FLOWR.root

FLOWR.root is a foundation model for structure-based drug design that unifies two tasks usually handled by separate systems: generating three-dimensional small molecules inside a protein binding pocket, and predicting how tightly the generated ligands will bind. Most generative models for pocket-aware ligand design produce geometries but leave scoring to external tools, while affinity predictors do not generate molecules. FLOWR.root couples both in a single equivariant network, and adds a calibrated confidence estimate for its predictions.

The model is an SE(3)-equivariant flow-matching architecture with three output heads — one for molecular structure, one for multi-endpoint affinity, and one for confidence. Flow matching provides a fast, simulation-free generative process that respects the rotational and translational symmetries of molecules in 3D space. FLOWR.root supports de novo generation, pocket-conditioned sampling, and fragment-level manipulation such as scaffold hopping, elaboration, and fragment- and core-growing, and it predicts several potency endpoints (pIC50, pKi, pKd, pEC50). It was developed primarily by the Machine Learning & Computational Sciences group at Pfizer Worldwide R&D, with collaborators at Jagiellonian University Medical College, Helmholtz Munich, and the Technical University of Munich; it was posted as a preprint in October 2025 and published in Nature Communications in July 2026. By joining generation, affinity prediction, and confidence in one pocket-aware model, it makes in-silico ligand design and its downstream evaluation a single coherent workflow.

FLOWR.root is the second generation of the FLOWR line. The original FLOWR, posted to arXiv in April 2025 and published in Nature Computational Science, established the generative core: continuous and categorical flow matching with equivariant optimal transport and a separate protein pocket encoder, released alongside SPINDR, a curated ligand-pocket co-crystal dataset built to address data-quality problems in existing benchmarks. It reported up to 70-fold faster inference than diffusion- and flow-based competitors and introduced FLOWR:multi for interaction- and fragment-constrained sampling without retraining. It generated structures only, however — the affinity and confidence heads that define FLOWR.root are new, making the two architecturally distinct checkpoints rather than versions of one network. The author roster shifted with the handoff: Julian Cremer and Djork-Arné Clevert appear on both papers, while the AstraZeneca and Chalmers coauthors of the original gave way to a Pfizer-centered team. The original repository is no longer maintained and directs users to FLOWR.root.

#Key Features

• Joint generation and affinity prediction: A single model produces 3D ligands and predicts multiple potency endpoints (pIC50, pKi, pKd, pEC50) with an accompanying confidence estimate.
• SE(3)-equivariant flow matching: The generative backbone respects 3D rotational and translational symmetry and samples molecules through fast, simulation-free flow-matching dynamics.
• Flexible conditional modes: Supports de novo design, pocket-conditioned sampling, scaffold hopping, elaboration, fragment-growing, core-growing, and interaction-constrained generation.
• Efficient adaptation and inference-time scaling: Offers parameter-efficient LoRA fine-tuning for domain adaptation and importance-sampling-based inference-time scaling to improve outputs.
• Open code and weights: The repository provides a downloadable pretrained checkpoint, a command-line workflow, and a FLOWR.ui web interface, with an MIT license declared in its README.

#Technical Details

FLOWR.root is an SE(3)-equivariant flow-matching model with roughly 33 million trainable parameters, structured as a shared backbone feeding structure, multi-affinity, and confidence heads. It was trained on large-scale ligand libraries together with mixed-fidelity protein-ligand complexes, then refined on curated co-crystal data. On the HiQBind affinity benchmark it reaches Pearson correlations of 0.92 for pIC50, 0.76 for pKi, and 0.57 for pKd; pocket-conditional generation achieves about 0.97 mean PoseBusters validity on CrossDocked2020 and SPINDR. On the FEP+/OpenFE benchmark it reports RMSE of 0.93 kcal/mol with Pearson 0.86, outperforming Boltz-2 and approaching FEP+. Inference of 100 ligands takes roughly 15 seconds on an H100 GPU. The authors describe the released checkpoint as an early release ahead of a fully converged model.

#Applications

FLOWR.root is aimed at computational chemists and drug-discovery teams doing structure-based design: proposing novel binders for a target pocket, growing or hopping scaffolds around known fragments, and prioritizing candidates by predicted potency and confidence in a single pass. Its LoRA fine-tuning path lets teams adapt the model to proprietary targets or chemistries, and the bundled web UI lowers the barrier for non-programmers to run interactive generation.

#Impact

By integrating 3D generation, multi-endpoint affinity prediction, and confidence estimation into one compact equivariant model — and reporting competitive results against physics-based free-energy methods and Boltz-2 — FLOWR.root points toward tighter, self-contained generative design loops for medicinal chemistry. Open weights, code, and a web interface under a permissive license lower the barrier to adoption, though the released checkpoint is an early one and real-world prospective validation is still ongoing.