Contact-Gated Flow Matching for Induced-Fit Peptide Co-Design

Generative models can now design a peptide binder — choosing its amino-acid sequence and its three-dimensional shape at the same time — but they treat the target protein as a statue. In reality, the target moves: side chains rotate and loops shift to wrap around the incoming binder, a phenomenon biologists call induced fit. The evidence that ignoring this matters is stark. Rigid docking methods that succeed on 57.94% of bound (holo) protein structures collapse to 13.74% when the same proteins are presented in their unbound (apo) shapes (APObind). Meanwhile, the models that do deform the receptor during generation have only ever been built for small-molecule drugs, not for polymeric binders that themselves carry a sequence (FlexSBDD; DynamicFlow). Every peptide co-design system in the literature freezes the receptor; PepFGLD says so outright as its foremost limitation.

The naive fix — put receptor coordinates, binder coordinates, and binder sequence into one synchronous generative flow — has a documented failure mode. THFlow showed that updating everything at once wastes capacity: early in generation the binder is far from the pocket, so fine-grained rotamer and sequence updates have no physical interaction to guide them, and simply splitting the timeline raised stability success from 4.16% to 12.33% over synchronous PepFlow. Adding receptor deformation compounds the problem: deforming the pocket before the binder arrives is unphysical, but waiting too long causes steric clashes. This proposal builds the first peptide sequence-structure co-design model whose receptor deforms during generation, with a contact-gated schedule that decides when each receptor degree of freedom turns on based on whether the binder has actually arrived near it.

Core Claim

I claim that a single equivariant flow-matching model can jointly generate the binder's amino-acid sequence, the binder's full-atom structure, and the receptor's induced-fit deformation, and that the key ingredient is not the union of these variables but the timing rule coupling them: each receptor side-chain torsion and local backbone frame receives gradient updates only once the binder has come within interaction range of that residue. Flow matching, concretely, is a generative method that learns a velocity field carrying noise toward data; here it carries noise on three coupled spaces at once — discrete sequence tokens, binder geometry, and receptor geometry — but on three staggered clocks set by geometry rather than by a fixed schedule. The receptor deformation is what makes this more than an incremental peptide generator: conditioned on an apo target, the model must simultaneously invent a binder and pull the pocket into a bindable state, which rigid-receptor models provably cannot do on non-preformed pockets (APObind).

The gate itself is simple. For each receptor residue $j$, let $d^{\min}_j(t)$ be the distance from its atoms to the nearest binder atom at generation time $t$. Its torsional and frame velocities are scaled by

$$\dot{\chi}_j = g_j(t)\, u_\chi(\cdot), \qquad g_j(t) = \sigma\!\left(-\beta\left(d^{\min}_j(t) - d_c\right)\right),$$

where $\sigma$ is a smooth step function, $d_c \approx 8$ Å is the interaction cutoff, and $\beta$ is a learned sharpness. When the binder is far away, $g_j \approx 0$ and residue $j$ stays put; as the binder approaches, the gate opens and the pocket relaxes around it. The gate is differentiable, so it trains end-to-end and the network can learn how sharply to switch. This answers the "when should the receptor move" question with an emergent, per-residue answer rather than a hand-set global schedule.

Research Program

Three components. First, the parameterization. Following the tension documented across the literature, receptor and binder side chains are modeled as torsion angles on tori and backbone geometry as rigid residue frames in SE(3), rather than free Cartesian coordinates; this prevents the "geometric drift" (broken bond lengths, steric explosion) that unconstrained coordinate updates cause (PepFGLD). This is the two-level split that FlexSBDD validated for small molecules and that PPFLOW and PepFlow use for binder generation; the contribution here is applying it to both sides of the interface at once. Binder sequence evolves as a discrete multinomial flow, as in PepFlow. Second, the network. An E(3)-equivariant graph network with scalar-vector dual representations processes the complex, with covalent receptor and binder edges fixed and a bipartite contact graph between them rebuilt at every generation step from current distances — because the set of receptor residues a peptide touches changes continuously as it folds into place, which fixed edge topologies cannot represent (the dynamic-topology open problem in graph-based molecular modeling). The velocity heads follow FlexSBDD's design, extended with the per-residue gate above; sampling targets roughly 20 steps, matching FlexSBDD's budget.

Third, the data. Paired experimental apo/holo peptide-binding structures are scarce (APObind), so training follows the established synthetic-relaxation recipe: holo peptide complexes from PepBench and PepBDB/Q-BioLip are relaxed with OpenMM and Rosetta into pseudo-apo receptors, and the model learns the apo→holo deformation conditioned on the binder that induces it — the same augmentation FlexSBDD and DynamicFlow used for small molecules. The receptor-deformation head is additionally pretrained on APObind's 10,599 experimental apo/holo pairs as a pure deformation-prediction task. Whether models trained on relaxed pockets generalize to experimental apo structures is itself an open question (Gap 5 in the field), and the evaluation below is designed to measure rather than assume it.

Baselines

The naive reference point is a cascade: run a rigid-receptor co-design model on the apo target, then relax the resulting complex with Rosetta flexible repacking. No reported result found on Lacuna for this exact apo-conditioned cascade, so it will be run as part of the study; its components' failures are documented (below). The strongest existing competitor is PepFGLD, which reports the best PepBench numbers in the file: binding energy $\Delta G$ of -19.86 (previous leaders: -13.46) and a physically plausible binder success rate of 59.35% (baselines 47–52%). It is the strongest competitor because it is the most capable co-design system and the one that explicitly tried to model flexibility — but only of the binder, leaving the receptor rigid. It will be rerun with its released code under my apo-conditioned protocol (same data splits, same sampling budget, same tuning sweep), because its reported numbers were not obtained under apo conditioning and an unfair conditioning mismatch would invalidate the comparison. Beating it means a higher plausible-binder success rate and a more negative $\Delta G$ on identical apo inputs, with paired significance tests. The second strongest is THFlow, which reports 12.33% stability success and 17.81% binding-affinity success on PepBDB/Q-BioLip — it is the closest prior work on scheduling and its numbers anchor the PepBDB-style evaluation. Three ablations isolate the proposal's ingredients: freezing receptor degrees of freedom (isolates induced fit); replacing the contact gate with a fixed clock schedule (isolates gating); and collapsing everything into one synchronous flow (the naive combination, predicted to fail). All runs share parameter count, sampling steps, and tuning budget.

Why not just…?

Why not just cascade a backbone generator, relaxation, and inverse folding? Because the cascade's failure is quantified. On PepBDB/Q-BioLip, decoupled RFDiffusion produced 4.17 Å Cα RMSD and a 26.71% binding-site ratio, versus 2.07 Å and 86.89% for joint co-design (PepFlow). Proteus diagnoses why: backbones generated without sequence carry unrealistic strains that threading then cannot pack. Relaxation after the fact cannot reopen decisions the generator already made. The cascade is nevertheless run as the naive baseline, so the comparison is empirical, not rhetorical.

Why not just co-design against the rigid pocket and repack receptor side chains afterward? Post-hoc repacking can resolve clashes inside a cavity that already exists, but it cannot open a pocket that requires coordinated motion during design. APObind shows the apo-conditioned collapse (57.94% → 13.74%); FlexSBDD shows in-generation flexibility cutting clashes from 6.43 (DecompDiff) to 1.39 and improving average Vina score from -8.39 to -9.12, and shows on KRAS$^{\mathrm{G12C}}$ that a single histidine side-chain rotation opens the cryptic switch-II pocket that rigid models are blind to. The cryptic-pocket experiment below is designed to settle this directly for peptides.

Why not just put receptor coordinates, binder coordinates, and sequence into one synchronous flow? This is the most serious alternative, and I will test it rather than argue around it: it is ablation three. THFlow already showed synchronous updating costs roughly 3× in stability success (4.16% vs. 12.33%), and receptor degrees of freedom add more unconstrained noise to early trajectories. But if the synchronous ablation matches the gated model under apo conditioning, the gate is not the key ingredient and the honest conclusion is that the contribution reduces to "receptor degrees of freedom help" — I would report that and adjust the claim.

Why not just use invariant networks with rotation augmentation? Augmentation only approximates symmetry and needs far more data; equivariant construction gives exactness by design, which matters at these dataset sizes (8,365 complexes in PepFlow's benchmark, 10,599 in APObind), as the de novo design survey notes (A Survey of Generative AI for De Novo Drug Design). This is a methodological constraint I accept rather than relitigate.

Proposed Experiments

The primary evaluation is an apo-conditioned version of PepBench: the receptor is presented in its relaxed apo conformation, and success is measured by binding energy $\Delta G$ (against PepFGLD's -19.86), plausible-binder success rate (against its 59.35%), stability and affinity success on the PepBDB/Q-BioLip protocol (against THFlow's 12.33% and 17.81%), re-docking RMSD (against its 4.09 Å), DockQ, and steric clash counts. PepFGLD's "high-flex" subset, where baselines fall below 46% and PepFGLD exceeds 50%, is the sharpest test: flexible receptors are exactly where induced fit should pay. A second metric is receptor recovery RMSD — how close the induced deformation lands to the experimental holo conformation — which no prior peptide system can even attempt. Third, a generalization probe on experimental (not relaxed) apo structures from the APObind pool, directly testing Gap 5. Fourth, a cryptic-pocket case study modeled on FlexSBDD's KRAS$^{\mathrm{G12C}}$ analysis: a pocket physically occluded by a side chain in the apo state, where rigid baselines and post-hoc repacking should fail and the gated model should open it; if all methods succeed, the induced-fit claim is overstated for peptides and I will say so. Finally, as a stretch goal, the two-state Calmodulin setting from AlloGen, where its decoupled score-then-generate pipeline found 5/10 holo-state binders and 0/10 apo-state binders: a model that generates sequence and structure while inducing the fit is the natural tool for designing binders to the unbound state, and even partial success there would be the clearest demonstration that coupling, not post-hoc scoring, is what the field has been missing.