ProChoreo: de novo Binder Design from Conformational Ensembles with Generative Deep Learning

ProChoreo is a generative deep learning framework that leverages multimodal contrastive learning to align protein sequences with conformational ensembles, enabling the design of dynamic, dynamics-informed protein binders that outperform static structure-based approaches.

The Gist

The Big Problem: Proteins Are Not Statues

Imagine you are trying to design a custom key to open a specific lock (a receptor on a cell). In the past, scientists designed these keys based on a single, frozen photo of the lock. They assumed the lock never moved.

But here's the catch: Proteins aren't statues; they are dancers. They wiggle, stretch, twist, and change shape constantly. A lock that looks closed in a photo might actually be opening and closing in real life. If you design a key based only on the "frozen photo," it might not fit when the lock starts dancing.

Most current AI tools for protein design are like photographers who only take one picture. They miss the dance.

The Solution: ProChoreo (The Choreographer)

The researchers created a new AI tool called ProChoreo. Think of it as a Choreographer who doesn't just look at a dancer's pose; they watch the entire dance routine (the "conformational ensemble").

ProChoreo is designed to create new protein "keys" (binders) that fit perfectly not just into one pose of the lock, but into the whole dance the lock does.

How It Works: The Three-Step Dance

1. Learning the Dance (Pretraining)

First, ProChoreo needs to learn the relationship between a protein's "recipe" (its amino acid sequence) and its "dance moves" (how it moves in 3D space).

• The Analogy: Imagine teaching a student to recognize a song. Instead of just listening to the lyrics (the sequence), the student also watches the music video (the molecular dynamics simulation) to see how the singer moves.
• The Tech: The AI uses a "contrastive learning" method. It's like a matching game where it learns to pair the correct recipe with the correct dance routine. It learns that "Recipe A" always leads to "Dance Moves B."

2. The Magic Translator (The Latent Space)

Once the AI has learned the connection, it creates a shared language (a "latent space") where it can translate between a recipe and a dance.

• The Analogy: Think of this as a universal translator. If you give the AI a recipe for a cake, it doesn't just see ingredients; it "sees" the fluffy texture and the way the cake rises. It understands the vibe of the protein, not just the letters.

3. Designing the New Key (Generation)

Now, the AI is ready to design. You give it a target receptor (the lock), and it generates a brand new protein sequence (the key) that is guaranteed to dance in harmony with that lock.

• The Analogy: Instead of guessing a key shape, the AI says, "Okay, this lock does a specific spin and a twist. I will design a key that has the exact right bumps and grooves to slide in while the lock is spinning."

Did It Work? (The Results)

The team tested ProChoreo on two very different "locks":

1. The Sweet Tooth (TAS1R2): A receptor that detects sweetness.
   • They designed a binder and compared it to Brazzein, a natural sweet protein found in a berry.
   • The Result: While Brazzein was a slightly stronger "hug" (higher binding energy), the ProChoreo-designed binder was amazing at mimicking the dance. It successfully triggered the receptor to open up and signal "sweet!" just like the natural protein did. It proved that understanding the movement is just as important as the strength of the hug.
2. The Growth Factor (FGFR2): A receptor involved in cell growth.
   • The Result: The designed binder locked onto this receptor tightly and stayed stable for a long time, proving the method works on different types of proteins, not just sweet receptors.

Why This Matters

This is a huge leap forward. Previous tools were like designing a suit based on a mannequin. ProChoreo designs a suit based on a living, breathing human who moves, sits, and stretches.

By teaching AI to respect the movement of proteins, we can design better medicines, more effective enzymes, and smarter biological tools that actually work in the messy, dynamic reality of the human body.

In short: ProChoreo teaches AI to stop looking at frozen photos and start watching the dance, so it can design the perfect partner to join in.

Technical Summary

1. Problem Statement

Current deep learning frameworks for protein design (e.g., RFdiffusion, ProteinMPNN) and structure prediction (e.g., AlphaFold) predominantly operate on single static conformations. However, proteins are dynamic entities that exist as ensembles of interconverting states, which are critical for their biological function, recognition, and regulation.

• The Gap: Existing generative models fail to incorporate conformational heterogeneity. They design sequences compatible with a single backbone, potentially missing the dynamic features required for high-affinity binding or functional modulation (e.g., stabilizing specific active states of receptors).
• The Challenge: Bridging the gap between sequence-based generative models and the complex, multi-state structural data derived from Molecular Dynamics (MD) simulations.

2. Methodology

The authors introduce ProChoreo, a generalizable framework that integrates conformational ensemble information into the binder design process through a multi-stage pipeline:

A. Data Construction

• MD Ensemble Dataset: Constructed from the ATLAS database (soluble proteins) and GPCRmd (membrane proteins).
  • Membrane Proteins: 117 GPCR trajectories (500 ns each), including CXCR4, 5-HT receptors, and A2AAR.
  • Non-Membrane Proteins: 4,170 soluble protein trajectories (100 ns each).
• Ensemble Generation: MD trajectories were processed using MDAnalysis. Principal Component Analysis (PCA) was performed on backbone atoms, retaining the top 5 components. K-Means clustering (k=10) partitioned the data into distinct conformational states, with representative frames selected for each cluster.
• Design Dataset: Derived from DIPS, filtered for high-quality protein-protein complexes (resolution < 3.5 Å, buried surface area > 500 Ų).

B. Model Architecture

ProChoreo consists of two primary modules:

1. Contrastive Pretraining (Seq2Ens-CLIP):

   • Sequence Encoder: Uses the pre-trained ESM2 3B model to extract residue-level embeddings from amino acid sequences.
   • Ensemble Encoder: Uses an Equivariant Graph Neural Network (EGNN) with Cα coordinates to encode structural ensembles derived from MD simulations.
   • Alignment: A bidirectional contrastive learning objective (inspired by CLIP) aligns the sequence embeddings with the ensemble embeddings. This forces the model to learn a shared latent space where a sequence is close to its corresponding dynamic structural states.
   • Distillation: A student network distills the mapping from sequence to geometry, creating a "geometry-aware" embedding.

2. Autoregressive Generator:

   • Input: A target receptor sequence is tokenized and processed by ESM2.
   • Fusion: The sequence embeddings are fused with the geometry-distilled features (via a trainable linear projection and element-wise addition) to create a fused representation ($F$) that captures both evolutionary context and dynamic structural constraints.
   • Generation: A Transformer-based autoregressive decoder generates the binder sequence ($B$) conditioned on the target ($A$) and the fused representation, predicting residues one by one.

C. Evaluation Pipeline

• Structural Validation: Designed binders are evaluated using Boltz-1 (an open-source alternative to AlphaFold3) to predict complex structures and assess interface quality (metrics: ptm, iptm, pLDDT, pDE).
• Dynamic Validation: Selected complexes undergo Molecular Dynamics (MD) simulations (500 ns for TAS1R2, 100 ns for FGFR2) using GROMACS/CHARMM36 to assess stability and binding free energy (MM-GBSA).

3. Key Contributions

• Novel Framework: ProChoreo is the first framework to explicitly incorporate conformational ensembles into de novo binder design using generative deep learning.
• Multimodal Contrastive Learning: Successfully aligns protein sequences with MD-derived structural ensembles, creating a latent space that captures the intrinsic relationship between amino acid composition and conformational variability.
• Dynamic-Informed Design: Demonstrates that incorporating ensemble data allows the model to design binders that not only bind but also stabilize specific functional states (e.g., active conformations of GPCRs).

4. Results

A. Pretraining Performance

• The contrastive model achieved high retrieval accuracy in bidirectional tasks:
  • Seq2Ens: R@1 = 0.7884, R@10 = 0.9580.
  • Ens2Seq: R@1 = 0.7603, R@10 = 0.9495.
• The pairwise distance distribution showed tight clustering, confirming effective alignment between sequence and structural modalities.

B. Benchmarking Against Baselines

ProChoreo was compared against ProChoreo-ΔAlign (without cross-modal alignment) and PepMLM (sequence-only language model).

• Performance: ProChoreo consistently outperformed baselines across multiple PDB targets.
• Metrics: Achieved a 4–8% improvement in confidence scores over ProChoreo-ΔAlign and up to 12% over PepMLM.
• Key Insight: Removing the alignment module resulted in significant drops in interface precision (iptm) and complex energy (ipDE), proving that ensemble conditioning is critical for generating thermodynamically plausible interfaces.

C. Case Studies

1. Human Sweet Taste Receptor (TAS1R2):
   • Binding: The designed binder showed strong interaction, though slightly weaker affinity (–50.79 kJ/mol) than the natural protein brazzein (–69.99 kJ/mol).
   • Conformational Dynamics: Crucially, the designed binder successfully induced the active-like conformation of TAS1R2. It triggered the characteristic closure of the Venus Flytrap Domain (VFTD) lobe 1 and the outward displacement of transmembrane helix 6 (TM6) from 13.24 Å (apo) to 15.09 Å (binder), mimicking natural activation.
2. Fibroblast Growth Factor Receptor 2 (FGFR2):
   • The designed binder formed a stable complex over 100 ns MD simulation with a binding free energy exceeding –100 kJ/mol, demonstrating robustness across different receptor classes (GPCR vs. RTK).

5. Significance

• Beyond Static Design: ProChoreo moves protein design beyond static "lock-and-key" models, acknowledging that function often depends on the ability to stabilize specific dynamic states.
• Therapeutic Potential: By designing binders that modulate receptor dynamics (e.g., stabilizing active vs. inactive states), this approach offers a pathway to develop more specific therapeutics with reduced off-target effects.
• Scalability: The framework is generalizable, as demonstrated by its success on both membrane proteins (GPCRs) and soluble enzymes, suggesting broad applicability in synthetic biology and drug discovery.
• Future Direction: The authors highlight that integrating experimental ensemble data (NMR, Cryo-EM) and diffusion-based conformational models could further refine the coverage of the protein energy landscape.

In summary, ProChoreo represents a significant paradigm shift in de novo protein design, leveraging the power of contrastive learning to bridge the gap between static sequences and dynamic structural realities.