Original Authors: Latent Labs Team (London & San Francisco)
Published: Technical Report, March 23, 2026

Abstract

Latent-Y, released by Latent Labs, is the first AI agent capable of autonomous end-to-end biologics design from natural language prompts. The system integrates literature retrieval, target structure analysis, epitope selection, candidate molecule generation, and quality assessment modules, based on their frontier generative model Latent-X2 for antibody design.

In wet-lab validation across 9 targets, 6 targets successfully produced nanomolar binders, achieving a 67% success rate. User studies show that experts using Latent-Y can compress design cycles from weeks to hours, achieving 56x efficiency improvement. This system marks a paradigm shift in drug discovery from the "molecular design" bottleneck to the "scalable execution" bottleneck, though its large-scale applicability and failure mode analysis remain to be further validated.

Key Metrics

  • Target-level success rate: 67% (6/9 targets)
  • Best affinity: PRL binder 5.44 nM, IL-6 binder 12.5 nM
  • Efficiency improvement: 56x acceleration (from 2 weeks to ~5 hours)
  • Literature review acceleration: ~4,300x
  • Structure analysis acceleration: ~350x

1. Background: The New Bottleneck in Drug Discovery

Traditional drug discovery relies on iterative expert workflows—a process that is slow and difficult to parallelize. In recent years, frontier AI models such as Latent-X2 and RFdiffusion have demonstrated the feasibility of zero-shot biologics design, capable of directly generating antibody and peptide candidates with drug-like properties.

This advance has shifted the bottleneck in early drug discovery from "finding candidate molecules" to a new level: the execution bandwidth of drug discovery organizations, and the accessibility of PhD-level domain experts required for scaling.

In this context, Latent Labs developed Latent-Y—an agent system for de novo drug design. Unlike single molecule generation models, Latent-Y operates in the same environment as human experts, with access to bioinformatics tools, biological databases, and scientific literature. It can autonomously execute the complete workflow from research objectives to lab-ready candidates, or operate in collaborative mode with researchers who review progress, provide feedback, and guide subsequent steps.

2. Technical Architecture and Methods

2.1 Hierarchical Collaborative Architecture

Latent-Y employs a hierarchical architecture design, complementing its foundation generative model Latent-X2. While Latent-X2 reasons at the atomic level, designing precise molecular interactions, Latent-Y reasons at the expert level, planning the complete path from research objectives to lab-ready candidates. This division allows the system to simultaneously handle the physicochemical details of molecular design and the biomedical logic of drug discovery strategy.

2.2 Workflow

Latent-Y's workflow encompasses six core stages:

In the epitope identification stage, the system analyzes target surfaces through spatial reasoning to identify candidate epitopes meeting functional criteria; in the generation stage, it calls Latent-X2 to produce candidate molecules, dynamically adjusting generation parameters, modifying epitope constraints, or switching design modalities based on intermediate results.

2.3 Autonomous Capability Extension

Notably, when standard functions are insufficient for specific design challenges, Latent-Y can generate custom computational methods based on natural language descriptions. In a cross-species binder design task, the system autonomously implemented and executed cross-species generation code through a single natural language instruction ("design a joint generation method ensuring candidates satisfy constraints for both human and cynomolgus monkey"), demonstrating flexibility in tool use and capability extension.

3. Wet-Lab Validation Results

3.1 Target-Level Success Rate and Binding Affinity

Among 9 test targets, Latent-Y successfully produced experimentally validated nanobody binders for 6 targets, achieving a 67% success rate. Single-target hit rates (confirmed by high-throughput surface plasmon resonance screening) ranged from 1% to 17%.

In terms of affinity, the best binders reached single-digit nanomolar levels:

All high-affinity binders were screened from less than one plate (~88 designs).

3.2 Efficiency Improvement Assessment

User studies compared the time required for independent experts versus Latent-Y-assisted experts to complete full computational design cycles. Baseline data came from self-reports of 10 independent PhD-level protein designers (median experience 9.8 years).

Results show that Latent-Y compressed an average of two weeks of expert work to approximately 5 hours, achieving ~56x acceleration. Maximum acceleration appeared in reasoning-intensive stages:

This efficiency improvement has multiplicative effects when running multiple activities in parallel.

3.3 Cross-Species Design Case

Cross-species reactivity is a common requirement in drug development but faces multiple challenges: human and cynomolgus monkey TNFL9 have 11 mutations (~5%), no experimental structure is available for cynomolgus monkey, and the target exists as a trimer with a relatively flat surface.

Latent-Y demonstrated the value of human-AI collaborative mode in this task: researchers provided high-level biological guidance (such as identifying and correcting a "reward hacking" issue caused by structure cropping), while the agent handled method implementation and iterative optimization. Ultimately, 3 out of 40 designs were confirmed as dual-reactive binders.

4. Discussion

4.1 Technical Contribution and Significance

Latent-Y represents significant progress in drug discovery automation, with its core value extending AI's role from "molecule generator" to "autonomous research executor." Previously, zero-shot molecular design models had proven feasibility, but integrating such capabilities into agents that can autonomously execute complete research workflows and validating them in wet labs is unprecedented.

Paradigm Shift

The 67% target-level success rate and 56x efficiency improvement indicate that the bottleneck in early drug discovery is shifting from "molecular design capability" to "research execution bandwidth." This paradigm shift may have profound implications for the operational models of drug discovery organizations. As discussed in our previous article, the early AIDD agent era will dramatically compress Design and Analysis cycles, while Making and Testing will become the primary constraints.

4.2 Limitations and Open Questions

Despite encouraging results, this study has several aspects requiring further validation:

4.3 Competitive Landscape Positioning

Compared to fixed generation pipelines (such as pure RFdiffusion or AlphaFold-derived workflows), Latent-Y's differentiated advantage lies in its adaptive decision-making capability—dynamically adjusting strategies, switching modalities, or generating custom tools based on intermediate results. This resonates with the "society of thought" concept proposed in a recent Science commentary: Latent-Y can be seen as an agent system implementing multi-step reasoning and self-correction in the drug discovery domain.

However, compared to widely validated structure prediction tools like AlphaFold, Latent-Y's independent validation data remains limited, and its reliability requires more third-party replication experiments for confirmation.

5. Conclusion

As the first wet-lab-validated autonomous drug design agent, Latent-Y demonstrates end-to-end design capability from text prompts to nanomolar binders. Its 67% target-level success rate and 56x efficiency improvement indicate that AI's role in drug discovery is evolving from assistive tool to autonomous executor.

However, this technology remains in early validation stages. Larger-scale target diversity testing, systematic failure mode analysis, long-term developability characterization, and deep integration assessment with existing drug discovery workflows will be key to determining its industrial value and sustainable competitive advantage.

For the drug discovery field, Latent-Y's emergence hints at a possible future: researchers' roles will shift from "personally executing designs" to "setting objectives and supervising agents," and the implications of this shift for talent development, organizational structure, and regulatory frameworks merit continued observation.

References

← Back to Blog