MAP4K4 is a serine/threonine kinase involved in multiple signaling pathways (JNK, p38 MAPK, ERK1/2) and implicated in cancer progression, TNF-α-driven insulin resistance, macrophage-mediated inflammation, and cardiomyocyte apoptosis in heart failure. The authors note that no selective small-molecule MAP4K4 inhibitor has reached clinical use and that epitope details for some existing antibodies remain incomplete. This motivated an effort to create small, site-specific protein binders as probes or potential reagents directed at defined surface sites on MAP4K4.
The design workflow combined generative and sequence-design methods with structure-prediction evaluation. Key tools cited in the source include RFdiffusion for backbone generation, ProteinMPNN for sequence design, and AlphaFold variants for structure and complex prediction. The authors also evaluated an integrated “one-shot” approach using BindCraft. These publicly available computational tools formed a pipeline intended to produce compact binders (~50–130 residues) to chosen surface hotspots on MAP4K4.
To select targetable surface sites, the team implemented an automated hotspot-determination algorithm. This algorithm weighted geometric features, chemical properties, structural rigidity, and AlphaFold pLDDT to prioritize surface regions likely to yield stable, specific interactions. The report indicates this approach was used to define the target sites for de novo binder generation.
From thousands of generated candidate sequences, the authors selected a subset for more rigorous evaluation. Candidate sizes spanned approximately 50–130 amino acids. Twenty of the most promising designs were chosen for downstream structure-based assessment with AlphaFold3, representing a narrowed, computationally prioritized set from the larger design pool.
All 20 selected designs were evaluated using AlphaFold3 (AF3) to model complexes with MAP4K4. The predicted interface-template modeling (ipTM) scores across these designs ranged from 0.16 to 0.90. Nine candidates achieved ipTM ≥ 0.80, and five scored ≥ 0.87, indicating high AF3 confidence for multiple binders. Two designs were noted to engage non-overlapping hotspots on opposite faces of MAP4K4, making them plausible candidates for a sandwich-pair format in assay development.
BLASTp searches of all designed protein sequences found only low-significance matches for approximately half of the designs. The authors interpret this as evidence that many designs represent novel solutions and exploration of previously unobserved regions of protein sequence space rather than simple recapitulation of known natural motifs.
To probe dynamic stability, five protein–MAP4K4 complexes spanning the AF3 confidence range were subjected to 100-ns explicit-solvent molecular dynamics simulations. Interchain contacts were retained across the simulations, though the degree of stability differed substantially between systems. These MD results provide a complementary, physics-based perspective on predicted complex stability beyond static AF3 models.
Specificity of high-confidence binders was assessed computationally by comparing AF3-predicted interactions with MAP4K4 against a negative-control kinase, CDK2. The comparison produced a consistent reduction in ipTM when the designed binders were modeled with CDK2; this reduction reached statistical significance (p = 0.0039). The authors present this as supporting evidence that the top-scoring designs are specific to MAP4K4 in AF3 evaluations.
The authors screened designs with ToxinPred2 and AlgPred 2.0 for potential safety concerns. These analyses suggested one candidate may be a potential allergen and two candidates may be potential toxins, highlighting the value of early in silico filtering for downstream experimental prioritization.
The study demonstrates that de novo design of small, site-specific protein binders against an under-served disease target such as MAP4K4 can be achieved using publicly available computational tools and modest resources. Multiple candidates showed high AF3 confidence and sequence novelty, two binders target non-overlapping sites suitable for sandwich assays, and MD plus negative-control AF3 comparisons supported specificity for several top designs. The authors suggest these findings point to an expanded role for public and amateur scientists in contributing to biotechnology through accessible computational methods.
Note: This work is reported as a preprint and has not been peer reviewed. Details beyond those reported in the source (for example experimental validation, binding affinities, or laboratory methods) were not provided in the source article body.