ORN

Use OR-gate Network (ORN) to model the impacts of somatic mutations on cancer transcriptomics

This implementation requires numba, numpy, and scipy installed. Alternatively, install Annaconda.

Example usage

After running the ORN.py, use function ORN to execute the algorithm. The arguments for ORN include:

• deg: the binary matrix of differential expression status, each row should be a sample, and each column should be a gene
• mut: the binary event matrix of somatic alterations, each row should be a sample. The number of samples should the same as deg.
• pathway_size: the number of pathways.
• alpha_X: the alpha prior for U (the relationship matrix between mut and path), default 1.0
• beta_X: the beta prior for U (the relationship matrix between mut and path), default 1.0
• alpha_Z: the alpha prior for Z (the relationship matrix between path and deg), default 1.0
• beta_Z: the beta prior for Z (the relationship matrix between path and deg), default 1.0
• leaky: the OR-gate function should be leaky or not whether or not, default True
• max_iter: the maximum number of iteration for optimization, default 100

Note: Need to remove samples with no mutations!

def synthetic_mut2deg(n_path=5, n_sample=1000, n_deg=1000, n_mutation=1000, p_mutation=0.01, p_m2p=(0.02,0.05), p_p2d=(0.05,0.1)):
    mut = np.random.binomial(1, p_mutation, (n_sample, n_mutation))
    U = np.zeros((n_mutation, n_path))
    Z = np.zeros((n_path, n_deg))
    prior_U = np.random.uniform(p_m2p[0], p_m2p[1], n_path)
    prior_Z = np.random.uniform(p_p2d[0], p_p2d[1], n_path)
    for i in range(n_path):
        U[:,i] = np.random.binomial(1, prior_U[i], n_mutation)
        Z[i,:] = np.random.binomial(1, prior_Z[i], n_deg)
    path = mut.dot(U)
    mut1 = np.zeros(mut.shape)
    path[path>1] = 1
    for i in range(path.shape[1]):
        present_sample = np.where(path[:,i]==1)[0] # Index of samples where the ith pathways was perturbed
        for j in present_sample:
            exclusion = np.where(U[:,np.where(path[j]==0)[0]])[0] # The index of driver mutation that should not be chosen
            seed_mut = set(np.where(U[:,i])[0]).difference(set(exclusion))
            mutation = np.random.choice(list(seed_mut)) # Index of the chosen mutations that dysregulate the pathways
            mut1[j, mutation] = 1
            #mut1[:,U.sum(1)==0] = mut[:,U.sum(1)==0] # Keep the mutation status of passenger mutations
    path1 = mut1.dot(U)
    path1[path1>1] = 1

    deg = path.dot(Z)
    deg[deg>1] = 1

    return mut, U, path, Z, deg

def jaccard_quality(X_hat, Z_hat, X_true, Z_true):
    X_and = X_hat.T.dot(X_true)
    Z_and = Z_hat.dot(Z_true.T)
    
    X_or = np.zeros((X_hat.shape[1], X_hat.shape[1]))
    for i in range(X_hat.shape[1]):
        temp = X_hat.T[i] + X_true.T
        temp[temp > 1] = 1
        X_or[i] = temp.sum(1)
    X_qual = X_and / X_or

    Z_or = np.zeros((Z_hat.shape[0], Z_hat.shape[0]))
    for i in range(Z_hat.shape[0]):
        temp = Z_hat[i] + Z_true
        temp[temp > 1] = 1 
        Z_or[i] = temp.sum(1)
    Z_qual = Z_and/Z_or

    return np.max(X_qual * Z_qual, 0)


# Simulation experiment repeated 10 times
res_mut2deg = []
cos_mut2deg = []
for _ in range(10):
    mut, U, path, Z, deg = synthetic_mut2deg(n_path=5)
    
    model1 = mut2deg(np.float64(deg), np.float64(mut), pathway_size=5)
    path_pred1 = reconstruct(np.float64(mut>0.5), model1[0])
    res_mut2deg.append(np.mean(np.abs(reconstruct(path_pred1, model1[1])- deg))) # Caching reconstruction error
    cos_mut2deg.append(jaccard_quality(model1[0][:-1], model1[1], U, Z).mean()) # Caching Jaccard score