From 7a1d3d031aa23fce723ac4f4c8e4bb5d64959447 Mon Sep 17 00:00:00 2001 From: Johannes Ranke Date: Thu, 4 Apr 2019 15:42:23 +0200 Subject: Direct error model fitting works - No IRLS required - Removed optimization algorithms other than Port - Removed the dependency on FME - Fitting the error model 'obs' is much faster for the FOCUS_2006_D dataset and the FOMC_SFO model (1 second versus 3.4 seconds) - Vignettes build slower. Compiled models needs 3 minutes instead of 1.5 - For other vignettes, the trend is less clear. Some fits are faster, even for error_model = "const". FOCUS_Z is faster (34.9 s versus 44.1 s) - Standard errors and confidence intervals are slightly smaller - Removed code for plotting during the fit, as I hardly ever used it - Merged the two cost functions (using transformed and untransformed parameters) into one log-likelihood function --- tests/testthat/test_error_models.R | 178 +++++++++++++++++++++++++++++++++++++ tests/testthat/test_irls.R | 142 ----------------------------- 2 files changed, 178 insertions(+), 142 deletions(-) create mode 100644 tests/testthat/test_error_models.R delete mode 100644 tests/testthat/test_irls.R (limited to 'tests') diff --git a/tests/testthat/test_error_models.R b/tests/testthat/test_error_models.R new file mode 100644 index 00000000..bda8ca7f --- /dev/null +++ b/tests/testthat/test_error_models.R @@ -0,0 +1,178 @@ +# Copyright (C) 2018 Johannes Ranke +# Contact: jranke@uni-bremen.de + +# This file is part of the R package mkin + +# mkin is free software: you can redistribute it and/or modify it under the +# terms of the GNU General Public License as published by the Free Software +# Foundation, either version 3 of the License, or (at your option) any later +# version. + +# This program is distributed in the hope that it will be useful, but WITHOUT +# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS +# FOR A PARTICULAR PURPOSE. See the GNU General Public License for more +# details. + +# You should have received a copy of the GNU General Public License along with +# this program. If not, see + +context("Error model fitting") + +m_synth_SFO_lin <- mkinmod(parent = mkinsub("SFO", "M1"), + M1 = mkinsub("SFO", "M2"), + M2 = mkinsub("SFO"), + use_of_ff = "max", quiet = TRUE) + +m_synth_DFOP_par <- mkinmod(parent = mkinsub("DFOP", c("M1", "M2")), + M1 = mkinsub("SFO"), + M2 = mkinsub("SFO"), + use_of_ff = "max", quiet = TRUE) + +SFO_lin_a <- synthetic_data_for_UBA_2014[[1]]$data + +DFOP_par_c <- synthetic_data_for_UBA_2014[[12]]$data + +test_that("Error model 'const' works", { + skip_on_cran() + fit_const_1 <- mkinfit(m_synth_SFO_lin, SFO_lin_a, error_model = "const", quiet = TRUE) + bpar_1 <- summary(fit_const_1)$bpar[, c("Estimate", "Lower", "Upper")] + # The reference used here is mkin 0.9.48.1 + bpar_1_mkin_0.9 <- read.table(text = +"parent_0 102.0000 98.6000 106.0000 +k_parent 0.7390 0.6770 0.8070 +k_M1 0.2990 0.2560 0.3490 +k_M2 0.0202 0.0176 0.0233 +f_parent_to_M1 0.7690 0.6640 0.8480 +f_M1_to_M2 0.7230 0.6030 0.8180", +col.names = c("parameter", "estimate", "lower", "upper")) + + expect_equivalent(signif(bpar_1[1:6, "Estimate"], 3), bpar_1_mkin_0.9$estimate) + # Relative difference of lower bound of confidence is < 0.02 + rel_diff <- function(v1, v2) { + (v1 - v2)/v2 + } + expect_equivalent(rel_diff(bpar_1[1:6, "Lower"], + bpar_1_mkin_0.9$lower), + rep(0, 6), tolerance = 0.02) +}) + +test_that("Error model 'obs' works", { + skip_on_cran() + fit_obs_1 <- mkinfit(m_synth_SFO_lin, SFO_lin_a, error_model = "obs", quiet = TRUE) + parms_2 <- round(fit_obs_1$bparms.optim, c(1, 4, 4, 4, 4, 4)) + expect_equivalent(parms_2, c(102.1, 0.7389, 0.2982, 0.0203, 0.7677, 0.7246)) +}) + +test_that("Error model 'tc' works", { + skip_on_cran() + fit_tc_1 <- mkinfit(m_synth_SFO_lin, SFO_lin_a, error_model = "tc", quiet = TRUE) + parms_3 <- round(fit_tc_1$bparms.optim, c(1, 4, 4, 4, 4, 4)) + expect_equivalent(parms_3, c(102.1, 0.7393, 0.2992, 0.0202, 0.7687, 0.7229)) +}) + +test_that("Error model 'obs_tc' works", { + skip_on_cran() + fit_obs_tc_1 <- expect_warning(mkinfit(m_synth_SFO_lin, SFO_lin_a, error_model = "obs_tc", quiet = TRUE), "NaN") + # Here the error model is overparameterised + expect_warning(summary(fit_obs_tc_1), "singular system") +}) + +test_that("Reweighting method 'tc' produces reasonable variance estimates", { + + # I need to make the tc method more robust against that + # skip_on_cran() + + # Check if we can approximately obtain the parameters and the error model + # components that were used in the data generation + + # Parent only + DFOP <- mkinmod(parent = mkinsub("DFOP")) + sampling_times = c(0, 1, 3, 7, 14, 28, 60, 90, 120) + parms_DFOP <- c(k1 = 0.2, k2 = 0.02, g = 0.5) + parms_DFOP_optim <- c(parent_0 = 100, parms_DFOP) + + d_DFOP <- mkinpredict(DFOP, + parms_DFOP, c(parent = 100), + sampling_times) + d_2_10 <- add_err(d_DFOP, + sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), + n = 10, reps = 2, digits = 5, LOD = -Inf, seed = 123456) + d_100_1 <- add_err(d_DFOP, + sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), + n = 1, reps = 100, digits = 5, LOD = -Inf, seed = 123456) + + # Unweighted fits + f_2_10 <- mmkin("DFOP", d_2_10, error_model = "const", quiet = TRUE, + cores = if (Sys.getenv("TRAVIS") != "") 1 else 15) + parms_2_10 <- apply(sapply(f_2_10, function(x) x$bparms.optim), 1, mean) + parm_errors_2_10 <- (parms_2_10 - parms_DFOP_optim) / parms_DFOP_optim + expect_true(all(abs(parm_errors_2_10) < 0.12)) + + f_2_10_tc <- mmkin("DFOP", d_2_10, error_model = "tc", quiet = TRUE, + cores = if (Sys.getenv("TRAVIS") != "") 1 else 15) + parms_2_10_tc <- apply(sapply(f_2_10_tc, function(x) x$bparms.optim), 1, mean) + parm_errors_2_10_tc <- (parms_2_10_tc - parms_DFOP_optim) / parms_DFOP_optim + expect_true(all(abs(parm_errors_2_10_tc) < 0.05)) + + tcf_2_10_tc <- apply(sapply(f_2_10_tc, function(x) x$errparms), 1, mean, na.rm = TRUE) + + tcf_2_10_error_model_errors <- (tcf_2_10_tc - c(0.5, 0.07)) / c(0.5, 0.07) + expect_true(all(abs(tcf_2_10_error_model_errors) < 0.2)) + + # When we have 100 replicates in the synthetic data, we can roundtrip + # the parameters with < 2% precision + f_tc_100_1 <- mkinfit(DFOP, d_100_1[[1]], error_model = "tc", quiet = TRUE) + parm_errors_100_1 <- (f_tc_100_1$bparms.optim - parms_DFOP_optim) / parms_DFOP_optim + expect_true(all(abs(parm_errors_100_1) < 0.02)) + + tcf_100_1_error_model_errors <- (f_tc_100_1$errparms - c(0.5, 0.07)) / + c(0.5, 0.07) + # When maximising the likelihood directly (not using IRLS), we get + # a precision of < 2% for the error model componentes as well + expect_true(all(abs(tcf_100_1_error_model_errors) < 0.02)) + + # Parent and two metabolites + m_synth_DFOP_lin <- mkinmod(parent = list(type = "DFOP", to = "M1"), + M1 = list(type = "SFO", to = "M2"), + M2 = list(type = "SFO"), use_of_ff = "max", + quiet = TRUE) + sampling_times = c(0, 1, 3, 7, 14, 28, 60, 90, 120) + parms_DFOP_lin <- c(k1 = 0.2, k2 = 0.02, g = 0.5, + f_parent_to_M1 = 0.5, k_M1 = 0.3, + f_M1_to_M2 = 0.7, k_M2 = 0.02) + d_synth_DFOP_lin <- mkinpredict(m_synth_DFOP_lin, + parms_DFOP_lin, + c(parent = 100, M1 = 0, M2 = 0), + sampling_times) + parms_DFOP_lin_optim = c(parent_0 = 100, parms_DFOP_lin) + + d_met_2_15 <- add_err(d_synth_DFOP_lin, + sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), + n = 15, reps = 100, digits = 5, LOD = -Inf, seed = 123456) + + # For a single fit, we get a relative error of less than 10% in the error + # model components + f_met_2_tc_e4 <- mkinfit(m_synth_DFOP_lin, d_met_2_15[[1]], quiet = TRUE, + error_model = "tc") + parm_errors_met_2_tc_e4 <- (f_met_2_tc_e4$errparms - c(0.5, 0.07)) / c(0.5, 0.07) + expect_true(all(abs(parm_errors_met_2_tc_e4) < 0.1)) + + # Doing more takes a lot of computing power + skip_on_travis() + f_met_2_15_tc_e4 <- mmkin(list(m_synth_DFOP_lin), d_met_2_15, quiet = TRUE, + error_model = "tc", cores = 15) + + parms_met_2_15_tc_e4 <- apply(sapply(f_met_2_15_tc_e4, function(x) x$bparms.optim), 1, mean) + parm_errors_met_2_15_tc_e4 <- (parms_met_2_15_tc_e4[names(parms_DFOP_lin_optim)] - + parms_DFOP_lin_optim) / parms_DFOP_lin_optim + expect_true(all(abs(parm_errors_met_2_15_tc_e4) < 0.01)) + + tcf_met_2_15_tc <- apply(sapply(f_met_2_15_tc_e4, function(x) x$errparms), 1, mean, na.rm = TRUE) + + tcf_met_2_15_tc_error_model_errors <- (tcf_met_2_15_tc - c(0.5, 0.07)) / + c(0.5, 0.07) + + # Here we get a precision < 15% for retrieving the original error model components + # from 15 datasets + expect_true(all(abs(tcf_met_2_15_tc_error_model_errors) < 0.15)) +}) diff --git a/tests/testthat/test_irls.R b/tests/testthat/test_irls.R deleted file mode 100644 index f61f793d..00000000 --- a/tests/testthat/test_irls.R +++ /dev/null @@ -1,142 +0,0 @@ -# Copyright (C) 2018 Johannes Ranke -# Contact: jranke@uni-bremen.de - -# This file is part of the R package mkin - -# mkin is free software: you can redistribute it and/or modify it under the -# terms of the GNU General Public License as published by the Free Software -# Foundation, either version 3 of the License, or (at your option) any later -# version. - -# This program is distributed in the hope that it will be useful, but WITHOUT -# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS -# FOR A PARTICULAR PURPOSE. See the GNU General Public License for more -# details. - -# You should have received a copy of the GNU General Public License along with -# this program. If not, see - -context("Iteratively reweighted least squares (IRLS) fitting") - - -m_synth_SFO_lin <- mkinmod(parent = mkinsub("SFO", "M1"), - M1 = mkinsub("SFO", "M2"), - M2 = mkinsub("SFO"), - use_of_ff = "max", quiet = TRUE) - -m_synth_DFOP_par <- mkinmod(parent = mkinsub("DFOP", c("M1", "M2")), - M1 = mkinsub("SFO"), - M2 = mkinsub("SFO"), - use_of_ff = "max", quiet = TRUE) - -SFO_lin_a <- synthetic_data_for_UBA_2014[[1]]$data - -DFOP_par_c <- synthetic_data_for_UBA_2014[[12]]$data - -test_that("Reweighting method 'obs' works", { - skip_on_cran() - fit_irls_1 <- mkinfit(m_synth_SFO_lin, SFO_lin_a, reweight.method = "obs", quiet = TRUE) - parms_1 <- round(fit_irls_1$bparms.optim, c(1, 4, 4, 4, 4, 4)) - expect_equivalent(parms_1, c(102.1, 0.7389, 0.2982, 0.0203, 0.7677, 0.7246)) -}) - -test_that("Reweighting method 'tc' works", { - fit_irls_2 <- mkinfit(m_synth_SFO_lin, SFO_lin_a, reweight.method = "tc", quiet = TRUE) - parms_2 <- round(fit_irls_2$bparms.optim, c(1, 4, 4, 4, 4, 4)) - expect_equivalent(parms_2, c(102.1, 0.7393, 0.2992, 0.0202, 0.7687, 0.7229)) - - skip("Too much trouble with datasets that are randomly generated") - # I need to make the tc method more robust against that - # skip_on_cran() - - # Check if we can approximately obtain the parameters and the error model - # components that were used in the data generation - - # Parent only - DFOP <- mkinmod(parent = mkinsub("DFOP")) - sampling_times = c(0, 1, 3, 7, 14, 28, 60, 90, 120) - parms_DFOP <- c(k1 = 0.2, k2 = 0.02, g = 0.5) - parms_DFOP_optim <- c(parent_0 = 100, parms_DFOP) - d_DFOP <- mkinpredict(DFOP, - parms_DFOP, c(parent = 100), - sampling_times) - d_2_10 <- add_err(d_DFOP, - sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), - n = 10, reps = 2, digits = 5, LOD = -Inf) - d_100_1 <- add_err(d_DFOP, - sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), - n = 1, reps = 100, digits = 5, LOD = -Inf) - - f_2_10 <- mmkin("DFOP", d_2_10, quiet = TRUE, - cores = if (Sys.getenv("TRAVIS") != "") 1 else 15) - parms_2_10 <- apply(sapply(f_2_10, function(x) x$bparms.optim), 1, mean) - parm_errors_2_10 <- (parms_2_10 - parms_DFOP_optim) / parms_DFOP_optim - expect_true(all(abs(parm_errors_2_10) < 0.45)) - - f_2_10_tc <- mmkin("DFOP", d_2_10, reweight.method = "tc", quiet = TRUE, - cores = if (Sys.getenv("TRAVIS") != "") 1 else 15) - parms_2_10_tc <- apply(sapply(f_2_10_tc, function(x) x$bparms.optim), 1, mean) - parm_errors_2_10_tc <- (parms_2_10_tc - parms_DFOP_optim) / parms_DFOP_optim - expect_true(all(abs(parm_errors_2_10_tc) < 0.25)) - - tcf_2_10_tc <- apply(sapply(f_2_10_tc, function(x) x$tc_fitted), 1, mean, na.rm = TRUE) - - tcf_2_10_error_model_errors <- (tcf_2_10_tc - c(0.5, 0.07)) / c(0.5, 0.07) - expect_true(all(abs(tcf_2_10_error_model_errors) < 0.5)) - - f_tc_100_1 <- suppressWarnings(mkinfit(DFOP, d_100_1[[1]], reweight.method = "tc", quiet = TRUE)) - parm_errors_100_1 <- (f_tc_100_1$bparms.optim - parms_DFOP_optim) / parms_DFOP_optim - expect_true(all(abs(parm_errors_100_1) < 0.1)) - - tcf_100_1_error_model_errors <- (f_tc_100_1$tc_fitted - c(0.5, 0.07)) / - c(0.5, 0.07) - # Even with 100 (or even 1000, not shown) replicates at each observation time - # we only get a precision of 15% to 30% for the error model components - expect_true(all(abs(tcf_100_1_error_model_errors) < 0.3)) - - # Parent and two metabolites - m_synth_DFOP_lin <- mkinmod(parent = list(type = "DFOP", to = "M1"), - M1 = list(type = "SFO", to = "M2"), - M2 = list(type = "SFO"), use_of_ff = "max", - quiet = TRUE) - sampling_times = c(0, 1, 3, 7, 14, 28, 60, 90, 120) - parms_DFOP_lin <- c(k1 = 0.2, k2 = 0.02, g = 0.5, - f_parent_to_M1 = 0.5, k_M1 = 0.3, - f_M1_to_M2 = 0.7, k_M2 = 0.02) - d_synth_DFOP_lin <- mkinpredict(m_synth_DFOP_lin, - parms_DFOP_lin, - c(parent = 100, M1 = 0, M2 = 0), - sampling_times) - parms_DFOP_lin_optim = c(parent_0 = 100, parms_DFOP_lin) - - d_met_2_15 <- add_err(d_synth_DFOP_lin, - sdfunc = function(x) sigma_twocomp(x, 0.5, 0.07), - n = 15, reps = 1000, digits = 5, LOD = -Inf) - - # For a single fit, we get a relative error of less than 30% in the error - # model components - f_met_2_tc_e4 <- mkinfit(m_synth_DFOP_lin, d_met_2_15[[1]], quiet = TRUE, - reweight.method = "tc", reweight.tol = 1e-4) - parm_errors_met_2_tc_e4 <- (f_met_2_tc_e4$tc_fitted - c(0.5, 0.07)) / c(0.5, 0.07) - expect_true(all(abs(parm_errors_met_2_tc_e4) < 0.3)) - - # Doing more takes a lot of computing power - skip_on_travis() - f_met_2_15_tc_e4 <- mmkin(list(m_synth_DFOP_lin), d_met_2_15, quiet = TRUE, - reweight.method = "tc", reweight.tol = 1e-4, - cores = 14) - - parms_met_2_15_tc_e4 <- apply(sapply(f_met_2_15_tc_e4, function(x) x$bparms.optim), 1, mean) - parm_errors_met_2_15_tc_e4 <- (parms_met_2_15_tc_e4[names(parms_DFOP_lin_optim)] - - parms_DFOP_lin_optim) / parms_DFOP_lin_optim - expect_true(all(abs(parm_errors_met_2_15_tc_e4) < 0.01)) - - tcf_met_2_15_tc <- apply(sapply(f_met_2_15_tc_e4, function(x) x$tc_fitted), 1, mean, na.rm = TRUE) - - tcf_met_2_15_tc_error_model_errors <- (tcf_met_2_15_tc - c(0.5, 0.07)) / - c(0.5, 0.07) - - # Here we only get a precision < 30% for retrieving the original error model components - # from 15 datasets - expect_true(all(abs(tcf_met_2_15_tc_error_model_errors) < 0.3)) -}) -- cgit v1.2.1