Use FLOWR Online
Commercially Available FLOWR No-Code Web Server
FLOWR: Flow Matching Foundation Model for Structure-Based 3D Ligand Generation and Affinity Prediction
FLOWR (the paper's model is named FLOWR.root) is an SE(3)-equivariant flow-matching foundation model that generates 3D ligands inside a protein pocket and predicts their binding affinity in the same network. It was developed by Julian Cremer, Tuan Le, Mohammad M. Ghahremanpour, Emilia Slugocka, Filipe Menezes and Djork-Arne Clevert, with affiliations at Pfizer Worldwide R&D, Jagiellonian University Medical College, Helmholtz Munich and the Technical University of Munich. The arXiv page does not list a code repository or license.
Structure-based drug design usually splits work across separate tools: one to propose molecules, another to dock or score them, and a third to estimate affinity. Generative models also tend to be specialized, so a model built for de novo design often cannot handle fragment growing or lead optimization.
FLOWR addresses this with a single model covering de novo generation, pocket-conditional sampling, fragment-based design and affinity prediction, from hit identification through lead optimization.
Item | Detail from the paper |
|---|---|
Model class | SE(3)-equivariant flow matching with an Euler sampler |
Inputs | Protein pocket (residues within 7 A), optional fragment anchors, number of ligand atoms |
Outputs | Ligand coordinates, atom types, bond orders, charges; pIC50, pKi, pKd, pEC50; confidence |
Size | About 33M trainable parameters |
Project adaptation | LoRA finetuning (rank 16, about 3.3M trainable parameters) |
How FLOWR Generates Ligands and Predicts Binding Affinity
Pocket-conditioned flow: the model transports a prior (noise or fragment anchors for coordinates, uniform for categorical features) to the ligand distribution in the pocket.
Fragment control: a mixed isotropic and anisotropic prior supports fragment growing, replacement and scaffold work.
Separate affinity heads: each affinity type has its own head, trained jointly with generation on a shared backbone.
Inference-time steering: importance sampling steers designs toward higher predicted affinity without an external scorer.
Staged training: pretraining on about 1.5B small-molecule conformations and about 2.5M protein-ligand complexes, then finetuning on SPINDR and HiQBind.
FLOWR Benchmark Results for Affinity Prediction and Ligand Quality
Benchmark | Reported result |
|---|---|
CrossDocked2020 | PoseBusters validity 0.97 |
HiQBind pIC50 | Pearson 0.92, R2 0.85 |
FEP+/OpenFE benchmark (n = 881) | RMSE 0.93, Pearson 0.86, versus Boltz-2 Pearson 0.62 |
Speed (as claimed) | About 200x faster than Boltz-2 |
Project Adaptation with LoRA
Zero-shot R2 was negative on four in-house projects. After LoRA finetuning, R2 ranged from 0.28 to 0.73, and on PDE10A the random split reached Spearman 0.97.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with FLOWR on Tamarind Bio
De novo ligand design: generate 3D molecules shaped to a defined binding pocket.
Fragment-based design: grow, replace or elaborate fragments and scaffolds.
Affinity ranking: predict pIC50, pKi, pKd and pEC50 for candidate ligands.
Selectivity studies: the authors show case studies on CK2a/CLK3, ERa, TYK2 and BACE1.
How to Use FLOWR on Tamarind Bio
Open the tool: log in to tamarind.bio and open FLOWR.
Provide the pocket: upload a protein structure and identify the binding site, ideally a holo conformation.
Choose the mode: de novo generation, fragment-conditioned design or affinity prediction, if exposed.
Set parameters: number of ligand atoms, fragment anchors and number of samples, if exposed.
Run the job: submit and wait for processing.
Download outputs: retrieve generated ligands, predicted affinities and confidence estimates.
Use downstream: prioritize compounds by affinity and confidence, then validate with docking, free-energy methods or assays.
Things to Keep in Mind
The binding pocket must be known, ideally in a holo conformation, and protein flexibility is left to future work.
Zero-shot generalization to new structure-activity landscapes is poor, and project adaptation needs enough assay data.
Water is treated implicitly.
The FEP+/OpenFE and PDE10A benchmarks may overlap with training data, especially for Boltz-2.
Predictions are in silico and need experimental validation.
Source: Cremer J, Le T, Ghahremanpour MM, Slugocka E, Menezes F, Clevert D-A. FLOWR.root: A flow matching based foundation model for joint multi-purpose structure-aware 3D ligand generation and affinity prediction. arXiv:2510.02578 (v6, 2026). https://arxiv.org/abs/2510.02578. No code repository is listed on the page.