Skip to main content

HubXchange

HubXChange logo

News Xchange

Page Banner 1

Zero-shot, drug-like antibody design: why Latent-X2 could reshape biologics R&D

Biologics discovery has long lived with a structural mismatch: we can generate binders at scale, but we still spend disproportionate time and money converting binders into drugs. Affinity is rarely the only constraint. Developability, manufacturability, and immunogenicity remain persistent sources of attrition, rework, and late-stage surprises—especially as the industry pushes into more complex formats (multispecifics, VHHs, engineered scaffolds) and harder targets.

Latent-X2 is notable because it explicitly targets that mismatch. In a recent preprint, the Latent Labs team describe an all-atom generative model that jointly designs sequence and structure in the bound complex, conditioned on target structure, epitope specification, and optionally an antibody framework (Kenlay et al., 2025). Unlike many prior AI systems, Latent-X2 reports zero-shot antibody designs that meet multiple “drug-like” criteria—without iterative optimization, filtering, or selection.

Critically, the same work reports experimental immunogenicity assessment in human donor panels for representative de novo VHH antibodies, showing no detectable T-cell proliferation or cytokine release relative to controls (Kenlay et al., 2025). To our knowledge, this represents the first disclosed human-cell immunogenicity data for antibodies generated entirely de novo by a generative AI system.

Why immunogenicity is the strategic hinge

Immunogenicity remains one of the most consequential—and least predictable—failure modes in biologic therapeutics. Anti-drug antibodies (ADAs) can neutralize efficacy, accelerate clearance, or trigger adverse immune responses, even for fully human or humanized antibodies (Vaisman-Mentesh et al., 2020). Regulatory and industry experience increasingly recognize immunogenicity as a multifactorial risk, influenced not only by sequence “foreignness,” but also by aggregation, polyreactivity, formulation, impurities, and dosing strategy (Carter & Quarmby, 2024).

Historical failures underscore the cost of misjudging this risk. The bococizumab program, for example, demonstrated how strong target engagement can be undermined by clinically significant immunogenicity (Ridker et al., 2017). As a result, immunogenicity is now viewed less as a niche safety concern and more as a core determinant of clinical viability.

Against this backdrop, the Latent-X2 results are strategically significant. Rather than relying solely on in silico immunogenicity prediction, the authors subjected AI-generated antibodies to ex vivo human PBMC assays, a commonly used translational proxy in biologics development (Kenlay et al., 2025). While such assays do not replace in vivo or clinical data, they represent a materially higher bar than computational heuristics alone.

From binding hits to lead-like molecules

The deeper claim made by Latent-X2 is that developability emerges directly from generation. Across VHH and scFv formats, the reported designs show favorable profiles for expression yield, aggregation propensity, hydrophobicity, polyreactivity, and thermal stability—benchmarked against approved and clinical-stage antibodies (Kenlay et al., 2025).

This matters because the field already has empirical definitions of what “clinical-like” antibodies look like. Large-scale analyses of clinical-stage antibodies have shown that successful molecules cluster within relatively narrow biophysical boundaries (Jain et al., 2017). Many discovery-stage binders fail not because affinity is insufficient, but because they fall outside these boundaries and cannot be rescued without sacrificing potency or specificity.

If generative models can internalize these constraints during design, they change the role of downstream optimization—from salvage to refinement.

Precision design versus brute-force screening

The macrocyclic peptide results reported alongside the antibody work further reinforce this shift. The authors show that Latent-X2-designed macrocycles can match or exceed the binding affinities of peptides discovered through trillion-member mRNA display libraries—while experimentally testing only ten designs per target (Kenlay et al., 2025).

This finding resonates with a broader transition underway in molecular discovery. Diffusion-based and structure-conditioned generative models have already demonstrated the ability to design functional proteins and binders de novo (Watson et al., 2023). In parallel, advances in structure prediction for complexes, such as AlphaFold 3, have lowered the barrier to structure-guided design at scale (Abramson et al., 2024).

For macrocyclic peptides—an increasingly attractive modality for targeting protein–protein interactions and intracellular targets—this raises a fundamental question: how much combinatorial scale is truly necessary when generative precision improves? The success of macrocycles progressing toward clinical relevance, including orally bioavailable examples, suggests this is not merely an academic consideration (Johns et al., 2023).

What this means for Pharma in 2026–2030

Discovery funnels will compress, shifting the critical path downstream

As zero-shot design yields candidates that are simultaneously potent, developable, and lower risk, the traditional early funnel (hit → triage → rescue optimization) will narrow. Competitive advantage will increasingly depend on mechanistic validation, translational relevance, and clinical strategy, rather than sheer hit-finding capacity.

Immunogenicity will become a first-class design constraint

Between 2026 and 2030, immunogenicity is likely to be treated on par with affinity and stability during early design. Systems that bias sequences away from immunogenic motifs at generation time—supported by human-relevant assays—will reduce late-stage risk and portfolio volatility (Carter & Quarmby, 2024).

Modality optionality will expand

A single generative architecture spanning VHHs, scFvs, and macrocycles supports rapid modality switching for a given target. This flexibility will be especially valuable for historically difficult targets, enabling parallel exploration of extracellular, shallow-surface, and intracellular strategies.

R&D operating models will evolve

Expect smaller, AI-native discovery teams paired with high-leverage validation pods focused on expression, biophysics, and translational assays. The bottleneck shifts from generating candidates to deciding which ones deserve to move forward.

Platform differentiation will hinge on feedback loops, not models alone

By the late 2020s, competitive moats will be defined less by access to generative models and more by proprietary experimental feedback—especially consistent developability and immunogenicity datasets that continuously refine design priors.

Regulatory scrutiny will increase—but may ultimately accelerate adoption

As AI-designed biologics enter the clinic, regulators will focus on control strategies, comparability, and traceability rather than algorithmic novelty. Platforms that demonstrate disciplined design governance and human-relevant risk mitigation may ultimately de-risk regulatory engagement rather than complicate it.

 

References

View all News Xchange
Loading
Search