GEARS model
| 45 | |
| 46 | |
| 47 | class GEARS_Model(torch.nn.Module): |
| 48 | """ |
| 49 | GEARS model |
| 50 | |
| 51 | """ |
| 52 | |
| 53 | def __init__(self, args): |
| 54 | """ |
| 55 | :param args: arguments dictionary |
| 56 | """ |
| 57 | |
| 58 | super(GEARS_Model, self).__init__() |
| 59 | self.args = args |
| 60 | self.num_genes = args['num_genes'] |
| 61 | self.num_perts = args['num_perts'] |
| 62 | hidden_size = args['hidden_size'] |
| 63 | self.uncertainty = args['uncertainty'] |
| 64 | self.num_layers = args['num_go_gnn_layers'] |
| 65 | self.indv_out_hidden_size = args['decoder_hidden_size'] |
| 66 | self.num_layers_gene_pos = args['num_gene_gnn_layers'] |
| 67 | self.no_perturb = args['no_perturb'] |
| 68 | self.pert_emb_lambda = 0.2 |
| 69 | |
| 70 | # perturbation positional embedding added only to the perturbed genes |
| 71 | self.pert_w = nn.Linear(1, hidden_size) |
| 72 | |
| 73 | # gene/globel perturbation embedding dictionary lookup |
| 74 | self.gene_emb = nn.Embedding(self.num_genes, hidden_size, max_norm=True) |
| 75 | self.pert_emb = nn.Embedding(self.num_perts, hidden_size, max_norm=True) |
| 76 | |
| 77 | # transformation layer |
| 78 | self.emb_trans = nn.ReLU() |
| 79 | self.pert_base_trans = nn.ReLU() |
| 80 | self.transform = nn.ReLU() |
| 81 | self.emb_trans_v2 = MLP([hidden_size, hidden_size, hidden_size], last_layer_act='ReLU') |
| 82 | self.pert_fuse = MLP([hidden_size, hidden_size, hidden_size], last_layer_act='ReLU') |
| 83 | |
| 84 | # gene co-expression GNN |
| 85 | self.G_coexpress = args['G_coexpress'].to(args['device']) |
| 86 | self.G_coexpress_weight = args['G_coexpress_weight'].to(args['device']) |
| 87 | |
| 88 | self.emb_pos = nn.Embedding(self.num_genes, hidden_size, max_norm=True) |
| 89 | self.layers_emb_pos = torch.nn.ModuleList() |
| 90 | for i in range(1, self.num_layers_gene_pos + 1): |
| 91 | self.layers_emb_pos.append(SGConv(hidden_size, hidden_size, 1)) |
| 92 | |
| 93 | ### perturbation gene ontology GNN |
| 94 | self.G_sim = args['G_go'].to(args['device']) |
| 95 | self.G_sim_weight = args['G_go_weight'].to(args['device']) |
| 96 | |
| 97 | self.sim_layers = torch.nn.ModuleList() |
| 98 | for i in range(1, self.num_layers + 1): |
| 99 | self.sim_layers.append(SGConv(hidden_size, hidden_size, 1)) |
| 100 | |
| 101 | # decoder shared MLP |
| 102 | self.recovery_w = MLP([hidden_size, hidden_size*2, hidden_size], last_layer_act='linear') |
| 103 | |
| 104 | # gene specific decoder |