Use FLOWR Online

FLOWR: Foundation Model for Joint 3D Ligand Generation & Affinity Prediction

FLOWR (specifically FLOWR.ROOT) is an SE(3)-equivariant flow-matching foundation model built for structure-based drug design (SBDD). Developed by researchers at Pfizer Worldwide R&D and academic partners , FLOWR.ROOT unifies pocket-aware 3D ligand generation with multi-endpoint binding affinity prediction and confidence estimation within a single generative backbone.

Whether you are navigating early-stage hit identification or intensive lead optimization, FLOWR.ROOT serves as an adaptive companion to guide your computational drug discovery campaigns.

What is FLOWR?

FLOWR.ROOT is a state-of-the-art framework that bridges the gap between generating physically realistic 3D molecular structures inside a protein pocket and accurately predicting their potency. Built upon an SE(3)-equivariant flow-matching backbone, the model maps a continuous and discrete transport map to morph prior distributions (noise or fragment anchors) directly into target ligand distributions inside the binding site.

Key Capabilities & Generative Modes

Unlike pipeline architectures that switch between standalone generators and external scorers, FLOWR.ROOT integrates everything natively into one architecture, supporting:

• De Novo Pocket-Conditional Generation: Explore entirely novel chemical matter optimized for the exact shape and contours of your target protein pocket.

• Interaction- & Pharmacophore-Conditional Sampling: Enforce and preserve crucial protein-ligand contact networks.

• Scaffold Hopping and Elaboration: Swap core structural scaffolds (e.g., Murcko scaffold replacements) while seamlessly preserving your active R-groups.

• Targeted Fragment Growing & Replacement: Execute spatially targeted modifications using a unique anisotropic Gaussian prior placement strategy that shifts generation precisely to the local editing site.

Core Technical Features

Natively Integrated Output Heads

The shared ligand decoder features three specialized downstream output heads:

1. Structure Head: Predicts exact atomic 3D coordinates, atom types, partial charges, formal charges, hybridization states, and bond types.

2. Multi-Affinity Head: Explicitly routes and separates predictions across four distinct experimental bioactivity endpoints - pC50, pKi, pKd, and pEC50 - treating them accurately according to their assay conditions rather than as interchangeable labels.

3. Confidence Head: Provides per-atom uncertainty estimates via pLDDT scores to evaluate the geometric reliability of the generated pose.

Advanced Three-Stage Training Pipeline

FLOWR.ROOT is pre-trained and refined using datasets sorted tightly by structural and annotation fidelity:

• Stage 1 (Large-Scale Pre-Training): Formulates broad structural and chemical priors by training on ~1.5 billion small-molecule conformations alongside ~2.5 million mixed-fidelity computational and experimental protein-ligand complexes.

• Stage 2 (High-Fidelity Refinement): Sharpens structural accuracy and initial affinity estimations by finetuning on ~30k highly curated, high-quality experimental co-crystal structures (SPINDR and HIQBIND datasets).

• Stage 3 (Project-Specific Domain Adaptation): Enables fast adaptation to proprietary or narrow assay chemical landscapes via parameter-efficient Low-Rank Adaptation (LoRA) and multi-objective guidance via inference-time importance sampling.

Benchmarked State-of-the-Art Performance

• Unrivaled Validity: Achieves a mean 0.97 PoseBusters-validity score for pocket-conditional ligand generation on demanding benchmarks.

• High-Throughput Accuracy: Matches or exceeds computationally expensive physics-based free energy methods, achieving Pearson correlations of up to 0.86 on industry standard FEP+/OpenFE benchmarks.

• Speed Advantage: Operates 3x faster than AEV-PLIG, 200x faster than Boltz-2, and over 10,000x faster than traditional FEP+ protocols.

What is Tamarind Bio?

Tamarind Bio is a leading computational platform that provides researchers with seamless access to advanced deep learning models and bioinformatics tools for molecular structure prediction, docking, and drug design. By translating cutting-edge, open-source architectures into intuitive, high-performance web applications, Tamarind Bio empowers medicinal chemists and biologists to accelerate their drug discovery pipelines without complex command-line or infrastructure configurations.

How to Use FLOWR on Tamarind Bio

Running your drug design campaigns with FLOWR.ROOT through Tamarind Bio's browser-based graphical user interface (FLOWR.UI) is simple and efficient:

1. Select Your Workflow & Checkpoint: Log into the interface and choose the structure-based drug design (SBDD) track. Select either the base pretrained model or upload/load a project-specific finetuned LoRA checkpoint.

2. Upload and Prepare Target Structures: Upload your protein pocket or whole holo complex (supported via standardized PDB/CIF/SDF inputs). Pockets are automatically framed using an optimized 7Å cutoff radius around the binding site.

3. Configure Generation Settings: Pick your design objective from the dropdown (e.g., De Novo, Scaffold Hopping, or Fragment Growing). For fragment or editing modes, use the interactive 3D viewer to directly highlight and select specific atoms for retention or replacement.

4. Steer Trajectories (Inference-Time Scaling): Set your steering preferences to guide the generation toward desired metrics, such as maximizing binding affinity, adjusting logP, modifying topological polar surface area (TPSA), or enforcing synthetic accessibility.

5. Analyze and Export Results: View generated compounds via automatically rendered 2D diagrams, interactive 3D protein-ligand contact maps (showing detailed hydrogen bonds and salt bridges), and comparative chemical property distributions. Rank your optimized, filtered leads and export them instantly as a comprehensive SDF file.

6. Run Iterative LoRA Active Learning Loops: Directly trigger project-specific finetuning within your active session by supplying available local bioactivity data to adjust model weights to your narrow SAR landscape, then instantly regenerate refined samples.

Source