Interaction Prediction

Interaction prediction estimates whether two or more entities interact, how strongly they interact, or what structured relationship they form. In Molecular Modeling this often means molecule-target, protein-ligand, protein-protein, residue-residue, or molecule-molecule interactions.

A pairwise interaction model can be written as:

$$\hat{y}=f_{\theta }(r_a,r_b,c)$$

where $r_a$ and $r_b$ are entity representations and $c$ is context such as assay, pocket, species, source, or geometric frame.

For a protein-ligand complex:

$$\hat{y}=f_{\theta }(r_P,r_L,r_{PL})$$

where $r_P$ is protein or pocket representation, $r_L$ is ligand representation, and $r_{PL}$ is an optional pair or interaction representation.

Output Variants

• Binary interaction: interact vs. not interact.
• Affinity or activity: scalar label such as binding affinity or assay response.
• Contact map: pairwise interaction matrix.
• Pose or geometry: structured coordinates or relative transform.
• Ranking score: prioritize candidates for screening.

For pairwise contacts:

$${\hat{C}}_{ij}=\sigma (s_{\theta }(h_i^{(a)},h_j^{(b)},e_{ij}))$$

where $h_i^{(a)}$ and $h_j^{(b)}$ are entity features, $e_{ij}$ is pair evidence such as distance or edge features, and ${\hat{C}}_{ij}$ is an interaction probability.

Evaluation Risks

• Random row splits can leak the same entity pair, close ligand scaffold, homologous protein, or shared assay context.
• Negative labels may mean measured non-interaction, unobserved interaction, or untested pair.
• A model can learn assay/source identity rather than interaction mechanism.
• Pose-based features may not be available at deployment time.
• Binding affinity, pose quality, and interaction existence are different targets.

Checks

• What are the interacting entities?
• Is the task binary prediction, affinity prediction, pose prediction, contact prediction, or ranking?
• Is the label measured, curated, inferred, censored, weak, or missing?
• What pair, scaffold, family, assay, source, or temporal split is required?
• Does the representation include only information available at prediction time?
• Are interaction evidence, label semantics, and metric aligned?

Related

• Property prediction
• Structured prediction
• Protein-ligand interaction
• Binding affinity
• Protein-ligand split
• Protein-ligand complex
• Target-assay-label contract
• Negative set
• Protein family split
• Scaffold split