def fit_clv(observations, time_points, held_out_rel_abun, held_out_time_points,
denom, method):
print("Estimating cLV parameters using", method)
rel_abun = estimate_relative_abundances(observations)
clv = CompositionalLotkaVolterra(rel_abun, time_points, denom=denom)
if method == "Elastic Net":
clv.train()
predictions = [
clv.predict(o[0], tpts)
for (o, tpts) in zip(held_out_rel_abun, held_out_time_points)
]
A, g, B = clv.get_params()
return A, g, predictions
elif method == "Ridge":
clv.train_ridge()
A, g, B = clv.get_params()
predictions = [
clv.predict(o[0], tpts)
for (o, tpts) in zip(held_out_rel_abun, held_out_time_points)
]
return A, g, predictions
else:
print("bad optimization method for cLV", file=sys.stderr)
return
def fit_clv(observations, time_points, effects, held_out_observations, held_out_time_points, held_out_effects, using_rel_abun=False, ret_params=False, folds=None):
# if observations are concentrations
rel_abun = []
held_out_rel_abun = []
if folds is None:
folds = len(observations)
if not using_rel_abun:
for obs in observations:
rel_abun.append(obs / obs.sum(axis=1,keepdims=True))
for obs in held_out_observations:
held_out_rel_abun.append(obs / obs.sum(axis=1,keepdims=True))
else:
rel_abun = observations
held_out_rel_abun = held_out_observations
clv = CompositionalLotkaVolterra(rel_abun, time_points, effects)
clv.train(folds=folds)
predictions = [clv.predict(o[0],tpts,e) for (o,tpts,e) in zip(held_out_rel_abun, held_out_time_points, held_out_effects)]
return predictions
# estimated previously
r_A = 1
r_g = 4
r_B = 0.5
P = []
Y_pc = []
log_Y = []
for y in Y:
mass = y.sum(axis=1)
p = y / y.sum(axis=1, keepdims=True)
p = (p + 1e-5) / (p + 1e-5).sum(axis=1, keepdims=True)
P.append(p)
Y_pc.append((mass.T * p.T).T)
log_Y.append(np.log(mass.T * p.T).T)
clv = CompositionalLotkaVolterra(P, T, U, pseudo_count=1e-5)
clv.r_A = r_A
clv.r_g = r_g
clv.r_B = r_B
clv.train_ridge()
A_clv, g_clv, B_clv = clv.get_params()
A_glv, g_glv, B_glv = ridge_regression_glv(log_Y, U, T, clv.r_A, clv.r_g,
clv.r_B)
A_rel, g_rel, B_rel = compute_relative_parameters(A_glv, g_glv, B_glv,
clv.denom)
plot_corr(A_rel, g_rel, B_rel, A_clv, g_clv, B_clv,
"plots/stein_correlation.pdf")
U = pkl.load(open("data/stein/U.pkl", "rb"))
T = pkl.load(open("data/stein/T.pkl", "rb"))
col_names = np.array(['und. Enterobacteriaceae', 'Blautia', 'Barnesiella',
'und. uncl. Mollicutes',
'und. Lachnospiraceae', 'Akkermansia',
'C. difficile', 'uncl. Lachnospiraceae', 'Coprobacillus',
'Enterococcus', 'Other'])
P = []
held_out_rel_abun = []
for y in Y:
P.append(y / y.sum(axis=1,keepdims=True))
clv = CompositionalLotkaVolterra(P, T, U)
A = np.loadtxt("pub-results/stein_A")
g = np.loadtxt("pub-results/stein_g")
B = np.loadtxt("pub-results/stein_B")
B = np.expand_dims(B, axis=1)
glv = GeneralizedLotkaVolterra(P, T, U)
A_glv = np.loadtxt("pub-results/stein_A_glv")
g_glv = np.loadtxt("pub-results/stein_g_glv")
B_glv = np.loadtxt("pub-results/stein_B_glv")
B_glv = np.expand_dims(B_glv, axis=1)
ntaxa = Y[0].shape[1]
old_denom = clv.denom
taxon_row_names = col_names[np.array([i for i in range(ntaxa) if i != old_denom])]