TDL for Docking, Screening, and Drug-Target Interaction

A review of Suay-García and Falcó (2026)

Primary source: Beatriz Suay-García and Antonio Falcó (Universidad CEU Cardenal Herrera), “Topological deep learning for drug–target interaction, virtual screening, and docking scoring: a practical, benchmark-driven review,” Briefings in Bioinformatics 27(4), bbag370, published 1 July 2026. Journal article

The publisher’s article page is behind a bot challenge that refuses non-interactive clients, so this review was read from the article’s PubMed Central deposit. That deposit is not cited above: Europe PMC has the PMID but has not yet ingested the PMC record, and the citation audit resolves PMCIDs only against Europe PMC, so the identifier would enter the corpus unchecked. The DOI and PMID both resolve and identify the same work.

Access and reuse: the version of record is gold open access under CC BY 4.0 — plain CC BY, with no non-commercial, no-derivatives, or share-alike term. That resolves verbatim-ok in the shared policy table, and this is the first note in the corpus to take the posture that permits.

The reason is specific rather than opportunistic. Most of what follows is assessment, and assessment belongs in our own words. But the central criticisms below are criticisms of the review’s own wording — what it tells a reader its table contains, and what its reproducibility rule actually asks for — and a criticism that paraphrases the sentence it objects to is asking the reader to trust the paraphrase. Those sentences are therefore quoted exactly, marked as quotations, and attributed. CC BY 4.0’s obligations are met by the attribution notice in the frontmatter and by the verbatim licence text at LICENSES/CC-BY-4.0.LICENSE, which this note is the first here to oblige. Everything outside quotation marks is ours; parameter values, dataset names, metric names, and counts are facts and are reproduced as such.

What it is, and what makes it different

This is a narrative review with a checklist attached, and the checklist is the contribution. It covers three tasks — drug–target interaction (DTI) prediction, virtual screening (VS), and docking scoring — and organizes the applied literature into four families: persistent homology as features, hybrid geometric deep learning with topological signals, end-to-end differentiable topology, and multimodal fusion. That taxonomy is competent and unsurprising.

What is not unsurprising is the second half. Sections on benchmarks, splits, metrics, baselines, and ablations set out per-task minimum standards, and they close with a reporting checklist. The review sets its own bar in a single sentence:

topology should be considered compelling only if it is validated under deployment-relevant splits, against strong baselines, with uncertainty reporting and transparent compute/tuning budgets.

That is guiding-principles‘s Be Honest About What Topology Buys, arrived at independently by authors with no connection to this Foundry. It is worth reading the two side by side, because they do not agree in every direction and the disagreements run both ways.

Where the review is stricter than this Foundry

Four places, and each is a live gap rather than a wording difference.

Capacity-matched baselines. Our principle asks for a matched non-topological baseline. The review asks for a capacity-matched one — a baseline of comparable model size, not merely a baseline of a different kind. A topological model beating a smaller conventional model has demonstrated capacity, not topology, and our phrasing does not currently exclude that.

Uncertainty on the difference, not on the endpoints. The review asks for the topology-on versus topology-off delta with paired bootstrap confidence intervals, at least 1000 resamples, and mean ± standard deviation over at least three seeds. Reporting a metric with a confidence interval on each arm is weaker, and it is what our replications currently do.

A leakage path specific to topology. This is the review’s most useful technical observation and the one a general ML checklist would not produce. Pocket geometry can act as a target identifier: distinct sequences can share near-identical binding sites, so a pocket-topology model split by sequence identity can still see the same pocket on both sides and score its own memory. The filtration is a second channel for the same failure — a filtration built on interaction fields can encode target identity or docking-protocol artifacts rather than transferable binding signal. Our version of the rule says identical split and stops there; identical is not sufficient when the feature is the thing the split was supposed to separate.

Ablation beyond on/off. The review treats filtration variants and vectorization variants as minimum requirements alongside topology on/off, on the grounds that a gain that only survives one filtration is a gain attributable to a hyperparameter. topological-deep-learning already says an ablation removing the topological block is the cheapest honest answer; the review’s point is that it is not a sufficient one.

Where this Foundry is stricter

One difference, and it is structural rather than a matter of degree: everything the review prescribes is advice, and nothing in it is a gate. Its standards are recommendations, its checklist is a checklist, and its reproducibility rule carries an explicit opt-out written into the rule itself: “Share code and configuration when possible (or state limitations clearly)”. The checklist row says the same thing in the passive — “Code/configuration shared when possible (or limitations clearly stated).”

guiding-principles distrusts exactly this form under Deterministic Tools Do Deterministic Work: a prose caveat is advisory, and an author can restate one accurately and violate it in the same breath. Here that is not a hypothetical. The review’s own methods matrix has a Code column, and across its eighteen representative methods, sixteen carry an em dash meaning no public repository was identified. Two link to GitHub. The review tabulates that its field is roughly nine-tenths unreproducible and then recommends, in a later section, that code be shared when possible. It never connects the two.

The corresponding difference in this corpus is that a replication is not complete until a named environment has produced the evidence, and that rule is enforced by a schema rather than trusted to an author. The review has no equivalent, and could not have one.

The defect at its centre

A benchmark-driven review is judged on whether it drives benchmarks, and this one does not close the loop.

Its opening guidance tells the reader where to look: “Table 1 summarizes representative methods (construction, integration strategy, and validation setup).” The third of those three is the one to hold on to. Table 1 has seven columns — task, method, family, topological object, filtration and signals, vectorization and learner, code — and none of them is validation setup. No split, no benchmark, no metric, no baseline appears against any of the eighteen methods.

So the review specifies a minimum evaluation standard in one section and, in another, tabulates eighteen methods without recording whether a single one of them meets it. The obvious and most valuable version of this paper — the standard applied retrospectively, each method graded, the field’s compliance rate reported as a number — is the version that was not written. What is left tells a practitioner how to evaluate their own next model and gives them nothing with which to re-read the results already in the literature. For a review whose stated contribution is an evaluation playbook offered “to reduce overoptimistic claims driven by leakage or incomparable setups”, that is the wrong half to leave out.

The illustrative worked example has the same shape at smaller scale. It specifies a complete docking-scoring protocol — PDBbind refined set filtered at 2.5 Å resolution, pockets as heavy atoms within 6 Å of the co-crystallized ligand, element-specific alpha-complex persistent homology in H0–H2 over C, N, O and S with 4–12 Å interaction cutoffs, 50×50 persistence images with cross-validated bandwidth, gradient boosting to pKd, CASF-2016 core set under a 30% sequence-identity target-cluster split, Pearson r with RMSE and Spearman ρ, bootstrap intervals over at least three seeds, a topology-off ECFP baseline and a classical docking-score baseline, and ablations over both filtration and vectorization. Every parameter a reader needs is present. No result is. It is a protocol, offered as an illustration, that nobody ran.

The argument worth keeping

Set the benchmarking apart and there is one genuinely sharp section, on when topological deep learning adds anything over 3D-equivariant models such as SchNet, PaiNN, DimeNet, NequIP and MACE.

The claim is mechanical rather than empirical, which is what makes it useful. Equivariant message passing encodes local geometry — distances, angles, torsions — with a fixed interaction cutoff. A binding-pocket channel that only becomes apparent at a scale larger than that cutoff is therefore invisible to the network, however deep, because recovering it would require aggregating across the whole structure at once. Persistent homology aggregates across scales by construction. The two are also robust to different things: equivariance is robustness to rigid motion, and persistence stability is robustness to coordinate noise.

The complement is stated with its cost attached. Topology discards orientation and metric precision, so two structures with the same diagram can differ in bond lengths, angles, and chirality; it is insensitive to hydrogen-bond geometry and dihedral preference; and in the features-based setting the filtration is a manual design step that an equivariant model does not have. That is a cleaner statement of the tradeoff than either survey already in this corpus manages, and it is the part of the paper most worth citing.

The practical conclusions follow from it and are stated plainly: prefer pocket topology to complex-level topology unless pose uncertainty is explicitly handled; keep homology dimensions low; and treat a gain that vanishes under modest perturbation as fragile. There is also a short section naming conditions under which topology should not be used at all — unreliable 3D, unavailable leakage controls, compute as the binding constraint — which is rarer in a survey of a method family than it should be.

What it means for the docking work here

The issue that proposed this note placed the review on score-docking-poses‘s exact territory. It is adjacent rather than identical, and the boundary is worth stating.

The review separates pose selection — choosing a near-native pose among candidates — from scoring and ranking, and explicitly deprioritizes the first, treating pose noise as a confounder to be controlled rather than as the task. topoqa, topodockq, and the Mold that wraps the open reproduction are interface quality assessment: they rank candidate structures for one target, which is the task the review sets aside. Its docking-scoring guidance is about predicting affinity.

Its evaluation standards still transfer, and two of them bite directly. Target-cluster splits with documented leakage controls are the default it prescribes for anything trained on PDBbind-derived structures. And the pocket-as-identifier warning applies with full force to interface featurizers, which are computed on precisely the geometry that can leak. Neither is a new idea to this corpus — topological-deep-learning already tells a reader to check the split and to ask whether correlations are pooled or per-target — but the review supplies an argument for why the topological case is worse than the general case, which that note does not currently give.

What a reader here should take from it

As a taxonomy it is a competent third entry beside tda-tdl-molecular-sciences and tda-tdl-beyond-persistent-homology, narrower than either and superseding neither. As an evaluation standard it is ahead of us in four specific places, and those are worth adopting rather than admiring.

The gap between what it prescribes and what it demonstrates is the lesson, and it is the same lesson this corpus reached from the other direction. A standard that is written down is not a standard that is met. The review says the right things about baselines, splits, and ablation, and then presents eighteen methods it never checked against any of them — while recording, in a column of its own table, that sixteen of them ship no code with which anyone else could.

Related

topological-deep-learning is the method this reviews, and the review’s equivariant-versus-PH argument belongs beside that note’s account of topology-as-features. guiding-principles holds the baseline rule the review independently reaches. tda-tdl-molecular-sciences and tda-tdl-beyond-persistent-homology are the broader surveys it narrows. score-docking-poses, topoqa, and topodockq are the docking work its standards apply to, across the scoring-versus-pose-selection boundary described above.