Latent-X2

Latent-X2 is a frontier AI model developed by Latent Labs for designing high-affinity antibodies with drug-like properties, as well as macrocyclic peptides. It is designed for pharmaceutical and biotech researchers who need to generate developable, low-immunogenicity therapeutic candidates from the very first generation—without the lengthy optimization cycles that characterise traditional drug discovery. The model jointly generates all-atom structures and sequences conditioned on multi-modal prompts, enabling precise control over the design process across multiple biological modalities.

Traditional drug discovery faces a critical bottleneck not in screening, but in development: hits rarely possess the properties needed for clinical success, and efforts to address one shortcoming frequently compromise another. Latent-X2 addresses this by producing molecules that not only bind with high affinity but can clear developability and immunogenicity hurdles from the outset, reducing the risk of costly downstream failure in clinical programs.

Supported Molecule Formats

• Nanobodies (VHHs): Single-domain antibodies derived from camelids, valued for their small size and ability to access epitopes inaccessible to conventional antibodies. Latent-X2 generates VHH binders with nanomolar affinities in lab experiments.
• Single-Chain Variable Fragments (scFvs): Combining the variable regions of antibody heavy and light chains in a single polypeptide, scFvs offer flexibility for therapeutic and diagnostic applications. Latent-X2 generates scFv binders with down to low picomolar binding affinities.
• Macrocyclic Peptides: Small cyclized proteins with potential oral bioavailability. Latent-X2 generates macrocyclic peptides that bind K-Ras—long considered undruggable—matching or exceeding the affinity of hits from trillion-scale mRNA display screens while testing 11 orders of magnitude fewer sequences.

Drug-Like Properties by Default

• Designed antibodies exhibit developability profiles—including expression yield, biophysical characteristics, and polyreactivity—comparable to approved therapeutics in head-to-head comparisons.
• These properties arise from the first generation, without optimization, filtering, or selection.
• In the first such assessment of any AI-generated antibody, de novo VHH binders targeting TNFL9 were evaluated across a ten-donor human panel in ex vivo T-cell activation and cytokine release assays, confirming both potent target engagement and low immunogenicity.
• This represents the first time an AI-generated antibody has been shown to clear immunogenicity assessment—a critical milestone for clinical translation.

Key Distinctive Features

• Drug-like from first generation: Developability and low ex vivo immunogenicity arise without optimization, filtering, or selection.
• First low immunogenicity demonstration: De novo antibody designs confirmed across human donor panels in ex vivo assays.
• Broad modality coverage: VHH, scFv, and macrocyclic peptide formats from a single model.
• Hits on difficult targets: Including historically challenging and undruggable targets, with affinities matching trillion-scale screens.
• Lab-compatible efficiency: Only 4 to 24 designs per target are needed, versus millions or trillions in conventional screening.
• Precise model steering: Condition on target structure, epitope specification, and antibody framework of choice.
• All-atom generation: Joint sequence and structure generation modeling non-covalent interactions in a single step.

How Latent-X2 Works

• Latent-X2 is an all-atom generative model that jointly generates sequences and structures while modeling the bound complex.
• By directly modeling non-covalent interactions, the model respects the biochemistry of binding and function at the atomic level.
• The model is conditioned on multi-modal prompts: the target structure defines what to bind, the epitope specification defines where to bind, and an optional antibody framework allows following a preferred scaffold.
• This enables precise control over the design process across different biological modalities in a single generative step.

Validation and Reproducibility

• Latent-X2 was validated across targets selected for diversity and difficulty, demonstrating high-affinity binders across multiple modalities.
• All results presented are preclinical; animal studies and clinical trials remain ahead.
• Immunogenicity assessment was conducted ex vivo using human donor panels—a state-of-the-art proxy for clinical immunogenicity risk.
• To support reproducibility, the sequence and structure of lab-validated antibody designs for each target are published on the Latent Labs Platform, accessible without sign-in.

Latent-X2 is available via the Latent Labs Platform, accessible through a web browser or via the Latent Labs API for system integration. The platform provides scientist-friendly, no-code workflows and builds on the success of Latent-X1, which has been adopted by industry and academic groups worldwide. Latent Labs is backed by a $50M funding round co-led by Radical Ventures and Sofinnova Partners, and the team includes former AlphaFold 2 co-developers and ex-DeepMind team leads. Early access is available by application at platform.latentlabs.com.