De Novo Minibinder Design at the Interface of AI and Structural Bioinformatics

We developed a novel, multi-stage pipeline for de novo minibinder design that integrates generative AI methods with structural bioinformatics and physics-based evaluation. The goal of this approach is to efficiently explore sequence–structure space while progressively enriching for candidates with realistic binding potential.

The pipeline begins with structural analysis of the target protein to identify a suitable binding region (hotspot). This step defines the geometric constraints for subsequent design and ensures that all generated binders are directed toward a biologically relevant epitope.

Backbone generation is performed using RFdiffusion, enabling the creation of diverse minibinder scaffolds compatible with the target surface. A two-phase sampling strategy is employed: an initial exploratory phase to identify promising parameter regimes, followed by a focused phase to generate a larger number of candidates under optimized conditions. This balances structural diversity with targeted design.

Sequences are assigned to the generated backbones using ProteinMPNN, which performs side-chain optimization and sequence design in a structure-aware manner. For refinement, interface-focused redesign is applied by selectively mutating residues at the binding interface while keeping the overall scaffold fixed. This enables targeted improvement of binding interactions without disrupting fold stability.

Monomer validation is performed using ESMFold to ensure that redesigned sequences maintain structural integrity. Candidates are evaluated for agreement with the designed backbone, overall model confidence, and the absence of poorly defined regions. In addition, surface properties such as hydrophobic exposure and charge distribution are assessed to eliminate designs with unfavorable physicochemical characteristics.

Complex formation is evaluated using AlphaFold3, which provides detailed predictions of binder–target interactions. Multiple confidence metrics are used to assess interface quality, including per-residue confidence, interface confidence, and inter-chain agreement. These metrics enable discrimination between plausible binding modes and incorrect docking configurations.

Interface quality is further evaluated using a combination of structural bioinformatics tools. CCP4 Shape Complementarity is used to assess geometric fit, FoldX AnalyseComplex provides an estimate of interaction energy, ARPEGGIO characterizes the interaction network at the atomic level, and PRODIGY predicts binding affinity based on interfacial contacts. Additionally, PepPatch is used to evaluate surface patch properties, including hydrophobic and electrostatic distributions. The combination of these tools provides complementary insights into interface geometry, energetics, and physicochemical plausibility.

Importantly, the pipeline is designed as an iterative optimization framework. Rather than relying on strict rejection criteria at intermediate stages, candidates are ranked based on a combination of metrics, and promising designs are further refined through targeted redesign cycles. This allows the identification of both high-confidence candidates and structurally diverse alternatives, which is critical for experimental success.

Overall, this work presents a novel pipeline that combines de novo generative design with rigorous structural and energetic validation. By integrating multiple orthogonal methods and emphasizing ranking over hard filtering, the approach enables efficient identification of realistic minibinder candidates with favorable binding characteristics.