Investigating ADR mechanisms with Explainable AI: a feasibility study with knowledge graph mining

Table 5 Numbers of drug features extractable from the knowledge graph, with different levels of filtering

		DILI	SCAR
	Neighbors	5,488,531	5,488,510
No filtering (deg = -1)	k	19	20
	t	21	21
	Neighbors	2,419,957	2,419,920
Filtering level 1 (deg = 500)	k	23	23
	t	21	21
Filtering level 2 (deg = 500, s_min = 5, k = 3, t = 3)	Neighbors	175,652	179,694
	Paths & path patterns	20,145,635	29,011,996
Filtering level 3 (deg = 500, s_min = 5, k = 3, t 3, m_DILI = pgm and m_SCAR = pg)
	Neighbors	4069	1594
	Paths & path patterns	102,674	86,753

The first line corresponds to the full neighborhood of drugs from DILI and SCAR expert classifications. \(\textit{deg} = -1\) means that all nodes are considered, regardless of their degree, whereas \(\textit{deg}=500\) in Filtering level 1 means that nodes with a degree \(>\,\textit{deg}\) are filtered out. In the two first lines (No filtering and Filtering level 1), k and t are unconstrained, so reported values are maximum k and t observed in the graph. Paths and paths pattern are computed only when deg and \(s_\text {min}\) (minimum support) are set, to avoid combinatorial explosion. Filtering level 2 and 3 share the following additional parameters: \(undirected={\texttt {false}}\), \(s_\text {max}=+\infty\). In Filtering level 3, m is set for additional filtering. Distinct values for m chosen respectively for DILI and SCAR are those associated with the best performances, e.g., \(m_{DILI}={\texttt {pgm}}\) and \(m_{SCAR}={\texttt {pg}}\)

ISSN: 1472-6947