Reading the results

Predicted engagement

Each compound in a screen gets a predicted engagement between 0 and 1: the model's calibrated estimate of how likely the compound is to engage the target. A screen ranks every compound in the library from highest to lowest predicted engagement.

Use predicted engagement to choose which compounds to test first. It is a prediction for prioritizing experiments, not a measured affinity.

Predicted engagement is computed the same way for every target and library, and doesn't depend on a protein's length, so values from screens run with the same model version can be compared across screens and targets. Use it first to rank compounds, and don't read a fixed cutoff such as 0.5 as "engages" versus "doesn't engage".

In downloads and in the API, the column is engagement. The API also returns score, the raw similarity at the highest-scoring residue, between -1 and 1. It is not calibrated, so rank by engagement, not by score. The two need not agree on order.

Highest-scoring residue

The residue of the target the compound scored best against, written as its amino acid (one-letter code) and its position in the protein's sequence: C145 is the cysteine at position 145. In the API these are predicted_residue_letter and predicted_residue (the position, counted from 1).

Many compounds can point to the same residue on one target, often a known active site. Compare compounds by predicted engagement rather than by their residue.

Residue predictions

On a compound's page, predicted engagement is shown residue by residue along the target's sequence. A residue marked not scored has no prediction: for some long proteins only part of the sequence is processed, and this is indicated in the target list on New screen ("Partial sequence: 1200 / 2000 residues scored"). The API returns the same values from GET /api/v1/screens/{screen_id}/compounds/{compound_id}?include_residues=true as residue_engagement: one number per residue, or null where not scored.

Reference proteome predictions

The reference proteome is the set of human proteins Ptarmigan-1 has predictions for, the same proteins targets are chosen from. A compound's page lists the five proteins in the reference proteome that the compound is most likely to engage, ranked by predicted engagement. The target you screened is always considered, and appears when it ranks among the five.

Use the list to see which other proteins a promising compound may engage besides your target. These predictions are in the web app only; the API and MCP don't return them.