VHH developability model predicting five biophysical endpoints from nanobody sequence, using the ESMC protein language model with conformal bounds.
No providers recorded yet. Browse all providers
A nanobody is a single immunoglobulin domain of about 120 residues, and the compactness that makes it attractive — tissue penetration, cheap microbial expression, easy fusion into multispecific formats — also means it has none of the paired-chain interface that stabilizes a conventional antibody. Developability therefore behaves differently: a VHH can express well and still run as a dimer on size-exclusion chromatography, or migrate at the wrong apparent mass on CE-SDS because an unpaired cysteine found a partner. Models trained on monoclonal antibody assay panels do not transfer to that regime, because the sequence space and the failure modes are both different.
NanoLab is the model Eli Lilly and Company built for it, released on 6 October 2026 in Lilly TuneLab's one-year-anniversary release. It takes a VHH heavy-chain amino acid sequence and returns five biophysical developability readouts, each with an uncertainty interval. Rather than learn protein sequence statistics from scratch on a dataset far too small for it, NanoLab sits on top of ESMC, the protein language model released by Biohub, and learns the mapping from those representations to measured developability. The supervision comes from a dataset generated specifically for this purpose by BigHat Biosciences on its Milliner platform.
That division of labor is the point. The scarce ingredient in developability modeling is not sequence data but physical measurement, and single-domain measurement at scale barely existed outside internal archives. NanoLab is the companion to Lilly's paired-chain model AbLab — whose documentation listed VHH models as forthcoming — and covers the format AbLab explicitly does not accept.
NanoLab is documented by Lilly as a developability model for camelid single-domain antibodies built on ESMC and trained on a purpose-built VHH dataset of approximately 2,500 sequences characterized across the five predicted endpoints. BigHat describes generating that dataset over six one-week design rounds on Milliner, a cell-free protein synthesis and automation platform, reaching roughly 90% field-level completeness across yield, HPLC-SEC monomeric purity, CE-SDS purity and thermal properties, with replicate coefficients of variation of 2–5% on nearly all metrics. Sequences that expressed too poorly for downstream assays were flagged rather than dropped, keeping negative examples in the training set. The head architecture, parameter count, ESMC checkpoint size and held-out accuracy are undisclosed, and no paper, code or weights accompany the release. Out-of-range sequence lengths are accepted but fall outside the training distribution.
The stated use is relative ranking during lead optimization: given a panel of VHH candidates from an engineering campaign, prioritize which ones to express, purify and characterize. Because yield, monomer purity and apparent mass are exactly the readouts that decide whether a single-domain construct survives into a multispecific or a CAR, a sequence-stage estimate removes a make-and-measure cycle from the triage loop. Lilly positions the predictions as advisory rather than a substitute for wet-lab confirmation, and not for regulatory submissions. The intervals matter operationally here: a candidate with a wide bound is a candidate to measure rather than to discard.
NanoLab's significance is that single-domain developability data of this kind has not been publicly modelable, and the entity that generated it was a specialist CRO-like platform rather than the pharmaceutical company that trained on it — a division of labor TuneLab has since extended with further data-generation partners. Reach is through membership: TuneLab counted 135 or more member companies at the time of release, with the models available through Benchling and Revvity and subsequently Collaborative Drug Discovery's CDD Vault. The limits are the same as for the rest of the platform and worth stating plainly. There is no publication, no public benchmark and no released artifact, so the model's accuracy rests on Lilly's own description; access is conditional on TuneLab membership and its data-contribution terms; and with five endpoints drawn from one partner's assay suite, the panel is narrower than the biophysical profile a development candidate ultimately has to clear.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.