Computational Design of Nipah Virus Binders: Analysis of the AdaptyvBio Competition Results

Introduction

Nipah virus (NiV) is one of the scariest viral threats out there today. With mortality rates hitting 40-75% and zero approved treatments available^[1], the WHO has labeled it a priority pathogen^[2]. Recently, AdaptyvBio launched a competition to design protein binders against the Nipah virus Glycoprotein G (a tricky, tetrameric target) giving us a fresh look at what actually works in computational protein design.

We conducted a comprehensive analysis of the complete competition dataset: 3,749 total submissions, of which 1,030 were experimentally tested. This gives us unprecedented insights into both the computational filtering process and experimental validation at scale. By analyzing everything from the computational screening pipeline to binding affinities and design strategies, we uncover what separates winning approaches from the rest.

Competition Design and Methodology

The challenge was straightforward: design a novel protein that binds tightly to NiV-G, the key protein Nipah uses to enter cells^[3]. To keep things grounded in real-world therapeutic constraints, Adaptyv set some rules:

• Maximum protein length: 250 amino acids
• Novelty requirement: At least 10 amino acids different from existing databases
• Submission limit: 10 designs per participating team
• Format flexibility: Both de novo designs and optimized pre-existing binders accepted

Adaptyv used high-throughput surface plasmon resonance (SPR) to measure the exact binding kinetics (on-rates, off-rates, and overall affinity) ^[5].

UniBio Intelligence Participation

UniBio Intelligence participated as one of the competing teams, submitting our maximum allowance of 10 designs. Our strategy was intentionally conservative, focusing on developing viable therapeutic candidates rather than pursuing high-risk, high-reward approaches that might not translate to clinical applications.

We chose to primarily use the scFv format and build upon known, clinically-relevant neutralizing scaffolds while also including a few de novo nanobody designs. We didn't specifically optimize for the Boltz2 iPSAE metric as we know from our previous experience (further confirmed from detailed analysis shown below) that this is a weak predictor of binding. As our designs did not pass the iPSAE metric threshold required for experimental validation, none of ours were tested experimentally.

Complete Dataset Analysis: From 3,749 Submissions to 103 Binders

To truly understand computational protein design effectiveness, we analyzed not just the final results, but the entire pipeline from submission to success ^[16].

This funnel reveals that computational filtering eliminated 72.5% of submissions. Understanding this filtering process is crucial for improving future computational design strategies.

Distribution comparison of key computational metrics between tested designs (blue) and filtered designs (gray).

Computational Filtering Effectiveness Analysis

Spearman correlation heatmap showing relationships between computational metrics and experimental binding parameters. Strong correlations reveal which metrics best predict binding success in experimentally validated designs.

These results show that the competition's filtering strategy of using the Boltz2 iPSAE metric has some merit, but the significant correlation between shape complementarity and binding success suggests that incorporating geometric compatibility metrics could substantially improve computational filtering accuracy.

Validation from Filter-Free Success Stories

These findings about aggressive filtering potentially eliminating successful designs receive compelling validation from Escalante Bio's approach. By completely removing computational filtering steps, Escalante achieved an extraordinary 90% binding success rate (9 out of 10 designs) and won 1st through 4th place in the de novo category ^[17]. Remarkably, most of their winning designs would have failed traditional computational filters, including BindCraft's "unsaturated hydrogen bond filter" with its overly stringent static cutoffs.

Binding Performance

Out of 1,030 experimentally validated designs, only 103 (10.0%) bound, and 882 of 1,030 (85.6%) expressed successfully. The top end of the leaderboard was fiercely competitive. A total of 26 designs reached single-digit nanomolar affinity, with the absolute best hitting a K_d of 8.7 × 10⁻¹⁰ M. This top binder was heavily inspired by ephrin-B2, Nipah’s natural receptor.^[6].

Design Class Performance

Complete design class analysis: submission distribution, computational filtering rates, and experimental binding success rates across all design types

Taking the top 1% of designs ranked by min_ipSAE resulted in over 60% of them successfully binding in the wet lab. However, post-competition analyses revealed significant nuances in using ipSAE. While it separates binders from non-binders, it struggles to rank which binders will have the highest true binding affinity (K_d). Researchers at Silico Biosciences also noted that ipSAE can be inconsistent, with identical sequences producing varying scores across folding runs due to extreme sensitivity to sub-Angstrom structural prediction variations (e.g., stochasticity in Boltz-2 random seeds)^[20]. Prof. Roland Dunbrack pointed out that exploring multiple ways to calculate ipSAE (ipsae_min, ipsae_max, or aligning on target vs. binder) can yield better correlations with actual binding affinity ^[21]. Other comparative analyses suggest alternative metrics, such as Chai-1 min_ipSAE, may serve as even stronger binary classifiers ^[18]. Adaptyv team also noted that score variability "was not uniform across designs and therefore problematic," potentially giving unintended advantages to designs with more stable computational predictions ^[16]. This highlights a fundamental challenge in computational screening: metrics that work well on average may still introduce systematic biases due to inconsistent reliability across design space.

Moving forward, the broader community is shifting toward consensus scoring across multiple models (AlphaFold3, Boltz-2, Chai-1) and ensemble calculations to overcome both the noise and reproducibility issues inherent in single-metric evaluations ^[20].

Similarly, ESMFold pLDDT scores did exactly what they're supposed to do: they predicted whether a protein would express successfully in the lab, but couldn't tell if it would actually bind the target^[8]. This outcome reinforces a core philosophy: while computational metrics are useful guardrails, maintaining structural integrity and prioritizing therapeutic viability should remain the primary drivers of the design process.

Shape Complementarity as Top Binding Predictor

Our correlation analysis of computational metrics against experimental binding data revealed that shape complementarityemerges as the strongest predictor of experimental binding success, showing significant correlations with both binding kinetics (Spearman r = 0.35, p < 0.001) and equilibrium dissociation constants (Spearman r = -0.31, p = 0.002).

This finding suggests that geometric compatibility between designed binders and the Nipah virus glycoprotein surface is more predictive of binding success than traditional structural confidence metrics. Shape complementarity captures the three-dimensional fit between binding interfaces, accounting for both positive and negative electrostatic interactions as well as hydrophobic contacts.

Distance from Known Binders: Similarity Analysis

To understand the relationship between design success and similarity to validated Nipah virus binders, we performed comprehensive sequence identity analysis comparing all 101 competition binder sequences with known neutralizing antibodies from PDB structures, including the potent 41-6 antibody (PDB: 8K3C) and mAb66 (PDB: 6T3F), as well as control sequences used in the competition.

Sequence identity heatmap showing all 101 competition binder sequences compared to known Nipah virus antibody chains. Values range 0-1 with colors indicating design class (left bar). High-similarity regions (yellow) reveal sequences closely related to validated antibodies.

Clinical Translation Potential

Multiple teams generated sub-nanomolar binders in this competition even though binding success rate was quite low. But binding is just step one. To actually get these molecules into patients, we also need co-optimization of:

• Neutralization breadth across Nipah virus strains
• In vivo stability and pharmacokinetics
• Manufacturing scalability and cost-effectiveness
• Safety profile and immunogenicity risk

Our computational pipelines are designed to co-optimize binding affinity with developability and immunogenicity risk profiles, ensuring that promising binders identified in competitions like this can successfully navigate the complex path from bench to bedside.

Conclusion

This data analysis highlights that the AI protein design still needs to figure out:

Affinity Prediction Accuracy: While computational methods can identify potential binders, quantitative prediction of binding affinity remains imprecise. The moderate correlation between computational scores and experimental K_d values indicates substantial room for improvement in predictive algorithms.

De Novo Design Limitations: Fully de novo approaches showed limited success compared to template-based methods, suggesting that current computational frameworks may not adequately capture the subtle sequence-structure-function relationships required for high-affinity binding.

Citing This Work

Huge props to the Adaptyv team for making this scale of analysis possible through transparent data sharing. If you found this analysis helpful or reference our insights on computational Nipah virus binder design in your work, please cite:

@article{unibio2026nipah,
  title={Computational Design of Nipah Virus Binders: Analysis of the AdaptyvBio Competition Results},
  author={Vivek Kohar},
  journal={UniBio Intelligence Blog},
  year={2026},
  month={5},
  url={https://unibiointelligence.com/blog/nipah-virus-binder-adaptyv},
  note={Accessed: 2026-05-12}
}

References