31 designs made it through the filter. Where do they land on PD-L1?
Previously, in Part 2: We swept binder length, hotspot configuration, and noise scale across 480 backbones. Thirty-one designs passed structural filtering. This page asks a pragmatic triage question: before spending more compute on minimization and MD, do those passing designs engage a coherent PD-L1 surface, recover intended hotspots, and form interfaces worth carrying forward?
Before quantitative fingerprinting, let's inspect the structures. The viewer loads sequence-threaded complexes: RFdiffusion docked geometry with binder residues renamed to the ProteinMPNN-designed sequence. Chain B is the binder; chain A is PD-L1.
These are sequence-threaded complexes. The viewer is best for docking mode and backbone shape. It is not an energy-minimized full physical model, and sidechain energetics should not be inferred from this view alone.
Alignment RMSD is the Cα RMSD used to place the full-sidechain ESMFold binder onto the RFdiffusion complex.
Ranked by Cα alignment RMSD of the full-sidechain ESMFold binder onto the RFdiffusion complex pose. Click a card to load the corresponding threaded complex above.
Interaction fingerprinting maps residue-level contacts between each passing binder and PD-L1, classifies contact chemistry, and compares hotspot contacts with the native PD-1/PD-L1 interface. The current analysis is deliberately a pre-relaxation triage analysis: it is useful for footprint convergence, approximate hotspot recovery, and prioritizing designs for minimization/MD, but fine-grained chemistry should be recomputed after relaxation.
| Configuration | Hotspot residues |
|---|---|
| Cluster A | 18, 20, 120, 122 |
| Cluster B | 26, 56, 113, 123, 125 |
| Distributed | 18, 20, 26, 56, 113, 120, 122, 125 |
The full-sidechain analysis superimposes ESMFold binder structures onto the RFdiffusion complex by Cα alignment, then measures heavy-atom residue contacts at a 4.5 Å cutoff. This is appropriate for topology and triage. It is not a force-field-relaxed energetic model, and sidechains were not packed in the context of PD-L1.
The designs converge strongly on the PD-L1 region centered around residues 119–122. Residue 120 is contacted by 93.8% of Cluster A designs and 100% of Cluster B and Distributed designs.
Figure 1. Contact frequency by PD-L1 residue. Bars show the fraction of designs in each hotspot configuration contacting each residue at least once. For small groups, 100% means 7/7 Cluster B designs or 8/8 Distributed designs, not a large-sample estimate.
The docking footprint is convergent. Residues 119–122 dominate across all three hotspot strategies. That is a strong pre-relaxation sanity check: the designs are finding a common PD-L1 surface rather than scattering across the target.
Hotspot coverage is real but incomplete. Cluster A performs best, Distributed is intermediate, and Cluster B recovers specified hotspots poorly.
Figure 2. Hotspot coverage using the RFdiffusion hotspot definitions. Full-sidechain analysis increases apparent coverage because sidechains bridge contacts invisible in backbone-only structures. This comparison is useful for triage, but contact chemistry should be repeated after relaxation.
Coverage is partial, not binary. Cluster A reaches 42.2%, Distributed 25.0%, and Cluster B 11.4%. The designs often land near the intended interface but do not recover all specified target residues. Cluster B appears to make larger but less epitope-faithful contact patches.
Not really. At hotspot-contact positions with a native PD-1 comparator, most designed residues are chemically distinct from native PD-1. This is not unexpected — there are multiple ways to achieve similar interactions. This makes them de novo hypotheses rather than PD-1 mimetics.
Figure 3. PD-1 mimicry at hotspot positions. Categories compare designed binder residue chemistry with native PD-1 residues contacting the same PD-L1 position in 4ZQK. This is a sequence/chemistry comparison, not a binding-energy estimate.
The corrected result is more novel than the earlier draft suggested. Across 81 hotspot contacts, 61.7% are novel class, 30.9% preserve the same physicochemical class, and 7.4% are exact amino-acid matches. ProteinMPNN partially recovers native-compatible chemistry, but most hotspot contacts are chemically distinct from PD-1.
Interaction types are residue-class compatibility labels, not geometrically validated interactions. “H-bond-compatible” means the residue identities could support that interaction if the geometry is correct. Charged–charged contacts may be attractive, repulsive, or solvent-exposed depending on sign, geometry, and environment. MD or geometric filters are needed to validate them.
Figure 4. Simplified interaction compatibility profile compared with native PD-1/PD-L1. These are pre-relaxation residue-class labels, not confirmed hydrogen bonds or salt bridges.
The chemistry labels are hypothesis-generating only. Native PD-1/PD-L1 has 17.4% H-bond-compatible contacts; the designs range from 7.5–10.1%. Charged–charged pairs are enriched relative to native in all three design groups, but this does not establish productive salt bridges without sign and geometry validation after relaxation.
Residue contacts tell us where the designs land. Extended metrics ask whether those contacts form substantial and geometrically plausible interfaces. These metrics are most useful here as triage features; the chemically detailed analysis should be repeated on minimized structures.
Figure 5. Extended interface metrics. BSA is computed with BioPython Shrake–Rupley. Normal-complementarity is a custom proxy, not validated Lawrence–Colman Sc. Raw BSA should be interpreted alongside binder length and repeated after relaxation.
Cluster B makes larger but less hotspot-faithful interfaces. Cluster B has the largest mean BSA in this set (697 Ų), even though it contacts specified hotspots poorly. However, all three configurations bury well below the 1,200–2,000 Ų often seen in stable protein–protein interfaces. These are measurable but small contact patches, not mature binding interfaces.
The interface-packing proxy remains below native. The designs average 0.178 on the custom normal-complementarity proxy, compared with 0.287 for native PD-1/PD-L1.
Figure 6. Closest atom-pair anatomy. Each contacting residue pair is represented by its single closest heavy-atom pair. This does not count all atom-pair contacts; it describes the nearest contact geometry for each residue pair.
Nearest-contact anatomy is a pre-relaxation warning signal, not final interface chemistry. In the threaded full-sidechain models, no designed residue-pair contact had a sidechain–sidechain pair as its closest heavy-atom interaction. Instead, nearest contacts were dominated by binder-sidechain to PD-L1-backbone pairs. Cluster B showed the same pattern despite having the largest mean BSA. Because these complexes have not yet been sidechain-repacked or energy-minimized, this should not be interpreted as proof that the designs cannot form sidechain–sidechain interactions after relaxation. Some SC-BB contacts may represent legitimate sidechain-to-backbone hydrogen-bond candidates, while others may disappear after minimization or MD. The right next test is to repeat this analysis after minimization and during MD.
This analysis turns the design set into experimentally testable hypotheses. The binders converge on a plausible PD-L1 surface, but hotspot recovery is partial and most hotspot chemistry is novel. That argues against treating these designs as PD-1 mimetics. Instead, they should be treated as de novo interface starting points.
Interface fingerprinting moved this project from structural filtering to design triage. The 31 passing designs converge on a shared PD-L1 surface patch centered on residues 119–122, indicating that the design process is not producing random docking modes. However, hotspot recovery is partial, interfaces are small relative to typical protein–protein interfaces, and most hotspot-contact chemistry is not PD-1-like. The absence of sidechain–sidechain nearest contacts in the unrelaxed threaded models raises a useful question about whether these helical binders are making shallow, backbone-facing contacts rather than well-packed sidechain interfaces. That question cannot be answered from the current models alone and should be revisited after energy minimization and MD.
One broader hypothesis from this analysis is that mini-binder topology may impose limits on what kinds of protein surfaces are easy to engage. Most of these designs are compact, helical binders. That topology is attractive because helices are stable, designable, and easy for current generative models to produce. But helices also have geometric constraints: they are relatively stiff, locally regular structures. They do not easily kink, fold back, or wrap around a target surface without disrupting the very secondary structure that makes them stable.
That matters for interface design. A compact helical bundle may be well suited to engaging a shallow patch or presenting a small number of sidechains into a defined epitope. But it may be less capable of making extensive contacts across convex surfaces, irregular grooves, or deeper clefts where the binder needs to bend, flatten, or conform around the target. In this dataset, the designs converge on the PD-L1 surface and make measurable but small interfaces (mean BSA 648 Ų), hotspot recovery remains partial, and native PD-1-like chemistry is limited. The pre-relaxation nearest-contact anatomy raises a useful question: are these helical binders mostly presenting sidechain tips toward a target backbone ridge rather than forming a deeply packed, interdigitated interface?
For some target surfaces, the right answer may not be “more filtering” or larger sample sizes of the same helical mini-binder topology. It may require a different binder architecture: loops, beta-rich scaffolds, repeat proteins, antibodies/nanobodies, or even intrinsically disordered regions that fold upon binding. In those cases, induced fit may be the relevant binding mechanism that enables high-affinity binding. The computational question then changes from “which rigid mini-binder docks best?” to “which scaffold class has the right physical degrees of freedom for this interface?”
This is not a conclusion proven by the current analysis, but a hypothesis supported by a consistent pre-relaxation pattern: coherent docking, partial hotspot recovery, limited native-chemistry mimicry, small buried surface areas, and a closest-contact anatomy that differs from native PD-1/PD-L1. The practical takeaway is that topology should be treated as a design variable, not just an output of the generative model.
In a compute-unlimited workflow, detailed interaction chemistry should be measured after sidechain repacking, energy minimization, and molecular dynamics. Here, the goal was more pragmatic: before spending modest Colab-scale compute on all 31 passing designs, use pre-relaxation fingerprinting to decide which models are worth carrying forward.
The full-sidechain models analyzed here remain unrelaxed. A contact classified as H-bond-compatible may have the wrong donor–acceptor geometry; a charged pair may be attractive, repulsive, or solvent-exposed; and nearest-contact anatomy may change after sidechain repacking. The next step is therefore not just to run MD, but to repeat the same fingerprinting workflow after minimization and during MD.
That repeated analysis is the key comparison: which contacts persist, which hotspots remain engaged, which apparent interactions disappear, and whether any designs develop more native-like sidechain packing after relaxation?
Contact definition: two residues are in contact if any heavy atom on the binder residue is within 4.5 Å of any heavy atom on the PD-L1 residue. Each residue pair is counted once; the closest heavy-atom pair is retained for distance and atom-label reporting.
Full-sidechain representation: ESMFold-predicted binder structures are superimposed onto the RFdiffusion docked binder backbone via Cα alignment, then paired with the PD-L1 target chain from the complex pose.
Interaction labels: hydrophobic/aromatic, H-bond-compatible, charged–charged, and mixed are residue-class labels. They are not validated interaction geometries.
Closest atom-pair anatomy: for each contacting residue pair, only the single closest heavy-atom pair is classified as BB-BB, BB-SC, SC-BB, or SC-SC. This does not enumerate all atom-pair contacts within the interface.
Extended metrics: BSA uses BioPython Shrake–Rupley. Closest atom-pair anatomy classifies the nearest atom pair for each residue-pair contact. Interface normal-complementarity is a custom proxy, not validated Lawrence–Colman shape complementarity.