Beyond Structure and Affinity: Context-Dependent Signals for de novo Binder Success

This study demonstrates that integrating biology-informed sequence features—specifically those capturing aggregation propensity, topology, and context-dependent signals—significantly improves the experimental success rate of de novo protein binder designs beyond traditional structure- and affinity-based evaluations.

The Gist

Imagine you are a chef trying to invent a brand-new recipe for a dish that has never existed before. You have a super-smart AI assistant that can generate thousands of potential recipes based on how ingredients look and how they taste together (structure and affinity).

The problem? The AI is great at making recipes that look perfect on paper, but when you actually try to cook them in the real kitchen, most of them turn into a burnt, inedible mess. They might taste good in theory, but they fall apart in the pan, or they make the whole kitchen smell bad, or they just don't work with the specific stove you're using.

This paper is about realizing that looking good isn't enough. To make a successful dish, you need to understand the "biology" of the kitchen: how the ingredients behave, how they interact with the heat, and what kind of pot you are cooking in.

Here is the breakdown of the study using simple analogies:

1. The Problem: The "Look Good" Trap

Scientists have been designing new proteins (tiny biological machines) to fight diseases. They used to judge these designs only by how well they fit together like puzzle pieces (structure) and how sticky they are (affinity).

But in real-world tests, most of these "perfect puzzle pieces" failed. Why? Because the scientists were ignoring the context.

• The CAR-T Context: Imagine a protein designed to be a flag attached to a soldier's shield (a T-cell). It needs to be flexible, stick out into the air, and not get tangled.
• The EGFR Context: Imagine a protein designed to be a standalone key that floats freely in a river. It needs to be a tight, solid ball that doesn't fall apart on its own.

The paper argues that a "good flag" looks very different from a "good key," even if they are both trying to unlock the same door.

2. The New Tool: "Biology-Informed" Filters

Instead of just checking if the puzzle pieces fit, the researchers used a new set of tools (AI models trained on nature's own proteins) to ask deeper questions:

• Will this clump together? (Aggregation/Amyloid)
• Is it too floppy or too stiff? (Disorder)
• Does it look like it belongs on a cell surface or inside a cell? (Topology)
• Does it have "chemical tags" that might cause trouble? (PTM sites)

3. The Three Layers of Discovery

The researchers found three types of clues that help predict success:

Layer 1: The Universal Rules (Transferable Signals)

These are rules that apply to every protein, no matter what it's doing.

• The Analogy: "Don't use ingredients that rot easily."
• The Finding: If a protein is likely to clump together (high aggregation), it will fail. This is true whether it's a flag or a key. Low clumping is the #1 sign of success.

Layer 2: The Context Rules (Architecture-Dependent)

These rules flip depending on where the protein is working. What works for a flag is a disaster for a key, and vice versa.

• The Analogy: "A swimmer needs a wetsuit, but a hiker needs boots. If you give the hiker a wetsuit, they'll overheat."
• The Finding:
  • Topography: For the "flag" (CAR-T), the protein needs to look like it belongs on the outside of a cell. For the "key" (EGFR), it needs to look like a tight, compact ball (inside-like).
  • Disorder: The "flag" needs some flexibility (disorder) to swing around. The "key" needs to be rigid and stiff to hold its shape.
  • Disulfides (Staples): The "key" loves having internal staples (disulfides) to stay strong. The "flag" hates them because they might get stuck or misfold in the complex cell environment.

Layer 3: The Specific Warnings (Context-Specific)

These are weird, specific signals that only happen in one specific test.

• The Analogy: "If you are cooking in a windy kitchen, you need a lid. If you are cooking indoors, a lid isn't needed."
• The Finding: In the CAR-T tests, proteins with too many "phosphorylation tags" (chemical markers) tended to make the T-cells die. This didn't happen in the other test. It's a specific warning for that specific job.

4. The Result: A Better Filter System

The researchers tested a new "screening process." Instead of just picking the best-looking puzzle pieces, they applied these new biological filters:

1. First: Throw away anything likely to clump (Universal).
2. Second: Check if it fits the specific job (e.g., is it flexible enough for a flag? Is it stiff enough for a key?).
3. Third: Check for specific red flags (like the phosphorylation tags).

The Outcome: By adding these biological checks, they increased the success rate of finding a working protein from 14% to 39%. That's nearly a 3x improvement.

The Big Takeaway

Designing new proteins isn't just about making a shape that fits. It's about making a shape that survives in its specific environment.

• If you design a protein for a cell surface, treat it like a flag.
• If you design a protein to float freely, treat it like a key.
• Don't use a one-size-fits-all rulebook.

By understanding the "personality" of the protein and the "room" it lives in, scientists can stop wasting time and money on designs that look good on a computer screen but fail in the real world.

Technical Summary

1. Problem Statement

Despite rapid advances in computational protein binder design, the majority of de novo designs fail experimentally. Current evaluation pipelines rely heavily on structure-based metrics (e.g., pLDDT, ipTM, Rosetta energies) and affinity proxies (e.g., docking scores, ProteinMPNN likelihoods).

• The Gap: Large-scale benchmarks (e.g., Bits to Binders, Adaptyv) have revealed a disconnect between in silico structural confidence and in vivo functional success. Simple sequence heuristics often outperform complex structural scoring functions.
• The Missing Layer: Successful binders must satisfy biological compatibility constraints beyond folding and docking: they must be expressible in specific host systems, avoid aggregation/misfolding, and function within specific architectural contexts (e.g., membrane-anchored CAR-T receptors vs. standalone soluble proteins). These constraints are not captured by structure prediction alone.

2. Methodology

The authors performed a retrospective re-analysis of two distinct public benchmarks using biology-informed machine learning (ML) features derived from models trained on natural proteins.

Datasets:

1. Bits to Binders (CAR-T CD20): 11,984 AI-designed 80-aa binders tested in primary human T cells.
   • Context: Membrane-displayed extracellular domain of a CAR28z receptor.
   • Gates: Expression (recovery), Enrichment (proliferation), Depletion (toxicity/dysfunction).
2. Adaptyv EGFR: 603 binder designs (13–250 aa) tested as standalone proteins.
   • Context: Cell-free expression and bio-layer interferometry (BLI).
   • Gates: Binding (affinity), Expression.

Feature Generation:
The authors utilized the Orbion Astra suite of ML models to generate sequence descriptors for every candidate. These models were trained on natural proteins but applied here as "biology-informed descriptors" to identify sequence patterns associated with biological compatibility:

• AstraUNFOLD: Intrinsic disorder, amyloidogenicity, membrane topology.
• AstraPTM2: Post-translational modification (PTM) site predictions (40 types).
• AstraSUIT2 & AstraROLE2: Protein classification (taxonomic, subcellular, functional).

Statistical Analysis:

• Features were tested at each experimental gate using Mann-Whitney U (continuous) and Chi-squared (categorical) tests.
• FDR correction (Benjamini-Hochberg) was applied.
• Cross-Benchmark Classification: Signals were categorized into three layers:
  1. Transferable: Same direction of association in both datasets.
  2. Architecture-Dependent: Significant in both but with opposite directions (due to different structural requirements).
  3. Context-Specific: Significant in only one dataset or gate.

3. Key Contributions

1. Identification of Three Signal Layers: The study establishes that binder success is driven by a hierarchy of signals: universal aggregation risks, architecture-dependent structural requirements, and context-specific assay biology.
2. Demonstration of Context-Dependency: It proves that features beneficial in one architectural context (e.g., membrane display) can be detrimental in another (e.g., standalone folding), challenging the notion of universal "good" sequence properties.
3. Quantifiable Screening Improvement: The authors demonstrate that stacking biology-informed filters on top of existing heuristics significantly improves hit rates in retrospective analysis.
4. Framework for Pre-Synthesis Screening: Proposes a layered, context-aware screening workflow to reduce wasted synthesis and testing.

4. Key Results

A. Transferable Signals (Universal)

• Low Aggregation Propensity: Lower amyloidogenicity is the most robust signal, significantly associated with success in both CAR-T and EGFR benchmarks.
• PTM Site Density: Higher predicted PTM site counts correlate with success in both datasets. However, in the heterogeneous EGFR dataset, this signal is partially confounded by sequence length, whereas it is a strong independent predictor in the fixed-length CAR-T set.

B. Architecture-Dependent Signals (Flipping Directions)

Features that are predictive in one context but predictive of failure in the other:

• Topology:
  • CAR-T: "Outside-like" topology (membrane-exposed character) correlates with enrichment.
  • EGFR: "Inside-like" topology (compact, globular character) correlates with binding.
• Disorder:
  • CAR-T: Higher C-terminal disorder correlates with enrichment (likely due to hinge flexibility requirements).
  • EGFR: Low disorder (rigid, ordered folds) is the strongest predictor of binding success.
• Disulfides:
  • CAR-T: Presence of disulfide potential correlates with failure (risk of misfolding in multi-domain constructs).
  • EGFR: Disulfide potential correlates with success (stabilizing standalone folds).

C. Context-Specific Signals

• CAR-T Depletion: High phosphorylation site density is strongly associated with T cell depletion (toxicity), a signal not observed in EGFR.
• Expression vs. Enrichment Tradeoff: In CAR-T, features like "Inside topology" and "Acetylation sites" flip direction between the Expression gate and the Enrichment gate, indicating that optimizing for one stage may compromise the other.

D. Filter Performance (CAR-T Benchmark)

Applying a stacked filter on a controlled subset (cysteine-free, K+E < 30%):

1. Baseline: 13.8% hit rate.
2. Filter 1 (Remove Amyloidogenic): 23.4% hit rate (1.7× lift).
3. Filter 2 (Require Outside Topology > 0.3): 30.1% hit rate (2.2× lift).
4. Filter 3 (Require ≥10 PTM sites): 38.6% hit rate (2.8× lift).

• This stack simultaneously reduced the depletion rate from 1.0% to 0.5%.

5. Significance and Implications

• Beyond Structure: The paper argues that successful binder design requires moving beyond structural plausibility to include biological compatibility (expression, aggregation, and architectural fit).
• Context-Aware Design: There is no "one-size-fits-all" scoring function. Screening rules must be adapted to the deployment context (e.g., membrane-anchored vs. soluble). A feature that is a "green light" for a standalone binder might be a "red light" for a CAR-T domain.
• Practical Utility: Biology-informed descriptors can serve as early warning systems to flag specific failure modes (e.g., aggregation vs. toxicity) before synthesis, saving resources.
• Future Directions: The authors propose integrating these context-aware constraints directly into generative design pipelines rather than using them solely as post-hoc filters. They call for prospective validation to confirm these correlations causally.

Conclusion:
The study demonstrates that biology-informed sequence features provide critical, transferable, and context-dependent signals that complement structural scoring. By adopting a multi-gate, context-aware screening framework, researchers can significantly improve the success rate of de novo binder design and reduce experimental attrition.