SPRINT

What is SPRINT?

SPRINT is a learned protein-ligand screening method that places proteins and compounds in a shared embedding space. It ranks a compound library for one protein target by cosine similarity. Unlike AutoDock Vina or GNINA, it does not generate 3D binding poses or calculate a docking energy.

This implementation runs the published SPRINT ultrafast-embed and ultrafast-topk workflow. It creates an embedding for the target, creates embeddings for the submitted compound library, and returns the compounds with the highest native CosineSimi values.

SPRINT is best used to prioritize compounds for follow-up. Its cosine similarity is a model score, not a calibrated binding probability, binding affinity, or experimental measurement.

How does SPRINT work?

SPRINT maps protein and ligand features into a shared vector space. The published Top K command normalizes those embeddings through cosine similarity and ranks larger values first.

Protein embedding

The selected checkpoint uses SaProt features. You can submit either:

• A plain sequence containing the 20 standard amino acid letters. SPRINT inserts masked structure tokens for this input.
• A paired SaProt sequence that alternates an uppercase amino acid with a lowercase 3Di token or # mask.

ProteinIQ does not predict a structure or generate 3Di tokens for a plain sequence. If you want the model to use known structural tokens, prepare and submit the paired SaProt representation yourself.

Compound embedding

The checkpoint uses 1024-bit Morgan fingerprints with radius 3. SPRINT projects these fingerprints into the same learned space as the protein representation.

Ranking

For each compound, SPRINT reports:

Higher values rank first. ProteinIQ does not transform this value into a probability.

How to use SPRINT online

Inputs

Text input example:

aspirin	CC(=O)Oc1ccccc1C(=O)O
ibuprofen	CC(C)Cc1ccc(cc1)C(C)C(=O)O
caffeine	Cn1cnc2c1c(=O)n(C)c(=O)n2C

CSV input example:

id,SMILES,collection
aspirin,CC(=O)Oc1ccccc1C(=O)O,reference
ethanol,CCO,reference

Additional CSV or TSV columns, including their whitespace, are preserved in the ranked output. UTF-8 BOM files are accepted. An existing id column takes precedence over name or compound_id.

Settings

Results

The results table preserves the source library columns and adds the native similarity. ProteinIQ adds rank as a display field. If the library already contains rank, the display field uses display_rank, with additional display_ prefixes only when needed to avoid another name collision.

Downloadable files include:

• topk_mol_data_<query_id>.csv, the native ranked compound table
• topk_mol_embeddings_<query_id>.npy, embeddings for the returned compounds
• sprint_target_embeddings.npy, the target embedding
• sprint_library_embeddings.npy, embeddings for the full submitted library
• sprint_run.log, command output for reproducibility and troubleshooting
• sprint_provenance.json, pinned source, model, and runtime identifiers

If a scientific command fails, sprint_run.log retains the failed command's output and any completed stages. The job remains failed; a diagnostic log does not indicate a completed screening result.

Interpreting scores

Use CosineSimi for relative ranking within a run. A larger value means the model placed that compound closer to the target in its learned embedding space. The score has no universal probability threshold and should not be interpreted as a dissociation constant, inhibition constant, or docking energy.

Screening workflow

One practical sequence is:

1. Filter compounds for drug-likeness using Lipinski's Rule of 5 or Molecular Descriptors
2. Run SPRINT to prioritize the library by cosine similarity
3. Assess top hits for toxicity and ADMET properties with ADMET-AI or eToxPred
4. Validate promising candidates with structure-based docking (DiffDock, AutoDock Vina, or GNINA)

The published Morgan featurizer substitutes a zero fingerprint when RDKit cannot parse a SMILES value. Such rows may still receive a model score. Check the run log and remove invalid structures before interpreting or advancing hits.

Limitations

• SPRINT ranks learned interaction similarity. It does not return calibrated probabilities, binding affinities, poses, residue attention maps, or binding sites.
• Plain protein sequences use masked structure tokens, so they do not contain the same structural information as prepared amino-acid/3Di pairs.
• Performance can decrease for protein families or chemical scaffolds unlike the training data.
• A single embedding cannot represent every conformational state, binding site, protonation state, tautomer, or stereochemical uncertainty.
• Follow up high-ranking compounds with orthogonal computational methods and experimental assays.