Advisor - Antibody Developability Validation & Benchmarking

Purpose

Lilly Tune Lab is an AI-powered drug discovery platform that provides biotech companies with access to machine learning models trained on Lilly's extensive proprietary pharmaceutical research data. Through federated learning, the platform enables Lilly to build models on broad, diverse datasets from across the biotech ecosystem while preserving partner data privacy and competitive advantages. Antibody develop ability prediction is a core workstream within Tune Lab — covering aggregation, self‑association, polyspecificity, thermal stability, viscosity, and chemical liabilities — that gates progression from discovery into lead optimization, cell line development, and formulation.

The Advisor/Senior Advisor – Antibody Develop ability Validation & Benchmarking plays an essential role in establishing whether Tune Lab's federated antibody models can be trusted to triage real candidates. The person in this seat must understand, at depth, how antibodies are characterized, what makes a sequence developable or not, and how predictions from a federated model translate into go/no‑go decisions in a discovery pipeline.

This is a validation‑led role that contributes to model design choices and partners closely with antibody modeling scientists on architecture, feature design, and uncertainty quantification.

• Antibody Develop ability Benchmark Suite: Build the canonical benchmark suite covering the full develop ability portfolio — aggregation propensity (AC‑SINS, SMAC, CIC), thermal stability (nanoDSF/DSF), polyspecificity (BVP‑ELISA, Heparin RT, PSR), self‑interaction, viscosity, chemical liabilities (deamidation, isomerization, oxidation, N‑glycosylation in CDRs), and immunogenicity surrogates. Define which endpoints are evaluated jointly versus independently and how multi‑endpoint reliability rolls up to a triage decision.
• Sequence‑Aware Federated Test Set Design: Architect privacy‑preserving protocols for constructing representative test sets across distributed partner datasets, with splitting strategies appropriate to antibody data — germline‑based, CDR‑similarity‑based, and clonotype‑based splits that genuinely test generalization rather than near‑duplicate memorization. Account for the structural asymmetry of antibody data (many sequences with shallow characterization, few sequences with deep characterization) when designing held‑out evaluation sets.
• Public Benchmark Integration: Systematically benchmark federated antibody models against established external resources — SAbDab, OAS, TAP, the Jain et al. clinical‑stage antibody panel, FLAb, and equivalent emerging datasets — to characterize generalization gaps and quantify where federated training delivers measurable lift over public‑only baselines.
• Cross‑Domain Validation: Develop validation strategies that assess model generalization across modalities and formats relevant to antibody develop ability — IgG vs. bispecific vs. fragment formats, different expression systems, different assay protocols across partners — while respecting partner data boundaries.
• Validation Frameworks: Implement temporal‑split and sequence‑similarity‑aware validation protocols that simulate prospective deployment, detect concept drift as partner data accumulates, and surface systematic failure modes across CDR length distributions, germline families, and physicochemical regimes.
• Model Design Partnership: Work alongside antibody modeling scientists on architectural and feature choices that have direct validation implications — uncertainty quantification approaches, calibration strategies, structure‑aware vs. sequence‑only representations, and how predictions from different endpoints should be combined or kept independent.
• Statistical Rigor: Design statistically powered validation studies that account for multiple testing across endpoints, hierarchical structure in antibody data (sequences clustered by germline, project, partner), and non‑independent observations. Provide honest confidence intervals on reported model performance.
• Reproducibility Infrastructure: Build robust MLOps pipelines ensuring complete reproducibility of federated experiments,…