Utilyze Submits Protein Designs to Anthropic x Adaptyv Competition

Kenneth Anshewitz | November 9, 2026

We generated 24 original computational protein designs with NVIDIA RFD4-Proteina, rejected two because the molecular target had changed, shortlisted 20 for the competition, and 19 passed Proteinbase's intial novelty screening process. We verified the underlying generation and structural-analysis records. No design has yet been shown experimentally to bind, or distinguish, the KRAS G12D target.

(two-tone protein ribbon complex)

Protein design is often presented as a moment of generation: a new amino-acid sequence appears, a predicted structure renders in three dimensions, and the machine seems to have done the important work. For Utilyze, the more demanding work began immediately afterward. Was the sequence really produced by the claimed model? Had the model silently altered the target? Did the proposed protein contact the part of the molecular complex that mattered? Would a novelty filter accept it? And which of those checks, if any, said something about real biological binding?

We explored those questions during the 5th challenge of the Anthropic × Adaptyv Protein Design Competition. The campaign became a case study for the Utilyze Protein Pipeline, our developing computational protein research workflow, and for Synchronicity, our broader evidence-organizing research system.

The result over the competition was a real, traceable generation-and-screening workflow with both encouraging platform feedback and scientifically consequential negative findings. Here is what we actually did.

The challenge was not simply “bind KRAS”

KRAS G12D is a mutation in which aspartate replaces glycine at position 12 of KRAS. In this competition, the target was a piece of the mutant KRAS protein displayed on a cell-surface immune presentation complex, not the complete free KRAS molecule.

The mutant peptide is ten residues long: VVVGADGVGK. Its wild-type counterpart, VVVGAGGVGK, differs at peptide position 6. Both are displayed by HLA-A*11:01, together with β2-microglobulin. A useful designed binder would need to recognize the mutant peptide–MHC complex while avoiding both the wild-type peptide–MHC complex and the same HLA/β2m scaffold without peptide.

That creates three independent questions: Does the design bind the intended mutant complex? Can it distinguish one amino-acid substitution? And is it actually peptide-dependent, rather than sticking to a shared region of HLA or β2m? The organizers plan to measure these conditions experimentally; our computational contact checks are not substitutes for those measurements.

We began with PDB 7OW6, a published X-ray structure of a T-cell receptor bound to the mutant KRAS peptide–HLA-A*11:01 complex. The original receptor is a useful structural example, but it was not our proposed binder. We removed the receptor chains from the conditioning input, retained the HLA heavy chain, β2m, and mutant peptide, and marked a new design region of roughly 70–100 residues.

We also encountered a construct mismatch worth documenting. The retained HLA heavy chain modeled from the public structure contains 276 residues, whereas the competition's target specification lists 281. That five-residue difference does not prove the models are unusable, but it does mean we cannot claim exact assay-construct equivalence without further reconciliation.

First: prove there was a real model output

A predicted-looking structure is not the same as a verified model output. During early pipeline development, we found that the surrounding software could produce a plausible sequence file or accept provider-shaped metadata without proving the underlying protein had actually been generated. Even an input structure could contain a chain placeholder labeled as the “design,” although no coordinates for that designed protein yet existed.

Our first successful authenticated NVIDIA RFD4-Proteina demonstration produced an unrelated 40-residue model output. This established that the transport and provider inference path worked; it did not establish a KRAS binder. We preserved the original provider response and structure and used that demonstration to test independent geometry screening.

For the KRAS campaign, we then switched to the publicly sourced 7OW6 peptide–MHC conditioning input. We used the RFD4-Proteina model, developed by the Baker Laboratory with NVIDIA, through authenticated hosted inference. NVIDIA's published implementation supports conditioned generation of both a candidate sequence and a modeled structure. It is their generative technology; Utilyze's contribution here was the research question, preparation, provenance checks, structural triage, and integration workflow.

Rather than manipulating one candidate into multiple cosmetic variants, we submitted bounded sampling batches and retained the original output for each result. Our source audit recovered seven provider-response archives containing 24 generated designs, with the corresponding workflow records, full response contents, and molecular coordinate files. One three-replicate workflow reported a downstream screening failure even though the original provider response contained three model outputs. We distinguish an authenticated generation from a later structural acceptance gate.

Second: reject candidates when the target changes

Each modeled complex contained the original target chains and a newly generated protein. We compared the output target-chain sequences against the input reference: 276-residue HLA heavy chain, 99-residue β2m, and the ten-residue mutant peptide.

Two of 24 outputs had changed the HLA heavy-chain sequence. We rejected those outputs. The remaining 22 retained the expected target-chain sequences and contained 22 distinct candidate sequences. From this set we shortlisted 20 computational proteins, each between 73 and 100 amino acids, preserving a direct hash-checked link from the candidate sequence to the original provider-returned structure.

This is a small but important scientific discipline. A model can produce an appealing interface while silently changing the thing it is supposedly binding. In a challenge with specificity requirements this strict, checking target identity is not optional.

The shortlist was an engineering triage, not a measured affinity leaderboard. Our internal score combined proximity and diversity signals. It was not a calibrated likelihood of binding, and it should not be interpreted as such.

Third: let geometry disagree with the pretty picture

We used two different geometry workflows and deliberately kept their results separate. First, we inspected an online MolProbity Next residue-level validation CSV for the earlier, separate 40-residue NVIDIA demonstration protein. A file-format issue with the raw mmCIF led us to make a PDB representation with unchanged atomic coordinates. The resulting residue table was not reassuring: reconstructed residue bookkeeping identified 30 of 39 peptide C–N connections as bond-length outliers; 25 of 40 residues carried bond-angle outlier assignments; and one chirality outlier was reported.

The CSV did not contain a complete MolProbity global score, and we have not independently recovered the original validator session. We therefore treat it as a residue-level diagnostic, not a formal certification.

The important lesson was that plausible-looking backbone and rotamer categories did not erase serious covalent geometry defects. We did not claim that this demonstration protein passed geometry validation.

Second, we independently examined the atomic coordinates of the 20 later KRAS-targeted models using local structural parsing and heavy-atom distances. This was not the same as running full MolProbity reports or all-hydrogen clash calculations on each submitted design.

Across those 20 models, 13 have at least one modeled binder–target heavy-atom separation below 2.5 Å when the HLA, peptide, and β2m chains are all considered. If we count HLA and peptide but leave out β2m, the total is 12. These are proximity warnings that require detailed atomic interpretation; some may represent steric overlap or artifacts, and a simple distance cutoff by itself is not a formal validation score.

One shortlisted model was particularly instructive: it showed a substantial modeled interface with β2m and an exceptionally short binder–β2m atom distance, approximately 0.23 Å. That is a reason to examine a candidate critically, not a reason to declare a stronger interaction. A model that concentrates on the conserved scaffold may have trouble with the competition's peptide-dependence requirement.

Fourth: distinguish contact with the mutation position from contact with the mutation chemistry

All 20 shortlisted models bring at least one binder heavy atom within 5 Å of the backbone of mutant peptide position 6. At first glance, that appears to be encouraging: they are close to the amino acid whose identity changes between KRAS G12D and wild type.

But when we separated backbone from side chain, the interpretation changed. 0/20 has a modeled binder heavy atom within 5 Å of the Asp6 side-chain atoms, and the nearest modeled contact to either Asp carboxylate oxygen was approximately 5.95 Å. The designs may still recognize a change in peptide conformation indirectly; the public structure itself shows that the Asp side chain is constrained by its environment. However, our local models do not establish direct recognition of the mutant-specific carboxylate group.

This was probably the campaign's most important scientific caution. Contacting the right residue index is not the same as demonstrating molecular discrimination between Asp and Gly. Neither distance-based screening nor a visual rendering measures differential binding affinity. We have not run matched wild-type and peptide-free controls for these candidates, nor have we measured any physical binding response.

Fifth: 19 of 20 passed the platform workflow, what does that tell us?

The founder observed 19 of the 20 shortlisted proteins pass Proteinbase's upload-screening workflow, with one excluded during submission. There is still an evidentiary gap: the available portal view did not expose individual novelty scores, which exact candidate failed the filter, the final 19-sequence mapping, or a competition-specific receipt listing all designs.

A high upload-screening pass rate indicates most candidates apparently cleared the platform's preliminary eligibility/novelty check. It does not mean 95% of proteins are experimentally novel folds or that 95% will bind. Adaptyv's published Proteinbase novelty system distinguishes sequence and structural similarity using a four-level scale; without the actual score report we cannot say that all 19 reached level 4.

We also detected several fairly similar sequences within our own shortlist. Exact uniqueness against the other generated outputs does not by itself establish external novelty, and external novelty does not by itself establish mutant-selective function.

Is 19 out of 20 actually an impressive novelty result?

It is encouraging that almost every preselected candidate made it through the reported upload workflow. But a reported 95% pass rate is not an unbiased estimate of a generator's success rate. We shortlisted 20 from 24 outputs after rejecting known target-sequence problems and applying our own criteria; only that selected subset entered the portal. The numerator therefore says something about this particular portfolio and platform interaction, not about all sequences RFD4 could produce.

There is also a more basic missing piece: we did not retrieve the platform's individual novelty grades or the exact per-design reason for rejection. A form can reject a record for syntax, molecule class, duplication, identity thresholds, or other validation issues. A human-observed pass should not be upgraded to an independently documented ProteinTyper 3/4 score without its row-level evidence.

Adaptyv's published novelty framework separates similarity to known sequences from similarity to known folds. In its general-protein rubric, level 3 is partly novel and level 4 is more distant from known sequence and structure references. Those are useful database-relative classifications. They do not measure whether a molecule folds, expresses, binds the intended target, or discriminates a single mutation. The criteria can also differ for antibody-like designs. Even two similar designs in our shortlist might both clear a comparison against an external database; such passes would not make those two sequences independently diverse of one another.

To compare performance scientifically, we would need the full distribution of level scores, the exact rejected candidate, all preselection outcomes, an external baseline pool screened using the same protocol, and eventual measured binding. Until then the correct verdict is promising screening progress, not a quantified biological breakthrough.

Where Utilyze Protein Pipeline and Synchronicity fit

Our work leveraged breakthroughs like RFD4-Proteina, MolProbity, Proteinbase, or the underlying structural biology. RFD4-Proteina produced new computational sequences and structures. RCSB PDB supplied the experimentally determined structural reference. MolProbity Next informed independent geometry diagnostics on the earlier demonstration.

Proteinbase provided the competition upload and screening workflow.

Utilyze Protein Pipeline connected preparation, provider-output provenance, target-sequence preservation, modeled geometry checks, and shortlist/export logic. Synchronicity helped us structure the research question, organize the supporting evidence and negative controls, and keep a decision trail. In a future controlled evaluation, we want to test whether that evidence-aware process selects better experiments than a fixed baseline or a generator's native score. That comparative performance test has not yet been done.

The broader lesson is not that a research agent should generate more designs. It is that it should know when a design is an actual output, when a structural claim is too weak, and when the next useful step requires a different kind of evidence.

What we can claim next will depend on real experiments

The next milestones are clear. We need to reconcile the exact HLA assay construct; obtain the individual Proteinbase novelty/acceptance details; run independent, protein-appropriate structural validation and refinement on preserved copies; and compare promising candidates with the matched wild-type and peptide-free targets. Ultimately, experiments must measure production, binding, and specificity.

The competition organizers distinguish submission from selection for experimental testing, and selection from measured binding. We have no verified selection notice or binding data, and we are not claiming that these computational proteins have therapeutic activity.

Still, the experiment achieved something useful: original model outputs with traceable sources, a transparent funnel that rejected changed targets, an explicit record of unfavorable geometry, and a real encounter with external platform filtering. That is the evidence we are prepared to discuss now.

In scientific AI, an attractive answer is the beginning of an investigation. A trustworthy workflow is the one that keeps testing it.

Primary sources: Challenge 05 · NVIDIA RFD4-Proteina · RCSB 7OW6 · MolProbity Next guidance · Adaptyv novelty framework

A note on evidence and attribution

This article combines original private provider-response archives and local coordinate calculations with founder-reported and authenticated account observations. The latter do not identify the exact 19 competition-submitted sequences. Molecular-generation technology belongs to the Baker Laboratory/NVIDIA collaboration; PDB 7OW6 is an experimental reference from its depositors; MolProbity and Proteinbase are independent tools operated by their respective organizations. None has endorsed Utilyze's proposed candidates. No original sequence list, proprietary coordinates or provider credential is published here.

For background on the receptor-bound KRAS G12D reference see Poole and colleagues, Nature Communications (2022); the resulting designed proteins are separate computational hypotheses and were not those experimentally characterized T-cell receptors.