SPRINT

Asking whether a compound has an off-target anywhere in the human proteome means docking it against every structured protein a cell makes — tens of thousands of targets, at seconds to minutes of physics apiece. That cost is why off-target screening is usually narrowed to a shortlist someone already suspected. Vector-based screening changes the arithmetic: if a protein and a molecule are each reduced once to a single vector in a shared space, "do these bind?" becomes a dot product, and a chemical library can be indexed the way a search engine indexes documents.

SPRINT (Structure-aware PRotein ligand INTeraction) is a drug-target interaction model built on that premise by researchers at the University of Pittsburgh and Carnegie Mellon University, first posted as a preprint in November 2024. Its contribution is where the protein vector comes from. Earlier co-embedding methods such as ConPLex average a protein language model's per-residue embeddings into one vector, diluting the pocket — the handful of residues that decide binding — across hundreds of residues that do not. SPRINT replaces the average with multi-head attention pooling, a learned, sequence-dependent aggregation, and takes its per-residue features from SaProt, whose vocabulary pairs each amino acid with a Foldseek structural token so fold information enters the representation without any 3D modelling at inference.

Both changes are ablated: swapping attention pooling back to averaging, or SaProt for ProtBert, each costs accuracy on the LIT-PCBA screening benchmark. The model sits alongside other retrieval-first screening systems such as CoNCISE, trading docking's physical realism for the ability to ask proteome-scale questions at all.

#Key Features

• Interaction as a dot product: Drug and target projections are normalised and compared by cosine similarity, so a scored pair costs one vector operation.
• Attention pooling over residues: A multi-head, learned aggregation replaces mean pooling, letting the model concentrate on information-rich positions.
• Structure-aware protein input: SaProt's paired residue and Foldseek structure tokens inject fold context into a sequence model.
• Library-scale retrieval: Embeddings live in a ChromaDB index, turning screening into nearest-neighbour search over pre-computed vectors.
• Residue-level interpretability: The pooling attention weights are exposed as maps over the target sequence, showing which residues drove a prediction.
• Affinity as well as classification: Swapping cosine similarity for a plain dot product turns the same architecture into a pKd regressor.

#Technical Details

Molecules are featurised as 2048-bit Morgan fingerprints (radius 2) and proteins as SaProt per-residue embeddings over structure-aware sequences derived from AlphaFold 2 models via Foldseek; both frozen encoders feed modality-specific MLPs projecting into the shared space. The headline SPRINT model adds 3-layer projection MLPs and 16M learned parameters on top of the frozen backbone, against 10M for the single-layer SPRINT-sm variant. Training is fully supervised on binding labels with binary cross-entropy, using the DAVIS, BIOSNAP and BindingDB benchmark splits plus a much larger MERGED corpus from PubChem, BindingDB and ChEMBL with 929,656 positive and 83.7M negative pairs across 11,958 targets. On LIT-PCBA, evaluated after MMseqs2 removes every training protein sharing 90% or more homology with the benchmark set, SPRINT reaches 73.4% AUROC and an enrichment factor of 10.78 at 1%, above SPRINT-Average (6.26) and DrugCLIP (5.51). On the Therapeutic Data Commons BindingDB_Patent leaderboard it reaches a Pearson correlation of 0.588, matching a four-model knowledge-graph ensemble. Retrieval is the point of the design: the top-100 binders for one target across ENAMINE REAL's 6.7 billion compounds take 7ms.

#Applications

The natural uses are the ones docking cannot reach — off-target profiling for a lead series (the whole human proteome against that 6.7 billion-compound library takes 16 minutes), drug repurposing, and screens against organisms rather than single proteins. The authors index 5,043 bacterial, 88 fungal and the human proteome, 4.3M sequences in all, to look for antimicrobials whose targets are absent from human biology. SPRINT also works as a prefilter ahead of physics: in the CACHE 2 setting, using it to pick 100,000 molecules for Gnina docking against SARS-CoV-2 NSP13 yielded higher-scoring, more diverse hits than the first round of an iterative deep-docking pipeline.

#Impact

SPRINT is a preprint, and its evaluation is retrospective apart from the CACHE 2 comparison, where docking scores rather than assays supply the ground truth. Two caveats matter in practice. The released checkpoints — SPRINT and SPRINT-ProtBert — are Google Drive downloads with no stated license, and the structure-aware input the model expects must be produced by running Foldseek over predicted structures; a bare amino acid sequence substitutes mask tokens for the structure half and forfeits the advantage the architecture was built for. The authors are candid, too, that the attention maps largely highlight non-conserved residues identifying the protein rather than mechanistically meaningful contacts. What the work establishes outlasts a leaderboard position: how per-residue protein embeddings are pooled is a first-order design decision in co-embedding models, and structure tokens are a cheap way to carry fold information into a screen that never builds a complex.