How Cradle’s 12 designs would have ranked top-12 in Adaptyv’s protein design competition 

Competition recap

In our previous blog post, we shared how Cradle secured first place in Adaptyv Bio’s second protein design competition. The challenge was to design a novel single-chain binder targeting the extracellular domain of EGFR, a key therapeutic target in cancer treatment. The main constraint was that the designs had to differ from known therapeutic binders in at least 10 positions. 

With just one click on Cradle’s platform, our team optimized the framework region of an scFv-formatted Cetuximab while preserving its CDRs.

Since no experimental data was available, our platform executed the first step of Cradle’s pipeline: zero-shot diversification. This process is designed to generate functional and diverse sequence variants from an existing lead. For this purpose, we selected an scFv-formatted Cetuximab as the lead. Unlike de novo design, which seeks to create entirely new binders, zero-shot diversification focuses on exploring sequence space around a known functional lead.

We submitted 10 designs, each with 10 mutations from the scFv-formatted Cetuximab lead. In total, 130 protein designers submitted 1,131 binders to the competition. Adaptyv ranked these binders using in-silico metrics, including the PAE and pTM metrics from AlphaFold2, as well as ESM2’s pseudo log-likelihood. 

Adaptyv selected the top 100 sequences based on this in-silico ranking for experimental validation. Unfortunately, our designs were overlooked by these in-silico metrics and placed between 322nd and 623rd. Luckily, Adaptyv still selected 1 to 5 designs per protein designer based on the quality and novelty of their approach. This allowed one of our variants to be selected, which went on to win the competition, achieving a binding affinity of 1.21 nM—eight times stronger than Cetuximab (9.94 nM) and significantly outperforming the nearest competitor's 5.18 nM.

Post competition

After winning the competition, Adaptyv gave us 11 free assays to test other designs. We decided to test the remaining 9 designs which we submitted to the competition—plus 2 more designs which we generated in the same way.

Deep dive: Generation Report

Cradle’s platform provides a detailed “Generation Report” for the selected set of sequences designed for lab testing. This report allows users to analyze overall mutation distribution.

 In green, we highlight the CDR regions, which we chose to block, while blue regions indicate where our models identified promising mutations.

Rather than optimizing individual sequences, Cradle strategically designs sets of diverse sequences to maximize the chance of a successful round. Across the 12 sequences, a total of 33 unique mutations at 25 distinct positions were suggested. The frequency of these mutations varies based on where our models identify promising mutations. Additionally, Cradle’s models account for epistasis—considering the combined effect of multiple mutations rather than evaluating each mutation in isolation.

[Figure A.] Heatmap showing pairwise sequence diversity based on BLOSUM62 substitution scores. Diversity is computed by subtracting pairwise similarity scores from each sequence’s self-alignment score. 

Although we designed 12 sequences, each exactly 10 mutations away from the same lead sequence, the submitted set maintains a balanced diversity, spanning sequences that are relatively similar to each other (<10 mutations) as well as those that are more distant (>10 mutations). Notably, the top-left quadrant of the heatmap in Figure A. highlights our top four sequences which exhibit high diversity based on BLOSUM62 substitution scores.

The results

All 12 sequences selected by Cradle’s pipeline proved to be strong binders, and they would have ranked as the top 12 in the competition. 

After correcting for batch normalization, we identified three subnanomolar binders that outperformed the competition winner. Notably, these top-performing sequences were 8, 14, and 19 mutations away from the winning sequence (cradle_3 refers to the competition-winning sequence).

[Figure B.] Bar plot showing the dissociation constant (KD) for different mutants, with error bars representing variability. Lower KD values indicate stronger binding affinity. Cradle_3 (in dark blue) is the competition-winning sequence and Cetuximab_scFv (in orange) is the chosen lead. 

These results demonstrate that Cradle’s zero-shot diversification pipeline effectively explores "headroom" beyond an initial lead. In this case, our results suggest that significant improvements were possible within the framework region of the scFv-formatted Cetuximab.

Furthermore, these findings highlight a key insight: for affinity maturation, lead optimization continues to be a more effective strategy than de novo design. Despite Adaptyv’s competition featuring many designs generated by state-of-the-art de novo design methods, the best results came from strategically optimizing an existing lead rather than designing from scratch.

Beyond binding affinity, Immunogenicity + Developability 

When diversifying a lead binder, we do not have direct control over the effect on the properties of interest. We might find variants with increased affinity at, for example, the cost of weaker expression or worse immunogenicity. Specifically for therapeutic development, it is crucial to consider the balance between improving target binding and preserving other key attributes such as manufacturability, expression levels, immunogenicity, etc.

We ran some experiments to confirm that we did not introduce any detrimental mutations in our 12 binders. In particular, we find that we improve expression levels of the scFv formatted Cetuximab, and we find no evidence of deleterious mutations linked to immunogenicity. 

Expression

In the competition experiments, the lead scFv-formatted Cetuximab exhibited low expression across both replicates, whereas our winning sequence consistently showed high expression in all three replicates. In the second experiment, all our sequences demonstrated medium to high expression levels, with one of our sub-nanomolar binders consistently displaying high expression. Note that Adaptyv uses a cell-free expression system, so results may vary in other systems. Further analysis is required to assess if these variants will express well in standard mammalian expression systems. 

Immunogenicity

Immunogenicity assessment evaluates the potential of therapeutic proteins to provoke unwanted immune responses, which can affect their safety and efficacy. This typically involves extensive in vitro assays and clinical evaluations, which can be time-consuming and costly. In-silico methods exist and offer a faster alternative by predicting immunogenicity using humanness predictions, which measure similarity to human proteins, and MHC class II binding predictions, which identify peptides likely to provoke immune reactions.

Humanness 

One common way to estimate immunogenicity is to assess the humanness of a variant, i.e., how closely its sequence resembles human antibodies. We predicted the humanness of the VH and Vk sequences using the LakePharma Antibody Analyzer which provides a T20 score, which quantifies how closely an antibody's variable region resembles human antibodies [0]. 

The predicted scores indicate that our mutations did not reduce the humanness of the variants. We note that the lead scFv-formatted Cetuximab consists of mouse CDRs grafted onto a human framework, setting a high bar for humanness.

[Figure C.] T20 humanness scores for cetuximab-derived VH sequences. [Figure D.] T20 humanness scores for cetuximab-derived Vκ domains. T20 scores are between 0 and 100, with 100 being “the most human”. 

MHC II binding

We can use MHCII binding predictions to assess the risk of immunogenicity. In particular, a higher number of MHCII binding peptides in a sequence indicates a greater potential for present T-cell epitopes, which increases the likelihood of an immune response, anti-drug antibodies, and potential efficacy loss.

Given an input sequence, NetMHCII predicts the binding affinity of each peptide within the sequence to different HLA-DR, HLA-DQ, and HLA-DP alleles [1]. The number of core peptides binding to MHC is determined by counting the unique core peptides that rank within the top 2% of 1,000,000 randomly chosen interactions. 

Predictions indicate that we did not introduce any mutations that could increase the risk of unwanted immunogenic responses. In fact, there may be fewer immunogenic peptides on average within our generated sequences, though the effect size is not large [see appendix Figure F.]. 

[Figure E.] Predicted number of MHCII-binding peptides in cetuximab-derived mutants based on the NetMHCII prediction tool. Less MHCII-binding peptides is better. 

Developability

While in silico metrics provide useful preliminary insights into factors like expression and immunogenicity, they do not capture the full complexity of developability. Properties such as stability, manufacturability, efficacy, etc. require comprehensive in vitro and in vivo studies to ensure therapeutic viability. 

References

[0] Gao, S.H.; Huang, K.; Tu, H.; Adler, A.S. Monoclonal antibody humanness score and its applications. BMC Biotechnol. 2013, 13, 55.

[1] Racle, J.; Michaux, J.; Rockinger, G.A.; Arnaud, M.; Bobisse, S.; Chong, C.; Guillaume, P.; Coukos, G.; Harari, A.; Jandus, C.; et al. Robust prediction of HLA class II epitopes by deep motif deconvolution of immunopeptidomes. Nat. Biotechnol. 2019, 37, 1283–1286.

Additional figures 

[Figure F.] Count distribution of the delta in MHCII binding core peptides across different alleles, compared to scFv-formatted Cetuximab.

——

Sign up for our webinar

Optimizing Cetuximab: How Cradle’s 12 designs would have ranked top-12 in Adaptyv’s protein design competition

Wednesday, April 16th, 2025 | 16:00-17:00 CET / 10:00-11:00 AM EDT

Sign up here