Figure 3.

Schematic illustration of the ‘Protein-Ligand Integration Module’ in ArkDTA which features attention regularization based on NCIs between protein and ligand. The query and key projections of m residues $\mathrm{R}$ and n chemical substructures $\mathrm{S}$ appended with $\overrightarrow{p}$ are $\left\{q_1,q_2,\dots ,q_m\right\}$ and $\left\{k_1,k_2,\dots ,k_n,k_p\right\}$, respectively. The auxiliary loss objective ${\mathcal{L}}_2$ enforces ArkDTA to focus most of the attention toward $k_p$ when given a query residue without NCI. On the other hand, the ${\mathcal{L}}_2$ encourages ArkDTA to only distribute its attention on actual chemical substructures from $k_1$ to $k_n$ in an unsupervised fashion.