Source-linked AI summary
ZetaDial: dialing net charge of protein binders at inference time for therapeutic developability
Mohammed Sameer Syed, Tamara Dinneen
TL;DR
Therapeutic binder design lacks a direct way to set net charge to a target despite charge-related developability criteria. ZetaDial adds post-sampling, per-protein secant feedback around fixed-backbone ProteinMPNN, reducing charge-hit error across matched benchmarks while exposing foldability and interface-cost limitations.
Problem
Inverse-folding and binder-design tools can influence composition but lack a per-protein feedback loop that measures realised charge and corrects it to a setpoint.
Method
ZetaDial uses a guarded secant controller around fixed-backbone ProteinMPNN, updating a scalar bias after each sampled design according to measured charge.
Results
The secant loop reduced mean charge-hit error versus fixed-slope feedback on both RCSB complexes and Cas13 monomers, while improving over matched global bias on heterogeneous RCSB data and remaining indistinguishable on Cas13.
Takeaways & Limitations
ZetaDial demonstrates inference-time per-protein charge control, but stronger bias carries foldability and target-dependent interface costs and binder-transfer results remain exploratory.
Takeaways & Limitations
BindCraft sweeps used small, unequal samples and predicted metrics only, while the selected-seed interface replication cannot estimate a population effect for all 52 complexes.
Abstract
from arXiv · showhide
Net charge is a developability-relevant property of therapeutic binders, linked to viscosity, clearance, nonspecific interaction and aggregation, and antibody screens already use charge-related criteria. Yet inverse-folding pipelines expose no way to set it to a target value. ProteinMPNN and BindCraft offer amino-acid biases, weight choices and custom losses, but neither supplies a per-protein feedback loop that measures realised charge after sampling and corrects it to a setpoint. ZetaDial contributes a post-sampling, per-protein secant controller around fixed-backbone ProteinMPNN. On matched stochastic benchmarks the secant loop reduced mean absolute error relative to a fixed-slope loop on RCSB complexes (5.17 vs 6.46 charge units) and Cas13 monomers (5.57 vs 8.23). Relative to the optimised matched global bias, it cut RCSB error from 11.71 to 5.17 (cluster bootstrap p < 0.001) and was statistically indistinguishable on Cas13 (5.47 vs 5.57). Across 800 eight-protein subsets, sensitivity heterogeneity was associated with calibration gain (Pearson r = 0.79); this is descriptive resampling, not a prospective decision rule. Foldability deteriorated as bias magnitude increased. In the full 52-complex seed-0 analysis, reference-based DockQ declined clearly at +/-3 but not at +/-1.5; a selected five-seed replication on eight complexes showed paired declines at every nonzero setting, but does not estimate the effect for all 52. In exploratory BindCraft sweeps, PD-L1 designs moved toward near-neutral charge at similar maximum interface pTM but with overlapping success-rate intervals; IL-7R-alpha responses were non-monotonic and RBD produced no strong designs. A fixed-backbone C-alpha-neighbour analysis found smaller same-sign charge-patch proxies near neutral charge, but this proxy is not a measured electrostatic surface or experimental developability endpoint.
1. Introduction
ZetaDial adds post-sampling, per-protein charge feedback to inverse-folding and binder-design workflows, addressing the absence of direct setpoint control. The secant loop improved charge-hit accuracy, while stronger bias introduced foldability and interface-related trade-offs.
- Motivation: Net charge is linked to therapeutic-binder developability, but existing tools provide composition controls without a post-sampling per-protein controller.Published charge-screening ranges are empirical heuristics rather than universal criteria for all protein binders.
- Method: ZetaDial measures realised charge after each fixed-backbone ProteinMPNN sample and updates a scalar bias with a guarded secant rule.The approach differs from global biasing and fixed-slope feedback by adapting to each protein’s observed response.
- Calibration analysis: Across 800 overlapping eight-protein subsets, sensitivity heterogeneity was associated with calibration gain at Pearson r = 0.79.The analysis is descriptive resampling, and the proposed warm-start use remains untested prospectively.
- Trade-offs: Foldability declined with stronger bias, while interface-related declines depended on setting and sampling design.In the full 52-complex seed-0 cohort, DockQ declined clearly at ±3; the selected eight-complex replication showed declines at every nonzero setting but cannot estimate the effect across all 52.
- Binder transfer: Exploratory BindCraft sweeps changed realised binder charge, but target-specific binding patterns and small samples did not establish tolerance windows.A fixed-backbone Cα-neighbour proxy was reported separately from experimental developability endpoints.
2. Related Work
ZetaDial belongs to a broader family of inference-time charge-steering methods but targets a narrower operational setting: post-sampling feedback around fixed-backbone ProteinMPNN. Its matched global-bias comparison measures residual error from one open-loop choice rather than all prior guidance methods.
- Controllable generation: Inference-time guidance methods steer protein properties through biased sampling, position-specific bias, rewards, or predictor conditioning without retraining.Examples include smooth net-charge control, pI steering, reward-based resampling, and requested-property conditioning.
- ZetaDial’s scope: ZetaDial measures realised charge after each design and updates a scalar bias with a guarded secant rule in a fixed-backbone ProteinMPNN loop.Its matched global-β comparator evaluates the residual from one open-loop bias choice.
- Related baselines: ProteinMPNN and BindCraft expose bias, weight, and loss controls, but the paper reports no prior matched evaluation of this exact guarded secant controller.The distinction is narrower than claiming that no prior method offers a signed charge target.
- Charge and interface context: Protein supercharging and interface-design studies motivate a trade-off between changing surface charge, solubility, and binding complementarity.The cited background includes increased negative surface charge correlating with solubility and positive patches predicting insolubility.
3. Methods
The methods implement charge steering by adding a signed scalar bias to ProteinMPNN and calibrating it per protein from sampled responses. Evaluation combines matched controller benchmarks, sensitivity modelling, complex-aware refolding, and exploratory BindCraft design sweeps.
- Charge bias: β > 0 pushes designs more positive by adding β to K,R logits and subtracting β from D,E logits, while β = 0 recovers vanilla ProteinMPNN.Net charge is scored as (#K + #R) − (#D + #E).
- Closed-loop calibration: ZetaDial probes a small β grid, estimates local sensitivity, samples iteratively, and uses a guarded secant update for a fixed three-iteration budget.The update re-estimates slope from recent (β, q) evaluations when the response is locally monotone and otherwise falls back to the probe slope.
- Sensitivity modelling: Sensitivity is predicted with a ridge regressor using length and titratable-residue composition, evaluated by grouped cross-validation within each dataset.Probe-free initialization is proposed rather than separately benchmarked as a controller.
- Datasets and metrics: The benchmarks comprise clustered RCSB complexes and 96 de-novo Cas13 binder monomers, with foldability measured by ESMFold self-consistency RMSD.RCSB targets use complex-level charge, whereas Cas13 and BindCraft designs are single chains.
- Complex-aware scoring: Complex-aware evaluation refolds 52 RCSB complexes across β ∈ {−3, −1.5, 0, 1.5, 3} and uses ipSAE plus native-reference DockQ with clustered bootstrap intervals.A five-seed replication was performed on eight complexes selected by baseline DockQ, making it a baseline-quality-selected stochastic replication.
- BindCraft transfer: BindCraft applies the identical bias during MPNN redesign across PD-L1, IL-7Rα, and SARS-CoV-2 RBD, reporting achieved charge, maximum interface pTM, and success rate.Interface residues are held fixed, so the bias acts on non-interface designed positions.
4. Results
ZetaDial’s adaptive secant loop reduced mean charge-hit error under stochastic ProteinMPNN sampling, outperforming the fixed-slope loop on both datasets and global β on RCSB complexes.
- Controller comparison: 5.17 versus 6.46 on RCSB complexes and 5.57 versus 8.23 on Cas13 monomers for secant versus fixed-slope mean MAE.The secant-minus-fixed cluster-bootstrap differences were −1.28 [−2.38, −0.40] and −2.67 [−3.17, −2.16], respectively.
- Controller comparison: The secant loop was closer in 42.9% of individual RCSB cells, so its mean advantage largely reflected reducing the fixed loop’s overshoot tail rather than uniform per-target dominance.This interpretation concerns mean error, not superiority on every target.
- Controller comparison: 11.71 to 5.17 was the RCSB MAE reduction for the secant loop relative to global β, whereas Cas13 showed similar MAE: 5.47 versus 5.57.The clustered secant-minus-global difference was −6.53 [−8.03, −5.04] on RCSB and +0.10 [−0.53, +0.73] on Cas13.
- Evaluation scope: Achieved and target charge correlated at r = 0.95, but band-hit performance remains the direct endpoint when landing inside a developability window is operative.The reported stagewise ablation used deterministic mean charge-response curves, which are optimistic relative to single-sample outcomes.
- Evaluation scope: The matched benchmark contained 912 RCSB and 768 Cas13 protein-target cells, with lower charge-hit error preferred and one sample per evaluation.Scoring was performed per protein where the target was reachable.
4.2. Sensitivity heterogeneity is associated with calibration gain in resampled subsets
Across resampled benchmark subsets, greater heterogeneity in charge sensitivity was associated with larger calibration gains from per-protein control. Composition and size predicted part of sensitivity variation, but the resampling analysis was descriptive and the model was not tested as a prospective warm start.
- Pearson r = 0.79 linked charge-sensitivity heterogeneity with observed calibration gain across 800 overlapping eight-protein subsets.Protein reuse and the two-source-dataset design make this a descriptive mechanism analysis rather than independent validation.
- The sensitivity model leaves unexplained variation, especially for Cas13, so it does not replace feedback control.
- Grouped cross-validated R2 was 0.77 for RCSB complexes and 0.53 for Cas13 monomers when predicting charge sensitivity from composition and size.The estimates could serve as controller initial values, but warm-start performance was not evaluated.
- As |β| increased, median scRMSD rose and the proportion below 5 ˚A fell, with the largest losses at ±3 and substantial Cas13 losses at ±1.5.This charge–foldability trade-off is separate from the sensitivity-calibration analysis.
4.5. Complex-aware refolding shows associated fold and interface changes
Complex-aware analyses associated stronger charge bias with losses in fold and interface metrics. The full 52-complex cohort showed clear DockQ losses only at ±3, whereas a selected eight-complex, five-seed replication showed negative paired DockQ changes at every nonzero setting; sparse conditional analyses did not establish interface preservation.
- In the full 52-complex cohort, ipSAE declined at every nonzero setting, while reference-based DockQ showed clear mean losses at −3 and +3.DockQ intervals at −1.5 and +1.5 included zero.
- Across five new seeds on eight selected complexes, mean paired DockQ changes were negative at every nonzero setting.The selection used seed-0, β = 0 DockQ, so these effect sizes cannot be generalized to all 52 complexes.
- The selected replication reported mean DockQ changes of −0.352 at −1.5 and −0.312 at +1.5, alongside larger losses at −3 and +3.
- Fold-tolerant conditional subsets contained few complexes and produced metric-dependent results, so they did not establish interface preservation.At ±1.5, conditional DockQ intervals included zero while conditional ipSAE intervals did not.
4.6. Exploratory BindCraft sweeps show different predicted responses by target
The BindCraft sweeps changed realised binder charge, but predicted binding responses differed by target and did not establish general charge-tolerance windows or experimental developability.
- PD-L1 shifted from mean charge −5.8 at β = 0 to −2.1 at β = 0.15 and −0.5 at β = 0.3, with maximum i pTM near 0.83.Success rates were 39.3%, 62.5%, and 37.5%, respectively, with overlapping bootstrap intervals between baseline and β = 0.15.
- IL-7Rα responses were non-monotonic: mean charge moved from −3.03 at β = 0 to −3.99 at β = 0.1 before becoming positive at larger settings.Hit rates included 31.3% at β = 0.2, 43.8% at β = 0.5, and 0% at β = 1.0.
- RBD produced no designs above i pTM 0.5 at any setting, with a maximum of 0.32, despite charge shifting.The figure compares maximum and mean interface i pTM and the fraction of designs reaching i pTM ≥0.5 against achieved charge.
- 0 of 28 PD-L1 baseline designs versus 5 of 24 designs at β = 0.15 met both the selected pI/charge window and i pTM >0.5.This is enrichment under finite-sample in silico criteria, not evidence of experimental developability or binding.
- The fixed-backbone Cα-neighbour analysis associated larger absolute net-charge magnitude with larger connected same-sign residue sets, with r = 0.81 for Cas13 and r = 0.89 for antibody Fv domains.The proxy is neither an electrostatic-potential surface nor an experimental developability endpoint.
4.8. Robustness and a surface-restricted variant
Sensitivity checks produced similar Cas13 secant-loop errors across temperatures, while surface-restricted bias showed task-dependent precision and a possible structural-footprint trade-off.
- Cas13 secant-loop MAE was 6.4, 5.6 and 5.6 at sampling temperatures 0.1, 0.2 and 0.3, respectively.These point estimates do not establish temperature invariance without uncertainty and complete output files.
- The surface-restricted comparison does not establish general precision parity because paired uncertainty was unavailable and the underlying result CSVs were absent from the audited snapshot.
5. Limitations
The study is a computational capability demonstration whose interface, BindCraft, charge-window, and spatial-proxy conclusions remain bounded by prediction-only analyses, small selected samples, and unvalidated heuristics.
- The 52-complex seed-0 analysis found fold and ipSAE declines at all nonzero settings, while DockQ declined clearly only at ±3.The eight-complex independent-seed replication used seed-0 DockQ for selection and therefore cannot estimate a population effect for all 52 complexes.
- BindCraft sweeps used small, unequal samples and predicted metrics only; PD-L1 intervals overlapped, IL-7Rα was non-monotonic, and RBD had no strong design.
- The charge windows were antibody-derived heuristics rather than validated cut-offs for small de novo binders.
- The Fig. 7 Cα-neighbour proxy was not an electrostatic-potential calculation or measured developability endpoint, and clustered uncertainty was not established.
Code and data availability
The authors plan public release of the control, analysis, and BindCraft resources, alongside prediction tables, replication outputs, scores, bootstrap summaries, and provenance information.
- The charge-control and per-protein calibration scripts, spatial-developability and antibody-Fv analyses, and charge-controlled BindCraft notebook will be released publicly upon publication.
- The analysis branch includes AlphaFold2-Multimer prediction tables, independent-seed replication outputs, native-reference DockQ scores, paired changes, cluster-bootstrap summaries, and a provenance manifest.
6. Conclusion
ZetaDial improves mean net-charge control under heterogeneous response, but its charge-tuning benefits are bounded by foldability costs and exploratory evidence that does not establish experimental developability gains.
- ZetaDial reduced mean charge-hit error on both benchmark datasets versus per-protein feedback controllers, and improved error on heterogeneous RCSB complexes versus matched global bias.It remained statistically indistinguishable from the matched global bias on Cas13.
- Foldability costs grew with |β|, with native-reference DockQ showing clear interface-structure loss at ±3.A selected-seed replication found substantial target-dependent interface costs at moderate settings but did not estimate effects across all 52 complexes.
- Exploratory BindCraft, Fv, and surface-patch analyses support exploratory rather than definitive conclusions about charge-controlled binder design.These analyses do not establish an experimental developability improvement.