🧬 biology

ZetaDial: dialing net charge of protein binders at inference time for therapeutic developability

ZetaDial introduces a post-sampling secant control framework for fixed-backbone ProteinMPNN that dynamically adjusts amino acid sequences to target specific net charges, significantly reducing charge prediction error compared to static bias methods while maintaining acceptable foldability and interface quality.

Original authors: Mohammed Sameer Syed, Tamara Dinneen

Published 2026-09-09✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammed Sameer Syed, Tamara Dinneen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the world of modern medicine, scientists are increasingly learning to design proteins from scratch. These molecules are the workhorses of biology, and when engineered correctly, they can act as precise tools to hunt down diseases like cancer or viral infections. The process of creating them often involves a digital leap: taking the rigid, three-dimensional skeleton of a protein and asking a computer to invent a new sequence of building blocks that will fold perfectly into that shape. This is known as inverse folding. For years, the most successful tools for this task have been able to generate sequences that fit the shape, but they have operated somewhat blindly regarding one critical physical property: the net electrical charge of the resulting molecule.

This electrical charge is not a trivial detail. In the real world, a protein that is too positively or negatively charged can become sticky, clumping together with itself or other molecules, or it might move too quickly through the body to be effective. These issues, known as developability problems, can cause a promising drug candidate to fail long before it reaches a patient. While researchers have long known that controlling this charge is vital, existing computer programs offered no built-in way to dial the charge to a specific target after the design was generated. They could nudge the composition slightly, but they could not guarantee the final product would land in the safe, neutral zone required for a viable therapy.

A new study introduces a method called ZetaDial, which acts as a precise tuning knob for this electrical charge. The researchers built a system that works after the computer has already generated a protein design. Instead of guessing the right settings from the start, the system measures the actual electrical charge of the generated protein. If the charge is too high or too low, the system calculates a small adjustment and asks the computer to try again, repeating this process just a few times until the design hits the exact target value. This approach functions like a thermostat: it checks the temperature, and if it is off, it makes a calculated correction to bring it back to the set point, all without needing to retrain the underlying computer models.

The team tested this method on two different types of protein structures: complex pairs of molecules found in nature and single, newly designed proteins. They compared their new tuning system against older methods that simply applied a single, fixed adjustment to all designs. The results showed that the new, adaptive system was significantly better at hitting the target charge. On the complex natural structures, the new method reduced the average error by more than half compared to the older fixed approach. On the single designed proteins, it performed just as well as the best possible fixed setting, proving that the extra step of checking and correcting was not just a theoretical improvement but a practical one.

However, the study also uncovered a clear trade-off. While the system could successfully tune the charge, pushing the electrical value too far in either direction began to damage the protein's ability to fold correctly. When the researchers tested extreme settings, the designs became unstable and lost their proper shape. Even at moderate settings, some of the designed proteins showed signs of structural weakness, suggesting that while the charge can be controlled, doing so too aggressively risks breaking the very structure the scientists are trying to build. This finding highlights that the ability to control charge is not a free pass; it must be balanced carefully against the need for the protein to remain stable.

The researchers also explored how this method might work in a more complex pipeline designed to create entirely new drug binders. They tested the system on three different targets, including proteins involved in immune response and viral infection. The results were mixed and depended heavily on the specific target. For one target, the system successfully shifted the charge toward a neutral range while maintaining the protein's ability to bind to its target. For another, the results were unpredictable, and for a third, the system failed to produce any strong designs at all. These experiments suggest that while the tool works, its success is not uniform across all types of proteins, and the ideal settings vary from one target to another.

Ultimately, this work demonstrates that it is possible to take a powerful protein design tool and add a layer of precise control over its electrical properties. The new method allows scientists to correct the charge of a design after it is generated, ensuring it meets the strict physical requirements needed for a drug. Yet, the study also serves as a cautionary note: the power to tune these properties comes with limits. The designs remain sensitive to how far the charge is pushed, and the relationship between electrical charge and drug success is complex. The findings offer a new, more controlled way to design proteins, but they also confirm that creating a viable therapeutic requires balancing many competing factors, not just electrical charge.

Technical Summary: ZetaDial

Problem Statement
Inverse-folding pipelines, such as those built on ProteinMPNN and BindCraft, have transformed protein design by generating sequences for fixed backbones. However, these generative models lack a built-in mechanism to condition outputs on specific biophysical properties critical for therapeutic developability, such as net charge. Net charge influences viscosity, clearance, nonspecific interactions, and aggregation. While existing tools allow for global or per-residue amino acid biases, custom losses, or weight adjustments, they do not provide a closed-loop, inference-time feedback mechanism to measure the realized charge of a sampled protein and iteratively correct the bias to hit a specific target setpoint. Current methods either rely on static heuristics or require retraining, leaving a gap between raw generative output and clinically viable therapeutics.

Methodology
The authors introduce ZetaDial, a post-sampling, per-protein controller that operates around fixed-backbone ProteinMPNN without retraining the underlying model.

Steering Primitive: ZetaDial utilizes ProteinMPNN's native capability to accept additive biases on per-position logits. It applies a single signed scalar bias, $\beta$ , which adds to the logits of positively charged residues (Lysine, Arginine) and subtracts from negatively charged residues (Aspartate, Glutamate) at every position.
Closed-Loop Control: Recognizing that a fixed $\beta$ $β$ produces protein-dependent charge shifts, ZetaDial implements an adaptive control loop.
1. Initialization: It probes the protein with a small grid of $\beta$ values to estimate the local slope ( $dq/d\beta$ ) using central differences.
2. Secant Update: Instead of a fixed-slope proportional loop (which risks overshooting, especially on charge-insensitive proteins), ZetaDial uses a guarded secant update. After each new sample, it re-estimates the local slope based on the two most recent $(\beta, q)$ evaluations.
3. Constraints: The update is "guarded" to proceed only when the response is locally monotone; otherwise, it falls back to the initial probe slope. This process converges within a fixed budget of three iterations, selecting the sample closest to the target charge.
Evaluation Framework: The method was evaluated on two datasets:
1. RCSB Complexes: 912 protein-protein complexes (filtered by 30% sequence identity).
2. Cas13 Monomers: 768 de-novo binder monomers.
  Comparisons were made against an optimized global $\beta$ (open-loop) and a per-protein fixed-slope loop. Foldability was assessed via ESMFold self-consistency (scRMSD) and AlphaFold2-Multimer interface metrics (ipSAE, DockQ).

Key Contributions

A Per-Protein Secant Controller: ZetaDial introduces a testable, post-sampling controller that measures realized charge and updates bias adaptively. It differs from prior steering methods by operating as a discrete, per-protein feedback loop around a fixed-backbone sampler rather than continuous analytic bias or retraining.
Descriptive Sensitivity Analysis: The authors characterize the relationship between protein composition/size and charge sensitivity. They demonstrate that sensitivity heterogeneity is associated with calibration gain, suggesting that per-protein controllers are most beneficial when response curves vary significantly across a dataset.
Characterization of Trade-offs: The study systematically maps the trade-off between charge controllability and foldability/interface preservation, quantifying the "cost" of bias magnitude ( $|\beta|$ ) on structural metrics.
Exploratory Integration: The authors demonstrate the injection of this bias into the BindCraft pipeline, observing target-specific responses in predicted binding metrics (i pTM) and success rates.

Results

Charge Control Accuracy: On matched stochastic benchmarks, the secant loop significantly reduced mean absolute error (MAE) compared to a fixed-slope loop.
- RCSB Complexes: MAE reduced from 6.46 (fixed-slope) to 5.17 (secant).
- Cas13 Monomers: MAE reduced from 8.23 (fixed-slope) to 5.57 (secant).
- Compared to an optimized global $\beta$ , the secant loop reduced error on the heterogeneous RCSB set (11.71 to 5.17) but was statistically indistinguishable on the Cas13 set (5.47 vs. 5.57).
Foldability and Interface Impact: Foldability deteriorated as $|\beta|$ $∣ β ∣$ increased.
- Self-Consistency: Median scRMSD increased significantly at $\beta = \pm 3$ .
- Interface Metrics: In a 52-complex cohort, reference-based DockQ declined clearly at $\beta = \pm 3$ . An independent replication on eight high-quality complexes showed paired DockQ declines at all nonzero settings, with the largest drops at $\beta = \pm 3$ .
- Moderate Bias: At $\beta = \pm 1.5$ , DockQ intervals often included zero in the full cohort, but a selected-seed replication showed consistent declines, suggesting target-dependent costs even at moderate settings.
BindCraft Sweeps: In exploratory sweeps against PD-L1, IL-7R $\alpha$ , and RBD, the bias successfully shifted net charge. However, binding outcomes were non-monotonic and target-specific. PD-L1 designs near neutral charge retained high maximum i pTM, while IL-7R $\alpha$ showed non-monotonic success rates, and RBD produced no strong designs at any setting.
Proxies: A fixed-backbone C $\alpha$ -neighbour analysis showed a correlation between absolute net charge and the size of same-sign charge patches, but the authors explicitly note this is a geometric proxy, not a measured electrostatic surface or experimental endpoint.

Significance and Claims
The paper claims ZetaDial provides a bounded improvement in the controllability of net charge during inference.

Primary Claim: The secant controller reduces mean charge-hit error relative to fixed-slope and global-bias baselines, particularly in heterogeneous datasets where protein responses to bias vary.
Modest Scope: The authors explicitly state that the results do not support an unqualified claim of superiority across all contexts or a guarantee of improved experimental developability. The study is a computational demonstration.
Trade-offs: The work highlights that while charge can be tuned, doing so incurs a cost in foldability and interface preservation that scales with the magnitude of the bias.
No Experimental Validation: The paper concludes that wet-lab validation is necessary. The reported "developability" improvements are limited to computational proxies (charge windows, geometric patch sizes) and do not constitute experimental proof of reduced viscosity, aggregation, or improved clearance.

In summary, ZetaDial offers a practical, retraining-free method to steer inverse-folding outputs toward specific charge targets with higher precision than static biasing, but it operates within the constraints of the underlying generative model's foldability limits.

Technical Summary: ZetaDial

More like this