Google sign-in. No card required.
Latch builds rigorous scientific benchmarks (benchmarks.bio http://benchmarks.bio) for AI agents across biology. We work with frontier labs and pharma to measure whether agents can handle real biological workflows: the messy, multi-layer reasoning that matters in the field.
We're seeking an AI Biologist to join our Protein Design & Engineering team, working at the frontier of what artificial intelligence can achieve in biology.
ABOUT THE ROLE
AI agents can now run the full protein design toolkit: generate backbones, design sequences, predict complexes, score interfaces and propose mutations. The open question is whether they make good decisions with it. We are building benchmarks that test exactly that, from designing a new binder from scratch to improving an existing one toward a measured function.
You will work across three connected areas:
- De novo design. Given a target and a goal, can an agent design, filter and choose candidates that actually work? Tasks test the decisions that separate a real hit from a confident-looking failure: which design tool to trust, which filter to apply, which candidates to send to the lab.
- Protein–protein interactions. Can an agent reason about a complex correctly? Tasks test whether it finds the true interface and its hotspots, tells a real interface from a crystal contact, reconciles co-folding models that disagree, and predicts how mutations change binding affinity and specificity.
- Optimization. Given a de novo or redesigned protein, can an agent improve affinity, stability, expression or developability without breaking something else?
Every task is graded against real measured data: binder screens, binding affinities, deep mutational scans, combinatorial landscapes, developability panels and published engineering campaigns. Your job is to find the decision point in that data and turn it into a question with one verifiable answer.
The work sits under our biosecurity program. You will build tasks on therapeutics such as antibodies, nanobodies, minibinders and enzymes. You will also help test whether agents correctly decline to design or improve pathogen and toxin proteins.
WHAT YOU'LL DO
Build the task
- Find measured design, binding and engineering data (binder screens, Kd titrations, SPR or BLI, DMS, combinatorial libraries, developability assays, directed-evolution lineages) and turn real design decisions into file-based tasks with automatic graders
- Write gold-standard answers from raw data, with success defined against replicate noise and measured hit rates
- Write Python pipelines that compute answer keys and check that an honest design route can reach them
Set the hard tasks
- Build interface tasks: hotspot calls, interface versus crystal contact, specificity against off-targets, and affinity changes on mutation
- Cover the full workflow: generate, filter, rank, combine, revert, triage pools and trade off competing objectives
- Set floors (the tool alone, the best single metric, the screen's base hit rate) so an agent has to beat its own tools
Judge the agent
- Read agent reasoning and code step by step, and find where a plausible design choice goes wrong
- Reward calibrated decisions over confident ones
REQUIREMENTS
- PhD (or equivalent experience) in protein design, protein engineering, structural biology, biophysics, computational biology or a related field
- Hands-on design or engineering in the last ~2 years, with experimental validation of your own designs: de novo binders or enzymes, antibody engineering, directed evolution, DMS, library design or ML-guided optimization
- Working knowledge of protein–protein interfaces: hotspots, shape and charge complementarity, affinity and specificity, and how binding is measured (SPR, BLI, ITC, display screens)
- Working knowledge of modern design and structure tools such as RFdiffusion, ProteinMPNN, BindCraft, AlphaFold, Boltz, Chai or ESM, and a clear sense of where each one fails
- Deep understanding of a