Latent-X: All-Atom Co-Design for Protein Binders

• The paper introduces Latent-X, an end-to-end co-design model that jointly generates binder and target interface geometries at an all-atom resolution.
• It demonstrates superior hit rates and structural diversity compared to backbone-centric systems by optimizing non-covalent interactions directly in the generated complex.
• The model streamlines protein binder discovery by reducing candidate counts, adapting target interface conformations, and accelerating design runtime.

Searching arXiv for the exact named topic and closely related comparison papers. Latent-X is an all-atom generative protein design model for de novo binder discovery that, given a target protein epitope, jointly generates the all atom structure and sequence of the protein binder and target, directly modelling the non-covalent interactions essential for specific binding. It is framed as a departure from screening millions of candidates with sub-1% success rates and from multistage backbone-first pipelines, toward an end-to-end co-design paradigm operating at atomistic resolution. In the reported study, Latent-X is validated on two therapeutically relevant modalities—macrocyclic peptides and mini-binders—with wet-lab campaigns testing as few as 30–100 designs per target and with direct comparisons against AlphaProteo, RFdiffusion, and RFpeptides under matched experimental conditions (Team et al., 25 Jul 2025).

1. Conceptual basis and design philosophy

Latent-X is presented as an “atom-level frontier model” and a “precision AI design” system because it is conditioned on a specified target epitope and directly generates binders intended to engage that site through atomically resolved interface geometry. Its central claim is that binder design should be performed at the level where specificity and affinity are physically determined: side chains, hydrogen-bond networks, packing, and other non-covalent interactions. In this framing, the key limitation of earlier pipelines is not merely sampling inefficiency, but the decomposition of design into separate backbone generation, sequence assignment, and structural validation stages, which can miss favorable sequence–structure–interface combinations (Team et al., 25 Jul 2025).

A defining feature of Latent-X is joint generation rather than post hoc docking or resequencing. The model does not treat the target as a fully fixed all-atom object. Instead, it is conditioned only on the target backbone, and it co-generates an all-atom bound complex in which the binder is newly designed while the target’s interface side-chain rotamers—and sometimes local flexible loop conformations—can adapt. This suggests a design regime in which interface complementarity is optimized in the context of a bound complex rather than imposed after a backbone-only stage.

The paper repeatedly distinguishes this approach from backbone-centric systems such as RFdiffusion and RFpeptides, which require ProteinMPNN resequencing, and from workflows that oversample large candidate sets to compensate for weak joint coordination between fold, sequence, and interface chemistry. Latent-X is therefore best understood as a co-generative binder–target complex model whose novelty lies less in generic generative modeling than in atomistic, end-to-end epitope-conditioned design.

2. Inputs, outputs, and supported design modalities

Latent-X is prompted with a target structure in mmCIF format, hotspot residues defining the desired binding epitope, binder length, and optionally target cropping. At least one hotspot is required. Target cropping is useful because the model has a total context window of 512 residues across target plus binder. A notable implementation detail is that the model receives only the target backbone atoms as input; target side chains are not fixed inputs (Team et al., 25 Jul 2025).

The generated output comprises the binder amino-acid sequence, the binder all-atom structure, and an all-atom bound complex containing both binder and target. The target is not redesigned in the sense of sequence engineering, but its bound-state all-atom conformation—especially at the interface—is allowed to adapt during generation. This is one of the paper’s core distinctions from rigid-target design pipelines.

The system is reported to support multiple therapeutic binder modalities without modality-specific retraining or fine-tuning. The paper experimentally validates two:

• Macrocyclic peptides: lengths 12–18 aa in the reported study.
• Mini-binders: lengths 80–120 aa in the reported study.

For macrocycles, Latent-X can generate cyclic topologies and disulfide-containing cycles, although cysteine-containing macrocycles were removed in the experimental synthesis pipeline for practical reasons. For mini-binders, the model is not restricted to canonical helical bundles; the paper emphasizes that it can generate structurally diverse binders, including complex beta-sheet folds.

3. Generation workflow and in silico filtering

The operational workflow begins with user-specified target structure, hotspots, binder length, and optional cropping, after which Latent-X samples candidate binder–target complexes. Generated designs are then screened using external structure prediction systems—Chai-1 in the main experiments, with Boltz-2 as an alternative benchmarked filter. In both cases, the target and binder sequences are supplied, the target structure is provided as a template, no MSAs are used, no binder template is supplied, 10 trunk recycles are run, and 5 outputs are sampled from the diffusion head. The selected structure is the one maximizing

$0.2 \cdot \mathrm{ipTM} + 0.8 \cdot \mathrm{pTM}.$

(Team et al., 25 Jul 2025)

For mini-binders, the Chai-1 filter is tuned to

$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$

For Boltz-2, the tuned thresholds are

$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.95,\qquad \mathrm{complex\_RMSD} < 2.5.$$

For comparison, the paper quotes AlphaFold 3 thresholds from AlphaProteo as

$$\mathrm{min\_iPAE} < 1.5,\qquad \mathrm{pTM\_binder} > 0.8,\qquad \mathrm{complex\_RMSD} < 2.5.$$

For macrocycles, the authors report that $\mathrm{pTM\_binder}$ is systematically low even for known true macrocycle binders, so they use a relaxed Chai-1 filter:

$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{complex\_RMSD} < 2.$$

Additional novelty and practicality filters differ by modality. Macrocycles are deduplicated, cysteine-containing or disulfide designs are removed, synthesis heuristics are applied, and MMseqs2 filtering against short PDB peptides rejects any cyclic permutation with more than 50% sequence identity to database peptides. Mini-binders are filtered against UniRef50, retaining only sequences with no match above 20% identity, and are structurally clustered with Foldseek using a TMscore $>0.6$ threshold to maximize diversity.

4. Experimental validation across macrocycles and mini-binders

The study evaluates seven benchmark targets previously used in prior binder-design literature. Macrocycle targets are MDM2, MCL-1, and PD-L1. Mini-binder targets are BHRF1, IL-7R$\alpha$, PD-L1, TrkA, and SARS-CoV-2 RBD. Macrocycles are tested by SPR after head-to-tail lactam cyclization, whereas mini-binders are screened by HT-BLI, quantified by 5-point BLI, and further assessed by mammalian display in HEK293T cells (Team et al., 25 Jul 2025).

Modality	Targets	Reported outcome
Macrocyclic peptides	MDM2, MCL-1, PD-L1	Hit rates 90.9%, 100.0%, 94.1%; best affinities 5.35 µM, 18.4 µM, 71.7 µM
Mini-binders	BHRF1, TrkA, PD-L1, IL-7R$\alpha$, SC2RBD	Hit rates 64%, 10%, 49%, 26%, 52%; best affinities 22.5 nM, 40 pM, 0.27 nM, <10 pM, <10 pM

For macrocycles, 700 designs per target were generated—100 per length for lengths 12–18—and the top 30 were selected per target. Synthesized and cyclized fractions were 77% for MCL-1, 57% for MDM2, and 87% for PD-L1, with 11–17 designs per target tested by SPR. The strongest macrocycle result is the uniformly high hit rate across all three targets. Affinity performance is more nuanced: Latent-X exceeds replicated RFpeptides controls on MDM2, but RFpeptides attains the stronger best single affinity on MCL-1.

For mini-binders, 20,000 designs per target were generated to obtain at least 100 final candidates per target after filtering and diversification. The best reported affinities reach low nanomolar and picomolar regimes, including 40 pM for TrkA and <10 pM for both IL-7R$\alpha$ and SC2RBD. In all-against-all mammalian display assays across IL-7R$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$0, PD-L1, and SC2RBD, the top Latent-X mini-binders showed no detectable off-target binding. Macrocycles also preferentially bound their intended targets, although some low-level off-target binding remained, which the authors note is expected for small canonical-only macrocycles.

5. Comparative position, structural diversity, and large-scale benchmarking

Latent-X is compared directly with RFpeptides on macrocycles and with AlphaProteo and RFdiffusion on mini-binders. On directly overlapping macrocycle benchmarks, Latent-X substantially exceeds RFpeptides in hit rate: 90.9% versus 37.5% on MDM2, and 100.0% versus 21.4% on MCL-1. On mini-binders, the picture is target-dependent but generally favorable. Relative to published AlphaProteo hit rates, Latent-X is lower on BHRF1, slightly higher on TrkA, much higher on PD-L1 and SC2RBD, and similar on IL-7R$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$1. Relative to RFdiffusion, Latent-X is higher on TrkA and PD-L1 and lower on the published IL-7R$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$2 hit rate, while achieving markedly stronger matched-assay affinities on TrkA, PD-L1, and IL-7R$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$3 (Team et al., 25 Jul 2025).

The paper also emphasizes structural diversity as a major differentiator. In held-out in silico analyses, RFdiffusion is described as producing mostly helical bundles, whereas Latent-X yields broader topology coverage, including substantial beta-sheet-rich and mixed-secondary-structure binders. This is presented as a practical advantage for surfaces that may not be well served by alpha-helical packing alone.

A larger in silico benchmark is constructed from 200 held-out post-cutoff PDB targets, each with 3 automatically selected epitopes and 100 binders per epitope, for

$$\mathrm{min\_iPAE} < 1,\qquad \mathrm{pTM\_binder} > 0.9,\qquad \mathrm{complex\_RMSD} < 2.$$4

designs per model per modality. Under Chai-1 filtering, average macrocycle hit rate is 8.26% for Latent-X versus 1.72% for RFpeptides, and average mini-binder hit rate is 5.11% for Latent-X versus 3.02% for RFdiffusion. Under Boltz-2 filtering, Latent-X still leads in both modalities. The fraction of targets with more than 10% hit rate is also higher for Latent-X, and the no-hit failure rate is lower for mini-binders.

Runtime is another reported advantage. Exact single-sample wall-clock values are shown only graphically, but the text states that Latent-X is approximately 10× faster than RFdiffusion on both A100 and H100 GPUs for a representative 80-aa PD-L1 binder design case. The system is also exposed through the hosted platform at https://platform.latentlabs.com, reflecting the paper’s emphasis on user-facing deployment rather than only offline benchmarking.

6. Limitations, disclosure level, and scientific significance

Despite strong wet-lab and in silico results, the paper is explicit about several limits. Latent-X requires a reasonable target structure, either experimental or predicted. Its context window is limited to 512 residues across target and binder, so cropping may be necessary and may omit distal context in large or highly context-dependent systems. The experimental validation is confined to macrocyclic peptides and mini-binders; broader modalities remain aspirational (Team et al., 25 Jul 2025).

The in silico filtering stack introduces its own biases. The Chai-1 and Boltz-2 filter thresholds were tuned on lab-validated mini-binders of 45–65 aa and predominantly helical folds, which may under- or over-value longer or non-helical designs. Novelty is also not absolute: generated proteins can still show sequence similarity to natural proteins despite MMseqs2-based filtering.

A further limitation is methodological transparency. The paper provides a high-level description of the model and its training data, including PDB and AlphaFold DB v4 with a cutoff excluding structures after 23 Nov 2023, but it does not disclose the internal training loss, diffusion transition kernels, score objective, or detailed coordinate-update equations in the text provided. This constrains architectural auditability relative to the granularity usually available for open academic design systems.

Scientifically, the importance of Latent-X lies in the combination of three claims that are supported together rather than separately: first, that all-atom co-generation of binder and target interface geometry is a viable design primitive; second, that this primitive can generalize across both macrocycles and mini-binders without modality-specific retraining or fine-tuning; and third, that the resulting workflow can reduce candidate counts to experimentally manageable scales while maintaining strong hit rates and high affinities. A plausible implication is that Latent-X represents not a generic replacement for structure-based design, but a concrete reorganization of the binder-design pipeline around end-to-end atomistic interface generation rather than backbone-first search.