ProbeMatchDTI

A drug's binding behaviour often turns on something small. Puerarin engages NOS3 through hydrogen bonds at ASN338, TRP356 and GLU361 — three positions in a sequence of more than a thousand residues. Most drug–target interaction (DTI) models refine the compound and the protein over several rounds and then match only the final representations, by which point the dominant scaffold or the global sequence context has absorbed that local signal. ProbeMatchDTI is built on the premise that the intermediate states should stay available until matching happens.

The model works in two stages. IterProbe refines drug and protein along independent paths but retains the contextual candidate produced at every depth, then uses a learned probe to weight those candidates at each SMILES token and residue position before anything crosses between the two entities. BindingProbe then matches the enhanced representations twice over: microscopically, as bidirectional atom–residue correspondences constrained by covalent connectivity and sequence neighbourhoods, and macroscopically, as a probe-weighted comparison of semantic and structure-derived summaries of the whole pair.

The work is from Beijing University of Technology, the Artemisinin Research Center of the China Academy of Chinese Medical Sciences, the National University of Singapore and the University of New South Wales, released in September 2026 as a preprint with code and trained checkpoints.

#Key Features

• Depth-retained entity refinement: IterProbe stores the initial state and every subsequent refinement candidate at each position, so functional-group, local-motif and residue-context patterns stay reachable instead of being overwritten.
• Position- and sample-specific probing: a learned query scores the retained candidates at each index, so which refinement depth matters is decided per molecule and per position rather than fixed by the architecture.
• Two-scale cross-entity matching: BindingProbe computes reciprocal atom–residue correspondence updates, then a macroscopic probe re-weights the semantic and structure-derived pair summaries before they are fused.
• Frozen upstream encoders: proteins and drugs enter as precomputed 1024-dimensional ProtBert-BFD and 768-dimensional ChemBERTa PubChem10M SMILES features built offline, and the released code trains only the probing and matching stack above them. The paper's setup section instead names ProteinBERT and SMILES-BERT as the backbone; the released code requires the ProtBert-BFD and ChemBERTa checkpoints, and the 1024- and 768-dimensional widths match those, so this entry follows the code.
• A checkpoint per training setting: the repository ships weights for BindingDB, DrugBank, Human and C. elegans individually plus one trained jointly across all four, the default for inference.

#Technical Details

Both IterProbe stacks project their language-model features to a width of 256 through grouped 1-D convolutions and run three refinement steps, each producing a candidate that a four-head probe aggregator scores against a learned query. A parallel branch embeds the 26 amino-acid symbols at width 75 and runs a five-layer convolutional encoder alongside message passing over the drug molecular graph, supplying the structure-derived evidence that BindingProbe matches separately from the semantic evidence. A two-class head predicts interaction from the fused pair representation, with auxiliary losses on the two pair summaries and a gradient-reversal discriminator over the four source datasets for joint training.

Across the public benchmarks the model reports AUC-ROC/AUC-PR of 0.991/0.992 on BindingDB, 0.906/0.906 on DrugBank, 0.997/0.997 on C. elegans and 0.992/0.994 on Human, against a strongest prior baseline of 0.971/0.972 on BindingDB and 0.901/0.902 on DrugBank. Ablations attribute the BindingDB gain mainly to BindingProbe, which lifts AUC-ROC from 0.971 to 0.988 alone, and the C. elegans and Human gains to IterProbe. Every reported number is within-dataset: the training entry point exposes per-dataset and jointly trained settings but no cross-dataset transfer split. On a private activity-based protein profiling (ABPP) set it reaches AUC-ROC 0.5917, and its scores fall monotonically as mutations accumulate in XPO2 and RUVBL1.

#Applications

The intended use is target fishing and candidate prioritisation — scoring one compound against a large protein pool and ranking the hits. The paper's case study runs puerarin against a reviewed human UniProt target pool; combining ProbeMatchDTI scores with literature and annotation evidence narrows 500 candidates to 291 while retaining 192 of 239 known targets, raising precision from 0.478 to 0.660 and cutting screening burden by 41.8%. NOS3 and PPARG were prioritised and examined in 100 ns molecular dynamics simulations. The repository's README also describes an agent layer that turns ranked predictions into docking and assay protocols, but that code is absent from the published tree.

#Impact

ProbeMatchDTI's contribution is a mechanism rather than a scale: it argues that what a DTI encoder discards between refinement steps matters as much as what it keeps, and it supplies stage-resolved evidence by tracing a molecular-dynamics-confirmed NOS3 contact back through both probing stages to the SMILES token and residue that carried it, the predicted hotspot landing within 6, 11 and 29 residues of the three contact sites. The margins are large but within-dataset, and the trained head is an ordinary two-class classifier over frozen third-party embeddings, so behaviour beyond the four benchmarks is untested and the work awaits peer review. The released code and checkpoints document the mechanism, but reusing it takes work: the documented inference entry point is missing, networks/utils is a submodule pointer the repository supplies no source for, and the preprocessed feature files it consumes, the ABPP measurements and a license are not part of the release.