FLOWR: Flow Matching for Structure-Aware De Novo, Interaction- and Fragment-Based Ligand Generation

The paper introduces FLOWR, a novel structure-based framework that combines flow matching with equivariant optimal transport and a curated SPINDR dataset to generate and optimize 3D ligands with superior validity, accuracy, and 70-fold faster inference than state-of-the-art methods, while also offering a versatile multi-purpose variant for fragment-based design without retraining.

The Gist

Imagine you are a master chef trying to invent a new recipe. You have a very specific, empty space in a pot (the protein pocket) where you need to fit a new ingredient (the ligand) perfectly. The ingredient must not only fit the shape of the space but also "taste" right by interacting with the specific spices already in the pot.

For a long time, computer programs trying to design these new ingredients have been like chefs who guess randomly, taste, and then try again thousands of times. This is slow, and often the resulting ingredients are weird, broken, or just don't fit well.

This paper introduces a new, faster, and smarter system called FLOWR, along with a new, high-quality cookbook called SPINDR.

1. The New Cookbook: SPINDR

Before you can teach a computer to cook, you need good recipes. The authors realized that the old cookbooks (datasets used to train AI) were full of errors. They had ingredients that didn't fit the pots, missing spices, or instructions that were contradictory.

To fix this, they created SPINDR. Think of this as a "Gold Standard" cookbook.

• The Cleanup: They took thousands of existing crystal structures (photos of how drugs actually sit in proteins) and scrubbed them clean. They fixed missing atoms, figured out exactly how the ingredients should be charged, and removed duplicates.
• The Result: A massive, high-quality collection of 35,000+ perfect examples of how a drug molecule fits into a protein pocket. This ensures the AI learns from reality, not from messy, broken data.

2. The New Chef: FLOWR

Now that they have a good cookbook, they built a new chef named FLOWR.

• The Old Way (Diffusion): Previous AI chefs worked like a sculptor chipping away at a block of marble. They started with a random cloud of noise and slowly chipped away the bad parts to reveal a molecule. This took a long time and sometimes left the sculpture looking a bit strained or unnatural.
• The FLOWR Way (Flow Matching): FLOWR works more like a GPS navigation system. Instead of chipping away, it draws a straight, direct line from "random noise" to the "perfect molecule."
  • Speed: Because it takes the most direct route, it is incredibly fast. The paper claims it is up to 70 times faster than the previous best methods.
  • Quality: Because it follows a straight path, the resulting molecules are less "strained" (they don't look like they are being twisted into uncomfortable shapes) and fit the protein pocket much better.
  • The "Pocket" Sense: FLOWR has a special "ear" that listens to the protein pocket once at the beginning. It memorizes the shape and chemistry of the pocket and then uses that memory to guide the creation of the molecule. This saves a massive amount of computing power.

3. The Special Tool: FLOWR.MULTI

Sometimes, a chef doesn't just want to invent a totally new dish; they want to modify an existing one. Maybe they want to keep the "sauce" (a specific chemical structure) but change the "garnish" to fit a new spice profile.

FLOWR.MULTI is a versatile mode of the chef that allows for this without needing to retrain the whole system.

• Interaction Targeting: You can tell FLOWR, "Make sure this new molecule touches these specific parts of the protein." The AI will keep those touching parts fixed and only invent the rest of the molecule around them.
• Scaffold Hopping: You can say, "Keep this core shape (scaffold) but change the rest."
• Fragment Growing: You can say, "Start with this small piece and grow a full molecule from it."

The paper shows that using this "targeted" approach results in molecules that are much more likely to actually work with the protein, compared to just guessing randomly.

4. The Results

When the authors tested FLOWR against the current best chefs (like PILOT):

• Validity: The molecules FLOWR created were much more likely to be chemically valid (they wouldn't fall apart).
• Fit: They fit into the protein pockets more accurately.
• Speed: It was dramatically faster (up to 70x speedup).
• Interactions: When asked to recreate specific chemical interactions (like a handshake between the drug and the protein), FLOWR.MULTI was much better at remembering and reproducing those handshakes.

Summary

In short, the authors built a better dataset (SPINDR) to teach the AI, and a smarter, faster AI (FLOWR) to design new drugs. They also added a specialized mode (FLOWR.MULTI) that lets scientists guide the AI to keep specific parts of a molecule while changing others. This makes the process of finding new drug candidates faster, more reliable, and more practical for real-world use.

Technical Summary

Technical Summary: FLOWR – Flow Matching for Structure-Aware De Novo, Interaction- and Fragment-Based Ligand Generation

Problem Statement

Structure-based drug discovery (SBDD) aims to design bioactive compounds by leveraging the three-dimensional structures of biological macromolecules. While traditional methods like molecular docking and virtual screening face challenges in exploring vast chemical spaces and accurately predicting binding poses, recent deep learning approaches, particularly diffusion models, have shown promise. However, existing diffusion-based methods suffer from several limitations: reliance on iterative stochastic sampling leads to prolonged inference times; generated molecules often exhibit strained conformations, uncommon substructures, and reduced drug-likeness; and they frequently struggle with accurate interaction recovery. Furthermore, the evaluation of these methodologies is hindered by the quality of widely used benchmark datasets (e.g., CROSSDOCKED2020), which often contain structural defects, unrealistic ligand-pocket constraints due to rigid cross-docking protocols, and significant data leakage between training and test sets.

Methodology

The authors introduce FLOWR, a generative framework designed for the de novo generation of three-dimensional ligands conditioned on protein pockets. The methodology integrates several advanced techniques:

1. Flow Matching with Equivariant Optimal Transport: Unlike diffusion models that rely on iterative stochastic processes, FLOWR employs continuous and categorical flow matching. It utilizes equivariant optimal transport (EOT) to reduce transport costs between the noise distribution and the data distribution. This approach allows for straighter integration paths and more efficient sampling.
2. Architecture: The model is built upon the SEMLA architecture, an E(3)-equivariant message passing framework. FLOWR extends this by incorporating a dedicated pocket encoder that processes protein pocket features independently of the flow time and noisy ligand. This encoding is computed once and reused, amortizing the cost. The ligand decoder utilizes equivariant cross-attention to integrate pocket information and optionally, interaction profiles.
3. Hybrid Modeling: FLOWR jointly models continuous features (ligand coordinates) and discrete features (atom types, bond orders, formal charges). It uses continuous flow matching for coordinates and discrete flow models for categorical properties, with formal charges predicted directly.
4. SPINDR Dataset: To address data quality issues, the authors present SPINDR (Small molecule Protein Interaction Dataset, Refined). Derived from the PLINDER dataset, SPINDR undergoes a rigorous pipeline including structure refinement (using Schrodinger's protein preparation wizard), protonation state assignment, atomic-resolution interaction inference (via ProLIF), and quality filtering. Crucially, it maintains the PLINDER data splits to minimize train-test leakage and provides explicit interaction profiles.
5. FLOWR.MULTI: An extension of the base model enabling multi-purpose conditional generation. It supports inpainting-style generation where specific fragments (e.g., scaffolds, functional groups, or atoms involved in specific interactions) are fixed (set to $t=1$), while the rest of the molecule is generated. This allows for scaffold hopping, elaboration, and interaction-conditional design without retraining.

Key Contributions

1. FLOWR Model: A flow-based generative model that outperforms state-of-the-art diffusion and flow-based methods in pose accuracy, validity, and interaction recovery, while offering significantly faster inference speeds (up to 70-fold).
2. SPINDR Dataset: A high-quality, curated benchmark dataset of ligand-pocket co-crystal complexes that addresses structural defects and data leakage prevalent in existing datasets like CROSSDOCKED2020 and PDBBind. It includes pre-computed atomic-resolution interaction profiles.
3. FLOWR.MULTI: A versatile framework for targeted ligand generation. It enables the generation of ligands adhering to predefined interaction profiles or chemical substructures (scaffolds, functional groups) in a single model, facilitating fragment-based design strategies like scaffold hopping and linking without computationally expensive re-sampling or retraining.

Results

Empirical evaluations were conducted on both the CROSSDOCKED2020 dataset (as a baseline comparison) and the proposed SPINDR dataset.

• Performance vs. Baselines: On CROSSDOCKED2020, FLOWR surpassed models like POCKET2MOL, DIFFSBDD, TARGETDIFF, DRUGFLOW, and PILOT across all metrics, achieving the highest PoseBusters-validity (0.92) and lowest strain energy.
• SPINDR Evaluation: When compared to PILOT (the strongest competitor) on SPINDR:
  • Validity: FLOWR achieved higher RDKit-validity (0.94 vs. 0.79) and PoseBusters-validity (0.88 vs. 0.71).
  • Pose Accuracy: FLOWR demonstrated superior AutoDock-Vina scores (-6.93 vs. -6.30) and lower strain energies (90.05 vs. 120.10 kcal/mol).
  • Distribution Learning: FLOWR showed lower Wasserstein distances for bond angles, bond lengths, and dihedral angles, indicating better alignment with the test set distribution.
  • Efficiency: FLOWR was approximately 70 times faster than PILOT when using 20 integration steps (compared to PILOT's 500 steps), and still 20 times faster with 100 steps.
  • Interaction Recovery: FLOWR outperformed PILOT in recovering interaction fingerprints (47.1% vs. 43.2% success rate).
• FLOWR.MULTI Capabilities:
  • Interaction-Conditional: Generated ligands with an average interaction recovery rate of 76.1% and improved Vina scores (-7.18) compared to unconditional generation.
  • Fragment-Based: Successfully generated ligands around specific scaffolds and functional groups, maintaining high validity (PB-validity > 0.85) and synthetic accessibility (SA scores ~0.80) while allowing for controlled chemical space exploration.
  • Explicit Hydrogens: While modeling explicit hydrogens reduced validity for both models (due to limited training data coverage), FLOWR maintained comparable performance to PILOT and showed improved strain energy metrics.

Significance and Claims

The paper claims that FLOWR and FLOWR.MULTI represent a significant advancement in AI-driven structure-based drug design. By combining state-of-the-art flow matching techniques with efficient protein pocket conditioning, the framework substantially enhances the reliability and applicability of de novo, interaction-, and fragment-based ligand generation.

The authors emphasize that the SPINDR dataset is a critical resource for the community, providing a robust benchmark that addresses the limitations of current datasets regarding pose quality and data leakage. They position FLOWR.MULTI as a practical tool for hit expansion, hit-to-lead, and lead optimization campaigns, capable of generating ligands that align with specific interaction patterns and chemical constraints without the need for model retraining.

However, the authors remain modest about the current limitations. They acknowledge that generated ligands still exhibit higher strain energies than reference co-crystal structures, that modeling explicit hydrogens remains challenging due to data scarcity, and that the current model operates on static protein structures, ignoring conformational flexibility. They view these models as ideation tools that complement, rather than replace, medicinal chemistry expertise and experimental validation.