Cropping the Target Cuts the Compute Cost of AI Protein Binder Design

Designing a binder requires checking many candidates by computer. A Google team examines whether those checks can focus on the contact surface while retaining useful information about the interaction. ^[1]

This is a preprint first submitted to arXiv on 23 October 2026. It has not been peer reviewed.

The finding

Cropping a target to its binding patch

In de novo binder design, a generative model proposes a backbone, a sequence-design step fills in amino acids, and a structure predictor then "refolds" the binder with its target to see whether the designed interface holds. The authors identify that refolding step as the main compute bottleneck ^[1].

Their example is fibroblast activation protein alpha (FAP), a 719-residue enzyme ectodomain. A binder contacts only a small rim around the catalytic opening, yet standard workflows diffuse and refold the whole protein every time. The paper's change, which it calls epitope excision, is to keep only the target residues near the binder and refold that fragment instead ^[1]. The saving depends on size: AlphaFold3 pads inputs to fixed token buckets, and in the authors' 15-target benchmark seven targets already fit the smallest bucket and gain no refolding speedup ^[1].

Computing savings depend on target size

Compute cost shapes which targets a lab will attempt. If cropping holds up elsewhere, big multi-domain proteins become less forbidding to design against, and the same budget screens more designs. The agent result is more tentative. The best arm, with a human setting strategy and the agent handling execution, beat both the researcher-only and fully autonomous arms on hit rate and efficiency, but on one target, and repeat runs of each arm are not reported. The sturdier claim is that cropping saved compute at small cost to interface geometry, with the caveat that targets already small enough for the smallest bucket gain nothing.

Also note that six of the authors are Google employees and may hold Alphabet stock, as their competing-interests statement says. The arXiv v1 text includes no data or code availability statement.

The evidence

Checking cropped protein interfaces

The crop has two parts: target residues with an atom within a fixed 5 angstrom contact radius of the binder, plus a surrounding "halo" shell of extra residues. The authors swept halos of 0, 3, 6, 9, and 11 angstroms, which correspond to total radii of 5 to 16 angstroms ^[1], Methods. The authors report that a window of roughly 85 residues is a safe floor, with the fraction of designs passing their geometry check climbing from 60 to 85 residues and levelling off above that ^[1].

The pipeline uses RFdiffusion for backbones (three-helix-bundle scaffolds), ProteinMPNN for sequences, and AlphaFold3 as the main refolding tool. Boltz-2 and ESMFold2 were also tested. A design counts as a computational "hit" when its interface-confidence score (ipSAE, max version) exceeds 0.45 and the interface RMSD to the intended pose is under 3.0 angstroms ^[1], Methods. A hit here is a prediction, not a measured binder.

The agent is a language-model controller (the paper names Gemini 3.1 Pro). It sees a structured summary, such as hit counts per surface patch and remaining budget, and returns actions like screen, expand, adjust the crop, backtrack, or stop. Three campaigns on FAP were compared: Arm A, a researcher running a fixed pipeline; Arm B, a human choosing the strategy while the agent handles tactical settings; and Arm C, the agent choosing strategy, patches, and hotspots from scratch. All arms shared a budget of 132.87 GPU-hours ^[1] and Table 1.

Predicted hits and yeast-display binders

Does cropping change the answer? On a benchmark of 15 targets and 3,666 designs (11 angstrom halo), the paper reports a pooled median interface RMSD of 0.37 angstroms for cropped refolds, measured against each design's reference pose rather than against a full-target refold, with 88.2% of designs under 1 angstrom ^[1], Figure 2. The direct crop-versus-full-target comparison is a separate, smaller test on 393 experimentally confirmed binders with a measurable interface: the median was 0.29 angstroms for both, and 98.5% (cropped) and 97.2% (full) passed the 3 angstrom gate ^[1]. For FAP, one epitope cluster saw per-unit time cut by 89% for RFdiffusion, 66% for ProteinMPNN, and 88% for AlphaFold3 ^[1].

Cropping is not free. Using the interface score to rank designs, average precision fell from 0.47 with the full target to 0.40 with an 11-angstrom halo, and ROC-AUC from 0.78 to 0.73. Refolding the crop without a multiple sequence alignment recovered some of the loss on targets with deep alignments but hurt performance on the SARS-CoV-2 receptor-binding domain, a shallow-alignment case ^[1].

Does the agent help? Under the shared budget, Arm A produced 1,672 hits from 33,024 designs, a 5.06% hit rate. Arm B produced 1,790 hits from 19,184 designs (9.33%) and finished in 77.18 GPU-hours, which the paper describes as about 42% fewer than Arm A. Arm C produced 1,803 hits from 29,176 designs (6.18%) and stopped itself at 124.43 GPU-hours. Hits per GPU-hour were 12.6, 23.2, and 14.5 for A, B, and C ^[1], Table 1. A fourth ablation arm is also reported in the table. Given only a neutral list of surface residues, the Arm C agent proposed targeting the rim of the catalytic channel ^[1].

Does it bind? The team ordered a 5,000-member yeast-display library with 1,250 designs per arm, built with random nucleotide substitutions (2 to 3 per sequence), sorted it against FAP over three rounds (300, 150, and 75 nM), and sequenced the survivors. They found 5,361 enriched unique sequences, of which 327 matched designs exactly. Those exact matches came from Arm A (50), Arm B (67), Arm C (105), and the ablation arm (105). Four representative clones showed apparent dissociation constants of 3.0 to 9.1 nM ^[1], Figure 6. Because of the deliberate mutagenesis, "enriched" does not always mean "unchanged design": per Figure 6h as labelled, all four clones come from Arm C, and only one is an exact design match; the other three are related sequences that differ from a design. So these affinities do not show that four unmodified designs bind.

Read the per-arm counts carefully: they are shares of the 327 recovered designs, not binding hit rates. The clone provenance above follows the figure labels.

Future directions

The authors themselves list soluble expression and FAP catalytic-inhibition assays as future work, and say their autonomy comparison covers one enzyme target, one scaffold class, and one language-model controller ^[1]. Testing other target types would be useful, especially ones with shallow alignments, where dropping the alignment hurt. Repeating each arm could show how much the 5.06% versus 9.33% gap moves between runs.

The crop itself also needs testing. The paper shows that interface geometry is preserved on the benchmark, but ranking accuracy dips. A pipeline that crops for the bulk screen and refolds only the top candidates on the full target would be a sensible design to test; the paper does not report that.