Latent-Y: A Lab-Validated Autonomous Agent for De Novo Drug Design

Abstract

Drug discovery relies on iterative expert workflows that are slow to parallelize and difficult to scale. Here we introduce Latent-Y, an AI agent that autonomously executes complete antibody design campaigns from text prompts, covering literature review, target analysis, epitope identification, candidate design, computational validation, and selection of lab-ready sequences. Latent-Y is integrated into the Latent Labs Platform, where it operates in the same environment as drug-discovery experts with access to bioinformatics tools, biological databases, and scientific literature. The agent can run fully autonomously end-to-end, or collaboratively, where researchers review progress, provide feedback, and direct subsequent steps. Candidate antibodies are generated using Latent-X2, our frontier generative model for drug-like antibody design. We demonstrate the agent's capability across three distinct campaign types: epitope discovery guided by therapeutic specifications, cross-species binder design, and autonomous design from a scientific publication targeting human transferrin receptor for blood-brain barrier crossing. Across nine targets, Latent-Y produced lab-confirmed nanobody binders against six, achieving a 67% target-level success rate with binding affinities reaching the single-digit nanomolar range, without human filtering or intervention. In user studies, experts working with Latent-Y completed design campaigns 56 times faster than independent expert time estimates, compressing weeks of work into hours. Because Latent-X2 is a general-purpose atomic-level model for biologics design, the same agent architecture naturally extends to macrocyclic peptide and mini-binder design campaigns, broadening autonomous discovery across therapeutic modalities. Latent-Y is available to selected partners at https://platform.latentlabs.com.

Figures (6)

• Figure : Fig. 1 |Latent-Y autonomously designs nanomolar-affinity antibodies from text prompts, accelerating expert workflows by over 50-fold.(a) Three independent VHH design campaigns, each initiated from a single natural-language prompt targeting , , and respectively. Designed structures of the top binder--target complex are shown alongside binding interface detail, with insets highlighting designed non-covalent interactions at the binding interface (pink dashed lines). All designs were experimentally characterized by . (b) Per-target experimental hit rates across all successful targets. cTNFL9, cynomolgus ; hTNFL9, human ; xrTNFL9, cross-reactive . (c) Estimated time required to complete a full computational protein design campaign, comparing expert alone versus expert assisted by Latent-Y. Expert baseline times were obtained by polling independent protein designers across academia and industry ($n = 10$); Latent-Y-assisted times reflect the average across agent-assisted runs ($n = 5$). Error bars indicate minimum and maximum values. (d) Time breakdown by major design stage, including literature review and search, structural analysis and epitope selection, computational binder generation, and quality assurance and final selection. Error bars indicate minimum and maximum values.
• Figure : Fig. 2 |Biophysical characterization of the best Latent-Y-designed VHHs against , , and , with affinities reaching the single-digit nanomolar range. Top-performing de novo VHH binders ranked by binding affinity, with corresponding designed bound structures and response curves. Binding affinities were measured using five analyte concentrations with kinetic fitting to determine $\mathrm{K}_{\mathrm{D}}$, see \ref{['sec:methods_spr']}. Reported $\mathrm{K}_{\mathrm{D}}$ values span the nanomolar range, with lower $\mathrm{K}_{\mathrm{D}}$ values corresponding to stronger binding.
• Figure : Fig. 3 |Latent-Y autonomously designs binders to to disrupt the / complex formation. A condensed trace of the full Latent-Y campaign that produced the lab-confirmed binders reported in this work, spanning close to 10000.0 lines of reasoning in its entirety. The trace shows the agent executing a high-level user prompt (grey box), with tool uses and subagent calls highlighted (purple boxes). Key stages include database search and target identification, hotspot analysis via a dedicated subagent, iterative pilot and scale-up batches, and a final quality assurance to select lab-ready candidates. Ellipses denote reasoning steps omitted for visual brevity.
• Figure : Fig. 4 |Latent-Y designs cross-species VHH binders through autonomous capability extension. Designed structures of the de novo VHH complexed with human (left) and cynomolgus (right) , with divergent mutations between species highlighted in orange. The VHH is shown bound to the trimeric unit. identified hits cross-reactive binding to both human and cynomolgus targets. To execute this campaign, Latent-Y developed a custom generative method from a one-line natural language description, with expert involvement limited to biological steering and logic verification.
• Figure : Fig. 5 |Latent-Y collaboratively designs cross-species binders through autonomous capability extension. A condensed trace of the full Latent-Y campaign that produced the lab-validated cross-species binders reported in this work, spanning over 30000.0 lines in its entirety. User prompts and interventions are shown in grey boxes, with tool uses and subagent calls highlighted in purple. Key stages include cyno structure prediction and alignment, hotspot analysis via a dedicated subagent, development of a custom generative method, and iterative wave-based exploration with intermediate findings summaries. Ellipses denote reasoning steps omitted for visual brevity.