs: framing the problem, choosing the representation and architecture, building the training and evaluation loop, and delivering results to the partner
• Train, fine-tune, and post-train biological foundation models on single-cell, perturbation, spatial, multi-omic, imaging, and clinical assay data
• Build models of cellular state and response: perturbation prediction, in-silico screening, cross-species and cell-line-to-patient translation, deconvolution, and related virtual-cell tasks
• Design evaluations that biologists trust: held-out perturbations, unseen cell types and tissues, cross-species and cross-cohort generalization, and wet-lab-in-the-loop validation with our partners
• Decide what data to acquire or generate next, and shape the data-generation experiments we run with academic and pharma collaborators
• Turn one-off project results into reusable model capabilities and product workflows
• Present results to partner scientists and translational teams, and defend the modeling choices behind them
• Work alongside our model-training and platform engineers so that research ideas become reliable, scalable systems

You may be a fit if you

Modeling depth

• Have trained or materially improved large models yourself, in any domain: vision, multimodal, language, speech, protein/molecule, or biology. You know what it takes to get a large training run to converge, what breaks at scale, and how to diagnose it
• Have strong applied ML judgment: representation choices, tokenization or encoding of non-text data, loss design, pretraining vs. fine-tuning vs. post-training trade-offs, and rigorous evaluation
• Are fluent in Python and PyTorch or JAX, and comfortable running experiments on multi-GPU infrastructure
• Have shipped models that people relied on in production or in a live program

Biology fluency

• Have worked hands-on with biological data such as scRNA-seq, Perturb-seq / CRISPR or compound screens, spatial transcriptomics, bulk or cell-free omics, histopathology, or high-content imaging, and understand its noise, batch effects, and failure modes
• Understand the biology behind the data well enough to ask whether a prediction is plausible, to choose sensible baselines and controls, and to talk credibly with pharma scientists
• Have followed or contributed to the virtual-cell / perturbation-modeling literature (single-cell foundation models, perturbation response prediction, in-silico screening) and have opinions about what works and what does not

Working style

• Want to own a project from ambiguous question to partner-facing result
• Are comfortable in a small team where priorities move toward whatever matters most this week
• Communicate clearly with both ML engineers and biologists

We'd be especially excited if you also have

• Trained vision or imaging models at scale and have since applied that to biology (cell painting, histopathology, spatial, microscopy)
• Built or evaluated a virtual-cell or perturbation-prediction model against held-out experiments
• A PhD or equivalent research record in computational biology, ML for biology, or a quantitative field applied to biology
• Worked with pharma or biotech partners and understand what a translational or discovery team needs from a model result
• Experience with cross-species, cell-line-to-patient, or preclinical-to-clinical translation problems
• Experience with LLM post-training (SFT, RL, tool use) applied to scientific reasoning

This role will shape what CellType's models can do. The right person will determine not only how well we model biology, but which problems we choose to solve and how quickly our partners see results they can act on.

This role is based in New York City.

If you have trained models at scale and want to point that skill at biology's hardest prediction problems, we'd love to talk.

About CellType

CellType is building foundation models and agent systems for biology.

We are building models that learn from biological data, experiments, and outcomes directly. Our goal is to develop AI systems that are genuinely useful for scientific discovery and drug development.

We work on problems such as preclinical-to-clinical translation, response prediction, biomarker discovery, and reasoning across complex biological datasets. These are technically difficult problems with real scientific and commercial importance.

We are an early team with high standards for technical depth, speed, and ambition. This is a good fit for people who want to work on hard model and systems problems in a domain that matters.