Researchers have introduced a breakthrough computational framework for the de novo design of small-molecule binding proteins, utilizing a zero-shot approach called Neural Iterative Selection-Expansion. The methodology leverages a specialized heterograph neural network that predicts protein sequences and side-chain conformations conditioned on three-dimensional protein-ligand co-structures. By iteratively refining sequences and co-structures through reciprocal conditional probability sampling, the algorithm maximizes tripartite self-consistency across the protein backbone, amino acid sequence, and target ligand. This closed-loop optimization eliminates reliance on traditional energy functions, enabling precise joint refinement of sequence, structure, and ligand conformation. The framework demonstrated unprecedented efficacy in designing high-affinity binders for two clinically relevant small-molecule drugs. For the anticancer agent exatecan, the algorithm generated a protein designated EPIC, achieving an initial dissociation constant of 120 nanomolar. Subsequent neural proofreading of the binding site yielded a double mutant with a binding affinity of 1.2 nanomolar, representing a 100-fold improvement over the initial design. Beyond affinity, the engineered protein successfully encapsulated the drug’s labile lactone ring, preventing hydrolysis in physiological conditions for over 50 hours. This functional property positions the binder as a potential delivery vehicle for maintaining the bioactive drug form in circulation. To validate algorithmic robustness across distinct protein folds and targets, researchers applied the system to apixaban, a blood-thinning anticoagulant. Utilizing NTF2 structural scaffolds and an advanced co-structure predictor, the system achieved a success rate exceeding 80 percent. The highest-affinity variant, APEX, bound apixaban with a dissociation constant of 80 picomolar, rivaling the affinity of the drug’s native biological target while being significantly smaller. Computational analysis confirmed that the designs maintained high specificity, showing negligible binding to structurally dissimilar off-target molecules. The protocol represents a significant departure from conventional protein design pipelines that depend heavily on high-throughput experimental screening or rigid energy minimization. By training exclusively on neural network confidence metrics and self-consistency scores, the algorithm efficiently navigates the joint probability distribution of protein-ligand systems. Comparative benchmarks revealed that the approach outperforms state-of-the-art alternatives by orders of magnitude in both success rate and binding affinity. The system’s architecture is inherently modular, allowing seamless integration of advancing co-structure predictors and sequence design models as they emerge. This advancement establishes a scalable, computationally driven pipeline for generating bespoke protein reagents. By decoupling binder design from exhaustive experimental iteration, the framework accelerates the development of targeted therapeutics, molecular sensors, and delivery systems. The demonstrated capacity to engineer proteins that not only bind with sub-nanomolar precision but also actively modulate drug stability marks a pivotal step toward programmable molecular biology and next-generation precision medicine.