Part 6: What Survives Physics
Post-MD interface fingerprinting across 28 consensus PD-L1 mini-binder candidates
Previously: Part 4 ran energy minimization and 1 ns MD on three representative designs to validate the simulation workflow. Part 5 used Boltz-2 to independently assess all 31 designs from sequence alone, identifying 28 consensus candidates — designs where RFdiffusion and Boltz-2 agree on the binder–target interface. This page scales MD to all 28 consensus candidates and asks: which interface contacts persist under an explicit force field, and what does the surviving interaction network tell us about design quality?
Scaling MD from 3 designs to 28
Part 4 demonstrated that our MD workflow produces stable, analyzable trajectories — but it tested only three designs. With 28 consensus candidates in hand, we ran the identical protocol (AMBER14/TIP3P, 1 ns production MD, 300 K, Colab T4 GPU) on every one. Each simulation takes roughly 1–1.5 hours on a T4, so the full campaign required approximately 30–40 GPU-hours across multiple Colab sessions. The notebook is restart-safe: completed simulations are detected on disk and skipped on restart.
All 28 simulations completed successfully with 100 trajectory frames each. Binder RMSD values ranged from 1.3 to 2.8 Å across the equilibrated window, indicating that the designed binder folds remained intact. Target-relative separation and persistent interface contacts showed that the binders also remained associated with PD-L1 throughout the analyzed window.
Post-MD fingerprinting: what survives physics?
The preceding models learn sequence and structural regularities from training data. Molecular dynamics adds a different test: the complexes are propagated under an explicit classical force field and solvent model, allowing the starting interfaces to relax and reorganize through atomic motion. Contacts that persist through this trajectory have therefore passed a qualitatively different filter from the models that generated them.
Parameters
| Parameter | Value | Rationale |
|---|---|---|
| Equilibrated window | Frames 76–99 (0.76–0.99 ns) | Final 24 frames after the RMSD plateau |
| Persistence threshold | ≥ 90% occupancy | Contact present in at least 22 of 24 frames |
| Contact cutoff | 4.5 Å | Minimum heavy-atom distance for each binder–target residue pair |
Original-contact retention: interfaces reorganize and consolidate
For each design, we matched the pre-MD contact set from Part 3 against persistent post-MD residue pairs. On average, 38.5% of the starting contacts survived unchanged. The designs gained an average of 21.2 new persistent contacts while losing 3.2 originals. Using the same residue-pair definition, the persistent post-MD contact networks are therefore larger than the starting-structure contact sets, indicating local interface reorganization and consolidation rather than widespread disengagement.
Original-contact retention by design (ranked)
Retention by hotspot configuration
The top design retained 83% of its pre-MD contacts and gained 28 new persistent contacts. Even the two designs with no surviving original contacts formed 14–26 new persistent contacts. Those binders remained target-associated during the trajectory, but their local contact networks reorganized substantially rather than retaining the original interface.
Binding modes can shift during directed evolution as mutations alter the local interaction network. The analogy here is limited but useful: the designed interface is the starting point, and MD shows how that fixed sequence and structure relax under the selected force field. It does not evolve or optimize the design, but it can reveal which interactions remain compatible with the locally sampled ensemble.
Hotspot engagement: the specified surface remains engaged
The PD-L1 hotspot residues were supplied to RFdiffusion during backbone generation. We therefore asked whether persistent post-MD contacts continued to involve those target positions.
Hotspot contact fate by configuration
Cluster A hotspots show strong retention: hotspot 120 achieves 70% survival and hotspot 122 reaches 67%. Cluster B shows a different pattern: no corresponding pre-MD hotspot contacts survived, but 66.7% of hotspot–design pairs formed new persistent contacts during MD. These designs therefore re-established engagement with the specified Cluster B surface through local interface reorganization.
This result has practical implications for design triage. Cluster A binders follow a retain-and-consolidate pattern: hotspot contacts already present in the static threaded structure remain persistent during MD. Cluster B binders follow a reorganize-and-engage pattern: persistent hotspot contacts emerge only after local relaxation. A scoring pipeline based exclusively on static fingerprints would capture Cluster A engagement but miss the MD-dependent contact reorganization observed for Cluster B. This is a concrete example of the additional information supplied by trajectory analysis.
PD-1 epitope overlap: 70% of contacts target native PD-L1 residues
PD-L1's natural binding partner PD-1 contacts 18 PD-L1 residues. We asked whether the designed binders contact those same target positions. For each designed contact to a PD-L1 residue in the native interface, the contacting binder residue was compared with the PD-1 residue that contacts the same PD-L1 position. RFdiffusion was not instructed to reproduce the native PD-1 contact pattern, so any overlap emerges from the generated interfaces rather than an explicit mimicry objective.
Pre-MD (161 contacts)
Post-MD (653 contacts)
The epitope-overlap fraction decreases modestly from 77.0% to 70.3%, so the persistent post-MD contact network remains concentrated on the PD-1-facing region of PD-L1. Within that larger post-MD contact set, the fraction classified as an exact residue-identity match rises from 4.3% to 15.6%. This means that interface reorganization brings more pre-existing binder residues into contact patterns that resemble the native PD-1 interface; the amino-acid identities themselves do not change during MD.
The increase in exact residue-identity matches from 4.3% to 15.6%, without any change to the binder sequences, indicates that local MD relaxation brings more pre-existing binder residues into contact patterns resembling the native PD-1 interface. The result does not imply that the trajectories independently searched binding space; rather, it shows that native-like local packing and complementarity are compatible with the persistent portions of these designed interfaces.
PD-1 contact-pattern classification: pre-MD vs post-MD
Persistent contact chemistry and amino acid convergence
All 59 residue pairs retained from the starting structures preserved the same chemistry-class annotation because the identities of the participating residues were unchanged. The informative result is therefore not whether a surviving pair changed category, but which chemically complementary residue pairs persisted through the equilibrated trajectory.
Examining which binder amino acids persistently contact each PD-L1 hotspot reveals recurring chemical solutions across independent designs. At hotspot residue 120 (Gly), 11 persistent contacts involve alanine, consistent with small-residue packing against the smallest target position. At hotspot residue 125 (Arg), 10 persistent contacts involve glutamate, a recurring charge-complementarity pattern across different backbones. Hotspot residue 56 (Tyr) draws hydrophobic contacts from Ala, Val, and Leu, while residue 113 (Arg) attracts a mix of glutamate and phenylalanine contacts. These cross-design patterns emerge from the intersection of ProteinMPNN sequence placement and MD persistence filtering.
In directed evolution, repeated recovery of the same amino acid at a functional position can identify a consensus hit. Here, recurring binder residue types against the same PD-L1 hotspot provide an analogous computational consensus signal. These positions are strong candidates for prioritization when defining fixed and diversified sites in a next-round combinatorial library.
Beyond binary contacts: interaction intensity over time
A fixed 4.5 Å cutoff is useful for summary statistics, but it forces a binary view of binding. We therefore used a switching function that assigns a continuous interaction intensity between 0 and 1 as a smooth function of distance: s(r) = 1 / (1 + (r/8Å)⁴). At 5 Å this gives 0.87; at 8 Å, 0.50; and at 12 Å, 0.17.
For each PD-L1 hotspot residue, the notebook computes the minimum heavy-atom distance to every binder residue across all 100 trajectory frames and converts those distances to interaction intensities. The resulting heatmaps preserve the time-resolved contact pattern: persistent horizontal bands indicate stable proximity, while bands that strengthen, weaken, or switch identify local interface reorganization.
Tight interface example — design_16_0_rank6
The dominant hotspot-facing contacts remain strong across the full trajectory, with persistent horizontal bands and limited redistribution after the initial relaxation period. This is the clearest example of a design retaining a compact, well-defined interaction network.
Reorganized interface example — design_15_0_rank2
Persistent contacts remain visible at the Cluster B target residues even though only 25% of the original contact set survives. The trajectory therefore supports substantial local rewiring while preserving engagement with the intended target surface.
Lower-ranked interface example — design_10_0_rank3
PD-L1 residues 18, 20, 120, and 122 remain engaged primarily through binder residues 79–90. The contact network reorganizes during the first ~150–200 ps and then settles into persistent horizontal bands. A secondary interaction network around binder residues 45–59 is broader and less intense, but the dominant hotspot-facing contacts remain well defined.
These heatmaps are conceptually reminiscent of smFRET traces because they preserve time-resolved information rather than collapsing the trajectory into a single contact count. Longer and replicated simulations could reveal local state transitions, rotamer switching, and transient contact networks. Quantitative association or dissociation kinetics would require substantially broader sampling and a dedicated kinetic analysis.
Where does each binder place its interaction patch?
ProteinMPNN and MD become most useful when their outputs are mapped back onto each binder structure. ProteinMPNN entropy identifies positions where the sequence model strongly preferred a narrow set of amino acids. MD identifies binder residues that persistently contact the configured PD-L1 hotspots. Viewing those residue sets together reveals where each design places its functional interaction patch.
The complexes below are aligned through PD-L1 and shown in a shared target-centered coordinate frame. Cycle through designs of the same length and hotspot configuration to see whether the interaction patch remains in one location or moves across alternative binder architectures.
The binder is shown as a molecular surface, with white indicating positions that meet neither criterion. Cornflower blue marks residues with ≥90% occupancy against at least one configured PD-L1 hotspot during the equilibrated MD window. Red marks binder positions with ProteinMPNN entropy below 0.5 bits, and purple marks residues satisfying both criteria. PD-L1 is shown in neutral gray for orientation. All complexes are aligned through PD-L1, and the camera is preserved while cycling through designs.
A persistent blue patch identifies the part of the binder that repeatedly engages the intended target surface. Purple residues are the strongest cross-stage constraints: ProteinMPNN converged on the position and MD retained its hotspot contact. Persistent blue residues that remain sequence-tolerant are candidates for controlled diversification in a next-round library.
Entropy × persistence: a four-quadrant classification
Overlaying ProteinMPNN sequence entropy with MD contact persistence produces four quadrants that classify binder–target contact pairs by sequence constraint and trajectory persistence. The two signals report on different features. Low entropy identifies binder positions where ProteinMPNN strongly preferred a narrow amino-acid set in the context of the supplied complex backbone; that constraint may arise from fold geometry, interface packing, or both. High persistence identifies residue pairs that remain in contact through the equilibrated MD window. Their intersection highlights contact pairs involving sequence-constrained binder positions that also maintain persistent target engagement.
Constrained + persistent pairs (222) involve low-entropy binder positions whose target contacts persisted during MD. These are high-priority candidates for fixed positions, but the pair-level counts must first be collapsed to unique binder residues before constructing a library. Tolerant + persistent pairs (431) involve sequence-tolerant binder positions that maintain target contacts and are the strongest candidates for controlled diversification. Constrained + lost pairs (288) involve low-entropy positions whose starting contacts did not persist; those residues may still be important for fold geometry or other local packing interactions. Tolerant + lost pairs (563) provide the lowest-priority signal for preserving the original contact identity.
The four-quadrant classification combines sequence constraint from ProteinMPNN with contact persistence from MD to support the next design round. After collapsing contact-pair results to unique binder positions, constrained + persistent sites are candidates to hold fixed, while tolerant + persistent sites define the clearest space for controlled diversification. The framework separates residues that are repeatedly selected by the sequence model from residues whose target engagement persists during local relaxation.
What this means for the campaign
Across the 28 consensus candidates, most binders remained structurally intact and target-associated during the 1 ns simulations. Approximately 38.5% of starting contact pairs persisted, while additional persistent contacts formed as the interfaces relaxed. The resulting contact networks remained concentrated on the PD-1-facing region of PD-L1 and continued to engage the RFdiffusion-specified hotspot residues.
The analysis supports four practical campaign decisions. Retain designs with stable folds, continued target association, and strong persistent hotspot-contact networks. Prioritize sequence-constrained binder positions that also maintain persistent target contacts when defining fixed sites. Use sequence-tolerant positions with persistent contacts as the primary space for controlled diversification. Deprioritize designs with weak persistent networks or extensive contact reorganization that is not supported by the other campaign metrics.
Together, these results provide a practical triage layer for the campaign. The next stage — affinity scoring and experimental prioritization — can focus on candidates whose interfaces are supported by sequence design, independent structure prediction, and persistent engagement during local MD relaxation.
Caveats
These 1 ns trajectories probe local relaxation around the starting complexes. They do not establish equilibrium binding stability or absolute affinity. The persistence metrics are comparative triage features for ranking this design set; longer replicated trajectories or dedicated free-energy methods would be required for stronger thermodynamic claims.