RangeFinder: Predicting Ligand Property Ranges from Protein Pockets

AI summary

Overview

Research area: Machine learning for drug discovery — specifically structure-based ligand discovery and protein pocket representation learning.

Technical level: Intermediate. The paper assumes familiarity with SE(3)-equivariant networks, attention, and standard evaluation metrics, but the core idea (pocket → property ranges) is easy to grasp.

Scope: RangeFinder is a pocket-level model that predicts intervals over nine physicochemical ligand properties directly from protein pocket structure, chemistry, and evolutionary context, evaluated against six baselines on BioLiP2.

What This Paper Is About

Finding a molecule that binds a protein pocket means searching a vast chemical space for a very small compatible set. Existing machine learning methods either generate individual candidate molecules or score individual protein–ligand pairs, which says little about the broader chemical space a pocket can accommodate. RangeFinder instead predicts, from the pocket alone, a range of values for nine ligand physicochemical properties, which can act as a cheap filter on candidate libraries before docking, generation, or synthesis.

Key Contributions

1. A new task formulation: The paper poses pocket-to-ligand modeling as a range-prediction problem — predicting intervals over nine ligand properties (molecular weight, logP, hydrogen-bond donors, hydrogen-bond acceptors, topological polar surface area, rotatable bonds, formal charge, aromatic rings, and fraction of sp³ carbons) — together with a corresponding metric suite (range-IoU, Reduction, Recall, Reduction-Recall F-Score, latency) and a training/evaluation protocol.
2. Empirical gains over deployable baselines: RangeFinder achieves a range-IoU of 0.464 ± 0.002 and an RRF of 0.543 ± 0.011, outperforming all deployable baselines, while requiring only 12.2 ms per pocket.
3. A diagnostic ablation study: Controlled variants isolate what drives performance, showing that input modality choices and specialized training strategies contribute little, whereas the density and quality of ligand supervision dominate.
4. An open, inexpensive filtering prior: The authors frame the model as a fast pre-screening step, releasing code at https://github.com/alannadels/RangeFinder.

Main Findings

• Best deployable accuracy on range-IoU: RangeFinder scores 0.464 ± 0.002, ahead of APoc (0.420), Feature k-NN (0.409 ± 0.000), Pocket cluster (0.403 ± 0.000), Global average (0.402), and Foldseek (0.390). The margin over APoc is 0.044, a ∼10.5% relative improvement.
• Best deployable RRF: RangeFinder reaches 0.543 ± 0.011 versus APoc (0.424), Pocket cluster (0.338 ± 0.004), Feature k-NN (0.333 ± 0.002), Global average (0.313), and Foldseek (0.278) — a 28.1% relative improvement over APoc.
• Roughly three orders of magnitude faster than the closest structural baselines: 12.2 ms per pocket versus 26.3 s for APoc and 127 ms for Foldseek.
• Nearly matches a non-deployable reference: The target-informed baseline, which clusters pockets using ground-truth ligand properties and so cannot be deployed, scores 0.470 ± 0.005 range-IoU and 0.602 ± 0.012 RRF.
• Trades less library reduction for much higher recall: RangeFinder's reduction is 2.1× with recall 0.557 ± 0.026, whereas APoc achieves 5.3× reduction but only 0.287 recall.
• Signal is modest overall: All methods, including the non-deployable reference, fall within ∼0.08 range-IoU of one another, and the pocket-blind global average alone reaches 0.402.
• Supervision density matters more than architecture: Holding the model fixed, the BioLiP2 variant gains +0.064 over its geometry-blind pocket-cluster baseline (0.467 vs 0.403), while a PDBbind variant gains only +0.006 (0.437 vs 0.431), both at seed 42. The global-average floor is essentially unchanged between datasets (0.402 vs 0.403).
• No single input modality is critical: Removing ESM-2 embeddings reduces range-IoU from 0.467 to 0.464; removing chemistry features also gives 0.464; removing 3D geometry gives 0.455 (and cuts latency to 6.8 ms). Removing any one modality lowers range-IoU by at most 0.012.
• Specialized training does not help: A denoising-pretrained encoder scores 0.461 and a mixture-of-experts (K = 6) scores 0.442, both below the 0.467 of the plain single model.
• Formal charge is an outlier property: In 81% of test pockets the target charge spread is below 0.1, making the interval nearly degenerate. RangeFinder's formal-charge range-IoU is only 0.124, versus 0.485–0.527 (mean 0.506) for the other eight properties; dropping charge raises the score to 0.506, an improvement of 0.042 (∼9.1%).

Methodology in Plain English

Defining the target. For each pocket, the authors gather its known ligands. For each of the nine properties they compute the mean and standard deviation across those ligands and define the target as mean ± 2 standard deviations. The model's job is to output a mean and spread for each property, and the predicted interval is formed the same way — with the absolute value of the predicted spread so the width is never negative. Only pockets with at least three known ligands are used, since fewer ligands give an unreliable range.

Representing the pocket. A pocket is the set of protein atoms within 10 Å of its ligand, kept as a 3D point cloud. Each atom carries a 1326-dimensional feature vector: 46 hand-crafted chemical and structural indicators (element, residue type, bonding flags) concatenated with a 1280-dimensional ESM-2 embedding of its residue.

The network. The model is an SE(3)-Transformer built with e3nn, so predictions are unaffected by rotating or translating the pocket. A linear layer projects each atom's features into 32 rotation-invariant scalar channels, while 16 directional vector channels start at zero. Four equivariant attention layers then operate over a radius graph connecting atoms within 5 Å, with edge directions encoded by spherical harmonics up to ℓ = 1 and distances by 16 Gaussian radial basis functions, plus 4-head attention over the scalar channels. Atom representations are mean-pooled into a 48-dimensional pocket embedding, and a small MLP maps that to 18 interval parameters plus one auxiliary logit predicting whether the pocket's ligands contain aromatic rings.

Training. The main loss is a one-dimensional generalized-IoU loss on the predicted versus target intervals, which keeps providing a gradient even when the two intervals do not overlap; it is a differentiable stand-in for the range-IoU metric actually used at evaluation. Per-pocket losses are averaged over valid properties and weighted by inverse protein frequency so over-represented proteins do not dominate. An auxiliary binary cross-entropy term is added with weight λ = 0.089, chosen by CMA-ES.

Evaluation. Training and testing use BioLiP2, split by protein: 2,627 training pockets, 562 validation, 564 test, corresponding to 10,905 / 2,313 / 3,823 structure-views. Six baselines range from a pocket-blind global average through feature k-NN (k = 50), pocket clustering (K = 6), and structural retrieval via APoc and Foldseek, to the non-deployable target-informed reference. Metrics pair accuracy (range-IoU) with the practical screening trade-off between shrinking a background library and retaining true binders (RRF), plus wall-clock latency.

Why This Matters

Research impact. The paper reframes structure-based ligand discovery from "propose a molecule" or "score a molecule" to "characterize the compatible chemical space," which is a complementary and much cheaper operation. Its most provocative finding is negative-leaning: the gap between a pocket-blind average and a sophisticated equivariant model is small on this data, suggesting that data density — not architecture — is the current bottleneck. That is a useful corrective for a field that often prioritizes model sophistication.

Real-world applications.

• Virtual library pre-filtering: Use predicted property boxes to discard most of a compound library before docking, with a tunable reduction/recall trade-off.
• Constraining generative design: Condition a ligand generator on the predicted ranges so it proposes molecules in a plausible property region rather than an arbitrary one.
• Pocket triage: Rank or prioritize many candidate pockets quickly (12.2 ms each) before committing expensive downstream analysis.
• Fragment or series selection in hit finding: Check whether an existing chemical series falls inside a target pocket's predicted property envelope.

Industry relevance. The latency figure is the commercially meaningful one: at roughly 12 ms per pocket versus 26.3 s for APoc, this class of filter can sit in front of existing docking pipelines without becoming a bottleneck. The honest framing — that the model is a cheap prior, not a replacement for docking or generation — makes it easier to slot into a real discovery workflow rather than compete with one.

Future Directions

• Denser, debiased supervision. Since data density and quality, not architecture, appear to be the binding constraint, the authors point to higher-quality, uncertainty-aware, and less biased pocket–ligand data as the most promising lever.
• Pockets with few or no known ligands. Training and evaluation are restricted to pockets with at least three known ligands; extending to poorly characterized pockets — where range estimates near the three-ligand floor are noisy — would broaden applicability.
• A non-equivariant control. The ablations do not establish whether the SE(3)-equivariant architecture is responsible for the gains; a non-equivariant version would settle this, and the small gaps between leave-one-out variants also motivate testing simpler, cheaper models.
• Disentangling pocket size from pocket chemistry. Because pockets are delimited by their bound ligand, size-related properties such as molecular weight may be partly inferable from pocket dimensions rather than chemistry — an effect the authors do not fully separate out.
• Multi-seed and homology-clustered splits. Splits are currently grouped by UniProt rather than sequence-identity clusters, and ablation studies use a single seed; both are flagged as open issues, along with incorporating cofactors as input features.

Target Audience

Researchers and practitioners in computational drug discovery and protein representation learning who are interested in structure-based screening, equivariant network design, or evaluating what actually drives performance in molecular machine learning. It is also valuable for methodologists who want an example of a task reformulation paired with a task-appropriate metric suite, and for industry teams assessing whether a cheap learned filter is worth inserting upstream of an existing docking pipeline. Readers without a background in geometric deep learning will follow the motivation and results but may need to consult the cited SE(3)-Transformer and e3nn work for the architectural details.

Authors’ abstract

Identifying ligands for a protein target requires searching a vast chemical space for a comparatively small set of compatible molecules. Existing machine learning approaches primarily address this problem by generating candidate ligands or scoring individual protein-ligand pairs, but provide limited information about the broader chemical space compatible with a binding pocket. We introduce RangeFinder, an SE(3)-equivariant transformer pocket-level model, trained and evaluated on BioLiP2 data, that predicts ranges over nine physicochemical ligand properties directly from pocket structure, chemistry, and evolutionary context. Using an intersection-over-union objective and evaluation metric (range-IoU), RangeFinder achieves a range-IoU of $0.464 \pm 0.002$, outperforming all deployable baselines while requiring only $12.2$ ms per pocket. We also quantify our method's ability to reduce the ligand search space while retaining compatible binders; in this regard, RangeFinder achieves a Reduction-Recall F-Score of $0.543 \pm 0.011$ --- once again outperforming all deployable baselines. Ablation experiments indicate that RangeFinder's performance depends more strongly on the density and quality of ligand supervision than on any individual input modality or specialized training strategy. Together, these results position RangeFinder as an inexpensive and effective molecular filtering prior, with demonstrated potential in narrowing the ligand search space of target pockets before docking, generation, or synthesis. The code for this project can be found at: https://github.com/alannadels/RangeFinder.