Latent-X

Latent-X is a state-of-the-art frontier AI model developed by Latent Labs for de novo protein binder design at the all-atom level. Designed for drug discovery scientists and therapeutic developers, it generates lab-functional protein binders across multiple therapeutic modalities — including macrocycles and mini-binders — with breakthrough performance validated in wet lab experiments. The platform is accessible to both AI experts and non-coding scientists, removing the need for dedicated AI infrastructure.

Traditional drug discovery relies on screening millions of random molecules, with hit rates typically well below 1% and experiments costing thousands of dollars over months. Latent-X transforms this paradigm by using precision AI design to solve the geometric puzzle of molecular binding at the all-atom level, directly generating the biochemistry required for high binding affinity and specificity. In extensive wet lab validation across 7 therapeutic targets, Latent-X achieved strong binding affinities — down to the picomolar range — while testing as few as 30–100 candidates per target.

Validated Performance Across Therapeutic Modalities

• Macrocycles: Latent-X generated binders for 100% of three selected benchmark targets, with 91–100% of tested macrocycles confirmed as binders. The strongest binders reached single-digit micromolar binding affinities and demonstrated target specificity, a prerequisite for low off-target effects.
• Mini-binders: Binders were found for 100% of five therapeutically relevant protein targets. The best mini-binders achieved picomolar binding affinities, surpassing the strongest reported results from RFdiffusion and AlphaProteo in head-to-head experimental comparisons. Generated mini-binders were shown to be highly target specific.
• Future modalities: Macrocycles and mini-binders represent the initial capabilities. Latent-X's deep understanding of protein binding opens pathways to additional therapeutic modalities such as nanobodies and antibodies, with partnerships welcomed to explore these expanded applications.

Distinctive Features and Capabilities

• Lab-validated binder hits on every tested target: Latent-X achieved functional binders on every tested target, with 91–100% hit rates for macrocycles and 10–64% hit rates for mini-binders.
• Stronger binding affinities than prior methods: Across all modalities, Latent-X produced a high incidence of strong binding affinities, including picomolar binders that exceeded the performance of designs from other models.
• Lab-validated target specificity: The model directly generates biochemical bonds to selectively bind user-specified epitopes, as confirmed through laboratory validation for both macrocycles and mini-binders.
• No-coding AI platform access: Latent-X is available on the Latent Labs Platform, providing a complete lab-validated workflow — target upload, hotspot selection, binder design, and computational ranking — without requiring coding skills or AI infrastructure.
• Joint protein sequence and structure generation: Latent-X generates all-atom structures by co-sampling sequence and structure simultaneously, outperforming prior methods that generate them sequentially.
• State-of-the-art in silico hit rates: Latent-X significantly outperforms prior methods on computational hit rates for held-out binder targets not seen during training, requiring fewer samples to arrive at lab-viable numbers of binders.
• Fast generation speed: Latent-X generates binders over 10x faster than previous methods, reducing generation time to seconds and enabling rapid computational experimentation. Speed is further improved in batched mode.
• Generalization across therapeutic binder types: The model successfully generates distinct therapeutically relevant binder modalities, with more to come.

How Latent-X Works

• Latent-X is a general-purpose frontier model that creates binders from scratch for unseen or previously untargeted proteins, generalizing beyond nature's repertoire.
• The model generates all-atom binder structures that obey atomic-level biochemical rules, including hydrogen bonds and pi-stacking of aromatic rings.
• Generation proceeds iteratively, placing every atom with precision to solve the geometric challenge of molecular binding.
• The model co-samples protein sequence and structure jointly, rather than treating them as separate sequential steps.

Latent Labs Platform Workflow

1. Upload a target protein structure to the platform.
2. Specify epitopes or hotspots of interest on the target.
3. Run binder design using Latent-X to generate candidate macrocycles or mini-binders.
4. Review computationally ranked designs using in silico success metrics.
5. Visualize predicted structures and overlays before selecting candidates for lab testing.

The Latent Labs Platform is available now at platform.latentlabs.com, with a free tier offering daily-renewing user credits. It supports both commercial and non-commercial use. Currently supporting macrocycles and mini-binders, the platform is built by a team including former AlphaFold 2 co-developers and ex-DeepMind team leads, backed by a $50M funding round co-led by Radical Ventures and Sofinnova Partners. Additional modalities and features are planned for future releases.