Member of Technical Staff (Epigenetics, Therapeutics, Single-cell)

LatchBio•San Francisco, CA

3d•$200,000•Onsite

About The Position

LatchBio is building AI agents that do real biology. Agents that can take raw sequencing data, run the right analysis, recognize when something is wrong, and return a defensible scientific claim. The hard part isn't model training. It's defining what good analysis actually looks like across the assays and decisions that matter in real R&D, and building benchmarks that can tell when an agent is doing it right. In this role, you will translate tacit workflow knowledge into structured, reproducible analysis patterns with clear ground truth and known failure modes, review agent outputs and engineer-built pipelines, catch what looks right but isn't, decide what an acceptable analysis looks like, generate reasoning traces, contribute training data, and iterate with the model team on where agents fall down. You will work directly with our CTO, head of product, and a small bioinformatics team. You will not be siloed. You will not be writing slides.

Requirements

3-5+ years of hands-on primary analysis experience in epigenetics, therapeutics R&D, or single-cell genomics
Deep domain expertise, not generalist
Fluent in tools like MACS, ArchR, Signac, ChromVAR, deepTools (Epigenetics track)
Fluent in tools like scipy curve fitting, MAGeCK, RDKit (Therapeutics track)
Fluent in tools like Scanpy, Seurat, Cell Ranger, scVI, Squidpy (Single-cell track)
Analyzed 3+ datasets in your domain with real consequences (publication, product decision, clinical or regulatory submission)
Built or contributed to open-source tools, internal platforms, or production pipelines used by people other than you
Python or R proficiency
Comfortable in pandas/numpy/scipy

Nice To Haves

Span two domains

Responsibilities

Translate the tacit workflow knowledge in your head into structured, reproducible analysis patterns with clear ground truth and known failure modes
Review agent outputs and engineer-built pipelines. Catch what looks right but isn't. Decide what an acceptable analysis looks like
Generate reasoning traces, contribute training data, and iterate with the model team on where agents fall down
Work directly with our CTO, head of product, and a small bioinformatics team
Personal end-to-end analysis of ATAC-seq, ChIP-seq, CUT&RUN/CUT&Tag, scATAC-seq, multiome, or Hi-C (Epigenetics track)
Explain why a peak opens, not just that it did (Epigenetics track)
Personal end-to-end analysis of preclinical drug discovery data: HTS plates, dose-response curves, CRISPR screens, PRISM/DepMap/GDSC response matrices, ADMET panels, PK/DMPK (Therapeutics track)
Know the difference between a real selective hit and broad cytotoxicity, and between a clean IC50 and a curve that shouldn't be reported (Therapeutics track)
Personal end-to-end analysis of scRNA, snRNA, CITE-seq, spatial transcriptomics (Xenium, Visium, MERFISH, Stereo-seq), or Perturb-seq (Single-cell track)
Have intuition for kit-specific QC thresholds (you know 100K cells from a 10X run means something is wrong) (Single-cell track)
Run a full analysis end-to-end in your domain, from raw platform output to a defended biological or therapeutic claim
Diagnose when an analysis is wrong. Given someone else's notebook or report, you can spot the bad QC threshold, the missing counterscreen, the lineage confound, the over-fit curve, the contaminating doublet. You know what artifacts look like in your domain because you've debugged them
Justify every analytical choice. Why this normalization, why this threshold, why this statistical test, why this filter. You can write the decision down, defend it, and say what would falsify your conclusion
Write reproducible code. Python or R. Comfortable in pandas/numpy/scipy and the standard tools of your domain. Your analyses should run twice and give the same answer
Translate your workflow into something someone else can grade. Given a real analysis task you've done, you can specify the input files, the correct output, the acceptable tolerance, and the common ways an analyst would get it wrong