LigUnity

Computational drug discovery usually splits protein-ligand affinity prediction into two disconnected stages. Virtual screening must rank millions of candidate compounds against a target and therefore favours fast, coarse scoring; hit-to-lead optimization must resolve small potency differences among close analogues and therefore favours slow, accurate physics such as free energy perturbation. LigUnity argues these tasks are complementary rather than incompatible, and trains a single model that handles both by embedding binding pockets and ligands into one shared representation space.

LigUnity was developed by Bin Feng, Zijing Liu, Hao Li, He Cao, Yu Li and Lei Zhang at the International Digital Economy Academy in Shenzhen, with Mingjun Yang and Junjie Zou at XtalPi and Sheng Wang at the University of Washington. The preprint appeared on bioRxiv in February 2025 and the peer-reviewed version was published in Patterns as the cover article of the October 2025 issue.

The model learns at two granularities simultaneously. Scaffold discrimination is a contrastive objective that teaches coarse active-versus-inactive separation, and pharmacophore ranking is a listwise objective that teaches fine-grained ordering of ligands by measured affinity within a single pocket. Together they let one checkpoint be applied zero-shot to targets never seen in training and then, optionally, few-shot fine-tuned when a handful of measurements for the new target become available.

#Key Features

• Shared pocket-ligand embedding space: Pockets and ligands are projected into a common space so affinity can be read off as a similarity, which removes the pose-generation step that dominates docking runtime and makes large-library screening tractable.
• Two-granularity training objective: A contrastive in-batch softmax loss for scaffold discrimination is optimized jointly with a Plackett-Luce listwise ranking loss for pharmacophore ranking, so the model learns both which compounds are active and how actives rank against each other.
• Zero-shot generalization to novel targets: The model is applied to held-out proteins without retraining, and the released screening checkpoint explicitly excludes DUD-E, DEKOIS and LIT-PCBA test proteins from its training set.
• Few-shot adaptation for lead optimization: A handful of measured analogues for a new target can fine-tune the ranking head, closing much of the gap to alchemical free energy calculations at a fraction of the cost.
• Active learning integration: An iterative loop on TYK2 uses model predictions to select the next compounds to evaluate, reaching optimal ligands within a few rounds.

#Technical Details

Both pockets and ligands are encoded with Uni-Mol, an SE(3)-invariant graph transformer over 3D molecular structures; ligand conformers are generated with RDKit. Virtual screening additionally uses a heterogeneous graph neural network head, while optimization tasks rank directly by embedding similarity. Training data comes from PocketAffDB, a curated resource assembled by the authors that joins affinity measurements from ChEMBL v34 and BindingDB 2024m5 to PDB structures: roughly 0.8 million structure-aware affinity data points spanning about 0.5 million unique ligands, 53,406 protein pockets and 26,748 assays.

Across eight benchmarks in six settings, LigUnity outperformed 24 competing methods. On DUD-E and DEKOIS 2.0 it improved the 1% enrichment factor by more than 50% over the strongest baselines, and it led on all 15 LIT-PCBA targets, where the active-to-inactive ratio reaches 1:1,000. Screening ran roughly 106 times faster than Glide-SP. On the Merck and JACS free energy perturbation benchmarks it beat Glide-SP, MM/GBSA and DrugCLIP zero-shot, and few-shot fine-tuning reached an r² of about 0.472 against FEP+ (OPLS4) at 0.528. Code is released under Apache-2.0, checkpoints on Hugging Face under MIT, and the PocketAffDB figshare deposits under CC BY 4.0.

#Applications

The intended user is a computational chemist running a hit-finding or lead-optimization campaign. For screening, LigUnity ranks large commercial or virtual libraries against a target pocket without docking each compound, making ultra-large library triage practical on modest hardware. For optimization, it prioritises analogues within a congeneric series as a cost-efficient stand-in for free energy perturbation, and the active learning setup lets it drive design-make-test cycles. Three Colab notebooks cover screening, direct inference and few-shot fine-tuning, so a team can evaluate the model on their own target without local installation.

#Impact

LigUnity's contribution is showing that virtual screening and hit-to-lead optimization can share a representation rather than requiring separate tools with incompatible speed-accuracy trade-offs. PocketAffDB is a useful artefact in its own right — a structure-linked affinity corpus at a scale that supports pretraining rather than per-target fitting — and its release lowers the barrier for others building affinity models. The caveats are the ones common to this class of model: all reported gains are retrospective and in silico, benchmark targets are covered by existing structural data so extrapolation to genuinely novel pockets remains unproven, and the curation pipeline behind PocketAffDB is not released, so the corpus cannot be regenerated or extended from source.