Network Medicine Framework for Identifying Drug Repurposing Opportunities for COVID-19

4.3 Network Proximity

$$d(v,t)$$ is the shortest distance. The expected distance between two randomly selected groups of proteins is follwing:

$$dc(V,T) = \frac{1}{|T|} \sum{t \in T} \min_{v \in V} d(v,t)$$

Then we do a normalization, from absolute distance $$dc$$ to the relavtive distance $$Z{d_c}$$

$$Z{d_c} = \frac{d_c - \mu{dc}(V,T)}{\sigma{d_c}(V,T)}$$

4.4 Diffusion State Distance

The diffusion state distance (DSD) uses graph diffusion peoperty to get node pair similarity. $$H_e(A,B)$$ means how many random walks starting from $$A$$ visits $$B$$, and each node is represented by

$$H_e(V_i) = [H_e(V_i, V_1), H_e(V_i, V_2), ..., H_e(V_i, V_n)]$$

Then the similarity between two nodes is L1-norm.

$$DSD(A,B) = | H_e(A) - H_e(B) |$$

Based on this, we try different metrics

1. Pipeline D1: $$I{DSD} = \frac{1}{|V|} \sum{t\in T} \min_{v\in V} DSD(t,v)$$
2. Pipeline D2: $$I{KL} = \sum{t \in T} \min_{v \in V} KL(t,v)$$
3. Pipeline D3: $$I{KL} = \sum{t \in T} \text{median}_{v \in V} KL(t,v)$$
4. Pipeline D4: $$I{JS} = \sum{t \in T} \min_{v\in V} JS(t,v)$$
5. Pipeline D5: $$I{JS} = \sum{t \in T} \text{median}_{v\in V} JS(t,v)$$

4.5 Graph Convolutional Networks

Use GNN based on the Protein-Protein, Drug-Protein, Disease-Protein, Drug-Disease.

COVID-19 treatment recommendation task Take this as a link prediction problem. 2 components:

1. Encoder: From graph G to node embedding
2. Decoder: optimizing embeddings for prediction

Frou decoders from Section 2.2.3

1. Pipeline A1: cosine similartiy between COVID-19 and all the drugs in the decoded embedding space. The decoding is based on N=10 nearest neighbor nodes with minimum distance of D=0.25.
2. Pipeline A2: D=0.8 and N=10.
3. Pipeline A3: N=5, D=0.5.
4. Pipeline A4: N=10, D=1.