AI Biologist - Peptide Discovery

Latch builds rigorous scientific benchmarks (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 Peptide Discovery team, pushing what AI can achieve in therapeutic peptide discovery.

About the Role. You will be the scientific authority on ground truth, testing whether agents can reason like an expert peptide discovery scientist. AI agents today struggle to interpret raw selection, biophysical, and structural data, to separate true hits from artifacts, and to make the calls that move a peptide from target to candidate. The data often misleads: enrichment can be a propagation artifact, protection is not contact, and binding is not function. Knowing what the evidence actually supports is the work. You will work with software engineers and scientists to build the tasks and datasets behind our benchmarks. They span linear, macrocyclic, and constrained peptides against protein targets, including protein–protein interactions, from target tractability to candidate nomination. Your judgment will set the standard agent outputs are held to, including:

Build the task

Define what expert reasoning looks like for a peptide discovery question, and write gold-standard answers with the rationale an experienced peptide scientist would defend

Design realistic scenarios across the discovery arc: target tractability, hit generation, hit confirmation and kinetics, binding-site mapping, functional confirmation, and lead optimization

Curate datasets from real programs: display-selection read counts, SPR/BLI sensorgrams, fluorescence-polarization titrations, affinity-selection MS, HDX-MS, crosslinking-MS, NMR chemical-shift perturbation, dose-response, and permeability and stability data

Build analysis pipelines in Python, from enrichment analysis to kinetic and dose-response fitting, that produce verifiable ground truth with tolerances that hold across replicates and fitting routines

Partner with software engineers to turn your judgment into new workflows

Set the hard calls

Test whether agents distinguish true signal from artifact, such as propagation or expression artifacts enriched across rounds, sensorgrams distorted by aggregation or superstoichiometric binding, or HDX protection mistaken for direct contact

Adjudicate conflicting evidence: reconcile orthogonal binding-site datasets, decide between competitive and allosteric engagement, and identify which experiment would settle the call

Test whether agents weigh the right tradeoffs: binding versus function, potency versus developability, and when added constraint or affinity costs a property that matters more

Judge the agent

Interrogate agent reasoning step by step, and challenge conclusions that sound right but would not survive a project review or lab meeting

Pinpoint where agents go wrong: why a step fails, what an expert would do instead, and what evidence would change the call

Score agent confidence, rewarding explicit uncertainty over overconfident calls on ambiguous data

Requirements

PhD in peptide chemistry, biochemistry, biophysics, structural biology, medicinal chemistry, or a related field

Senior, hands-on peptide discovery experience: you have personally analyzed real program data, not just supervised it

Direct experience with at least one of: display selection (phage, mRNA, ribosome, RaPID, or DNA-encoded libraries), SPR/BLI or FP binding kinetics, binding-site mapping by structural MS or NMR, or peptide developability (PAMPA/Caco-2 permeability, proteolytic and serum stability)

Deep expertise in at least one area (display and selection, biophysics, structural mapping, or peptide medicinal chemistry) and working fluency across the others

Strong command of experimental design, quantitative data analysis in Python or R, and statistics

AI-ready from day one: daily use of Claude, Cursor, or similar agents, with the judgment to spot where model reasoning is technically correct but scientifically unsound