KinetiDiff: Docking-Guided Diffusion for De Novo ACVR1 Inhibitor Design in Fibrodysplasia Ossificans Progressiva

The paper introduces KinetiDiff, a structure-based framework that integrates geometry-complete diffusion with real-time AutoDock Vina gradient guidance to successfully generate potent, synthetically accessible, and diverse de novo inhibitors for the ACVR1 kinase target in Fibrodysplasia Ossificans Progressiva, outperforming both neural proxy and unguided approaches.

The Gist

Imagine you are trying to design a custom key to unlock a very specific, broken door. This door belongs to a rare disease called Fibrodysplasia Ossificans Progressiva (FOP). In people with FOP, a specific protein (a "lock" called ACVR1) is stuck in the "on" position, causing their muscles and tendons to slowly turn into bone, eventually trapping them in their own skeletons.

The goal of this paper is to design a new, perfect key (a drug molecule) that fits this lock perfectly to turn it off, stopping the bone growth.

Here is how the authors, led by high school student Aaryan Patel, built a machine to do this, explained simply:

1. The Problem: The Old Way is Too Slow

Traditionally, scientists try to find a drug by looking at a giant library of millions of existing keys, hoping one fits. It's like searching for a needle in a haystack.

• The New Idea: Instead of searching, why not invent a brand new key from scratch?
• The Tool: They used a type of AI called a Diffusion Model. Think of this AI like a sculptor who starts with a block of pure, chaotic noise (static on a TV screen) and slowly chips away the noise to reveal a perfect statue. In this case, the "statue" is a molecule.

2. The Innovation: The "GPS" for Molecules

The tricky part is that if you just let the AI sculpt randomly, it might make a beautiful statue that doesn't fit the lock at all. Most AI drug designers make the shape first, then check if it fits later. If it doesn't fit, they throw it away and start over. This is slow and wasteful.

KinetiDiff (the new system) does something different. It gives the sculptor a real-time GPS.

• The Metaphor: Imagine the sculptor is in a dark room trying to carve a key. Usually, they carve blindly, then turn on the lights to see if it fits.
• KinetiDiff's Trick: It turns on a "magnetic compass" (called AutoDock Vina) while the sculptor is still carving. Every time the sculptor moves a piece of the molecule, the compass instantly tells them: "You're getting closer to the lock! Move a little to the left!" or "No, that's too far, pull back!"

This "compass" is based on physics. It calculates exactly how well the molecule would stick to the protein lock, guiding the AI to carve a key that fits perfectly as it is being made.

3. The Results: A Master Key

The team ran this process 10,000 times.

• Success Rate: 9,997 of those attempts resulted in valid, usable molecules.
• The Best Key: The top molecule they found fits the lock 19% better than the best key currently known in science (the one found in the crystal structure).
• Practicality: Not only does it fit better, but it's also easy to build in a real lab (it's not too complex) and follows all the safety rules for drugs.

4. The "Shortcut" vs. The "Real Deal"

The authors also tested a "shortcut." They tried to use a fast, smart guess (a neural network) instead of the slow, physics-based compass.

• The Analogy: It's like asking a friend who has seen a thousand keys to guess the shape of a new one (fast), versus actually measuring the lock with a ruler (slow but accurate).
• The Result: The "friend" was fast (60 times faster!), but their guesses were often wrong because they didn't understand the 3D geometry as well as the physics compass. The physics-based "GPS" was the only one that could guarantee a perfect fit.

5. Why This Matters

This paper proves that we can use AI + Physics to design drugs for rare diseases much faster and better than before.

• For FOP: This offers hope for a new treatment that could stop the painful bone growth.
• For Science: It shows that if you guide AI with real-world physics while it learns, you get much better results than just letting it guess.

In a nutshell: They built an AI that doesn't just guess what a drug looks like; it "feels" its way toward the perfect fit in real-time, creating a custom key for a broken lock that could save lives.

Technical Summary

1. Problem Statement

Fibrodysplasia Ossificans Progressiva (FOP) is a rare, debilitating genetic disorder caused by gain-of-function mutations (specifically R206H) in the ACVR1 gene, which encodes the ALK2 kinase. This mutation renders the kinase constitutively active, triggering aberrant bone formation in soft tissues.

• Current Limitations: Existing inhibitors (e.g., saracatinib, zilurgisertib) suffer from off-target effects, incomplete mechanistic targeting, and safety concerns. Traditional medicinal chemistry explores only a narrow slice of chemical space.
• Generative AI Gap: While diffusion models can generate 3D molecules conditioned on protein pockets, most existing pipelines decouple generation from binding evaluation. Molecules are generated first and filtered later via docking, meaning the generative model never "sees" the actual binding affinity during the sampling process. This limits the ability to steer generation toward high-affinity conformations.

2. Methodology: KinetiDiff Framework

The authors propose KinetiDiff, a structure-based framework that integrates a Geometry-Complete Diffusion Model (GCDM) with real-time AutoDock Vina gradient guidance.

A. Core Architecture: Geometry-Complete Diffusion Model (GCDM)

• SE(3)-Equivariance: The model uses an SE(3)-equivariant graph neural network (GCPNet) to ensure that molecular generation is invariant to rotation and translation. It decomposes hidden states into scalar (invariant) and vector (equivariant) channels.
• Pocket Conditioning: The denoising process is conditioned on the 3D structure of the ACVR1 binding pocket (PDB: 3MTF). Residue features are weighted by distance to the binding site centroid and injected into the network layers.
• Diffusion Process: Uses a forward diffusion process with a cosine noise schedule to corrupt molecular coordinates and atom types, and a reverse process to denoise them back to valid structures.

B. Key Innovation: Real-Time Vina Gradient Guidance

The central contribution is the injection of AutoDock Vina docking gradients directly into the reverse diffusion loop.

• Mechanism: At specific timesteps during denoising, the current (partially formed) molecule is scored against the ACVR1 pocket.
• Gradient Calculation: Since Vina is not natively differentiable, gradients are computed via central finite differences ($\hat{g}_t$). For an $N$-atom molecule, this requires $6N+1$ scoring calls per step.
• Guided Update: The predicted clean coordinates ($\hat{x}_0^{(t)}$) are shifted in the direction of improved binding affinity:
  $$x_{t-1} = \hat{x}_0^{(t)} - \lambda_t \hat{g}_t$$
  where $\lambda_t$ is a timestep-dependent guidance strength that activates only during the middle-to-late stages of generation (when geometry is coherent but not yet finalized).

C. Alternative Strategy: HNN-Denovo

To address the computational cost of Vina, the authors evaluated HNN-Denovo, a lightweight neural proxy trained on BindingDB data.

• Speed: Offers a 60-fold speedup per guidance step compared to Vina.
• Limitation: It relies on a differentiable proxy score derived from 3D coordinates (contact counts, distances) to bridge its SMILES-based training to the diffusion space. This proxy showed a low correlation ($r=0.224$) with actual Vina scores, leading to inferior performance in the final metrics.

D. Multi-Objective Ranking

Final candidates are ranked using a composite score balancing:

1. Affinity ($pK_d$): Gaussian target centered at 7.5.
2. Kinetics ($k_{off}$): Box-exponential weighting for optimal dissociation rates.
3. Synthetic Accessibility (SA): Penalizing complex synthesis.

3. Key Results

The framework was evaluated on 10,000 diffusion samples for the ACVR1 target.

• Validity & Diversity:

  • 9,997 valid molecules (99.97% validity rate) were generated.
  • High Diversity: The top 100 candidates showed an internal Tanimoto diversity of 0.790, indicating no "mode collapse" to a single scaffold.
  • Drug-Likeness: 100% of the top 100 candidates were Lipinski-compliant (zero violations).

• Binding Affinity (The "Best" Candidate):

  • The top molecule achieved a Vina score of −11.05 kcal/mol ($pK_d = 8.10$).
  • This represents a 19.2% improvement over the crystallographic reference inhibitor (−9.27 kcal/mol).
  • All top 5 candidates exceeded the reference by >1.2 kcal/mol.

• Synthetic Accessibility:

  • The Vina-Direct guided molecules had a median SA score of 2.26, significantly lower (better) than the HNN-Denovo (4.97) and Multi-Objective (4.37) strategies.
  • Insight: Physics-based docking naturally favors compact, well-organized geometries that correlate with synthetic tractability.

• Ablation Study:

  • Vina-Direct dominated all metrics (affinity, drug-likeness, SA).
  • HNN-Denovo improved over unguided baselines but failed to match Vina-Direct due to the domain mismatch between the proxy score and true docking physics.
  • Multi-Objective (unguided generation + post-hoc scoring) performed the worst, confirming that real-time guidance is superior to post-hoc filtering.

4. Significance and Contributions

1. First Real-Time Integration: This is the first framework to integrate real-time AutoDock Vina gradients directly into the denoising loop of a diffusion model for a therapeutic target, proving that physics-based feedback can steer generative models more effectively than learned proxies.
2. Rare Disease Application: Successfully demonstrated the design of potent inhibitors for FOP, a rare disease with limited known scaffolds, providing a computational pipeline for targets where data is scarce.
3. Superiority of Physics over Proxies: The study empirically demonstrates that while neural proxies (HNN-Denovo) are faster, they suffer from domain mismatches when bridging 1D sequence/SMILES data to 3D coordinate space. Direct physics-based guidance (Vina) yields higher affinity and better synthetic accessibility.
4. Practical Workflow: The framework balances high performance with practical constraints, generating molecules that are not only high-affinity but also synthetically accessible (low SA scores) and chemically diverse.

5. Limitations and Future Work

• Rigid Receptor: The docking assumes a rigid protein receptor, ignoring induced-fit effects and solvation dynamics.
• Lack of ADMET: No assessment of absorption, distribution, metabolism, excretion, or toxicity (ADMET) was performed.
• Computational Cost: The Vina-Direct approach is computationally expensive (seconds per molecule per step), though the HNN-Denovo alternative offers a faster, albeit less accurate, path for early exploration.
• Validation: The results are computational; experimental synthesis and biophysical validation (e.g., ALK2 activity assays) are required for clinical translation.

Conclusion

KinetiDiff establishes gradient-guided geometric diffusion as a practical and powerful tool for de novo drug design. By injecting docking physics directly into the generative process, it overcomes the limitations of post-hoc filtering and neural proxies, delivering high-affinity, synthetically accessible inhibitors for challenging targets like ACVR1.