In a remarkable advancement at the intersection of computational biology and drug design, researchers have unveiled a novel approach that enables the zero-shot design of drug-binding proteins with exquisite specificity and affinity. By leveraging neural iterative selection-expansion (NISE), a cutting-edge computational method, scientists have effectively tailored proteins to bind tightly and selectively to the chemotherapy agent exatecan, a derivative of camptothecin. This breakthrough not only represents a leap forward in protein engineering but also illuminates a new path towards the rational design of therapeutic binders without the need for prior experimental templates.
The team synthesized DNA sequences encoding proteins developed via NISE and expressed them in Escherichia coli, achieving robust protein yields. Initial biophysical characterization through size-exclusion chromatography affirmed that all four NISE-designed proteins existed predominantly as monomers, a critical feature ensuring the stability and functional integrity of the binders. Exploiting the intrinsic fluorescence properties of exatecan, the researchers conducted fluorescence polarization assays to quantify the binding affinities, obtaining dissociation constants (K_d) spanning from an impressive 0.12 µM to 17 µM. Notably, three designs manifested K_d values below 10 µM, significantly outperforming human serum albumin (HSA), a known non-specific binder with a considerably weaker K_d of 43 µM.
Among the array of NISE constructs, the exatecan-protein interaction construct (EPIC) emerged as the highest-affinity binder, exhibiting approximately 360-fold tighter association with exatecan compared to HSA. Structural analysis revealed that EPIC’s binding site was distinctively characterized by an optimized balance: it contained the fewest polar residues relative to other designs but conserved key interactions such as a buried glutamine residue positioned near the lactone ring of the ligand and an aspartic acid residue engaging the exatecan amine group. This intricate design resulted in a more extensive burial of the ligand’s apolar surface area, an attribute closely correlated with enhanced binding affinity.
For broader context, the investigators also explored proteins designed through the earlier COMBS methodology. Sixteen COMBS-designed proteins were expressed, yet only three demonstrated measurable binding to exatecan with K_d values of 8 µM, 12 µM, and 44 µM respectively. Among these, aggregation tendencies diminished their practical utility, and detailed analysis suggested that while COMBS could generate binders, their affinities were generally weaker than those produced by NISE. Comparative metrics such as ligand root-mean-square deviation (r.m.s.d.), backbone Cα r.m.s.d., and ligand predicted local-distance difference test (pLDDT) scores underscored how NISE optimized both protein folding and ligand interaction in a synergistic fashion.
Crucially, the stark contrast in binding efficacy between NISE and COMBS designs could not be chalked up to the use of residue-fluctuation-aware algorithms (RFAA) for filtering, as both approaches incorporated this quality control measure. Instead, the definitive advantage of NISE stemmed from its capacity to dynamically remodel protein backbones and binding pockets, crafting sequences better attuned to both structural stability and ligand complementarity. Case in point, EPIC’s binding site harbored a dramatically altered architecture compared with the baseline COMBS model, eschewing numerous original hydrogen-bond-forming residues in favor of a refined ensemble that judiciously balanced hydrophobic packing with targeted polar interactions.
Thermostability assays further attested to EPIC’s resilience under physiologically relevant conditions, bolstering its potential utility in therapeutic or diagnostic contexts. Intriguingly, the specificity profile of EPIC aligned tightly with chemical and steric resemblance among camptothecin-based drugs. Binding affinity titrations revealed a gradient of weakening interactions from exatecan to structurally related molecules such as FL118, belotecan, and camptothecin, establishing a clear structure-activity relationship revolving around the ligand’s unique fluoro-phenyl ring and amine functionalities.
The analysis suggested that molecular features contributing most substantially to binding free energy changes were hydrophobic contributions from desolvated fluoro and methyl groups, exceeding the influence of polar amine engagement. On the other hand, EPIC displayed no detectable affinity for bulky prodrug derivatives like irinotecan, nor for off-target ligands from unrelated drug classes such as the anticoagulant apixaban or the steroid dexamethasone. This intrinsic selectivity emerged despite the absence of explicit negative design constraints against off-target interactions, indicating that NISE inherently fosters specificity by optimizing for ligand compatibility during design iterations.
Remarkably, the NISE design trajectory reflected these precise binding preferences in silico, as ligand pLDDT values progressively increased when EPIC was co-folded with exatecan, yet remained low for off-target binders. This correlation between computational confidence metrics and experimental binding underscored the power of the integrative approach to predict and fine-tune protein-ligand interactions from first principles, without resorting to prior structural data or extensive screening.
This study spotlights the transformative potential of marrying deep learning with biophysical principles and synthetic biology, effectively enabling researchers to sculpt protein architectures from scratch tailored to challenging small molecule targets. The remarkable performance of EPIC exemplifies how iterative optimization frameworks can surmount longstanding hurdles in de novo binder design, which traditionally relied on laborious laboratory evolution or fragment-based screening techniques.
Looking ahead, the success of NISE could catalyze a new era in drug development where bespoke protein binders serve as versatile agents for targeted drug delivery, molecular sensing, or even in the modulation of pharmacokinetics. Beyond oncology, such an approach could be readily adapted to diverse therapeutic areas and emerging pharmaceutical modalities, drastically accelerating the pace of innovation.
Moreover, this method’s reliance on computational infrastructure, combined with validation through synthetic biology techniques, offers a scalable pipeline ideal for rapidly prototyping protein-ligand pairs against novel pharmacophores. While further studies will be necessary to explore in vivo stability, immunogenicity, and therapeutic indices, the foundational work presented here constitutes a compelling proof of concept with wide-reaching implications.
In sum, this research marks a pivotal advance in protein engineering, demonstrating that zero-shot design models empowered by neural iterative selection-expansion can generate high-affinity, highly specific drug-binding proteins. The confluence of cutting-edge machine learning, structural bioinformatics, and experimental biochemistry heralds a paradigm shift, enabling the rational creation of precision therapeutics that were previously beyond reach.
Subject of Research: Protein engineering and drug discovery; computational design of high-affinity drug-binding proteins.
Article Title: Zero-shot design of drug-binding proteins via neural iterative selection−expansion.
Article References:
Fry, B., Slaw, K. & Polizzi, N.F. Zero-shot design of drug-binding proteins via neural iterative selection−expansion. Nature (2026). https://doi.org/10.1038/s41586-026-10670-w
Image Credits: AI Generated
