NEWS.md

rxode2 (development version)

New features

• library(rxode2) is faster. .onLoad() no longer calls requireNamespace() on the suggested packages (pillar, tibble, arrow, dplyr, nlme, units, digest) before registering their S3 methods. The registration helper already installs an onLoad hook when the other namespace is not loaded yet, so the eager loads only added startup cost; the methods are still registered at the same point in time from the user's perspective.

Bug fixes

Solving

• Fixed an out-of-bounds thread index that could segfault a solve. The internal thread id used to slice the per-thread solving buffers was not bounded by the number of threads those buffers were allocated for (op$cores). A larger id read past the end of gInfusionRate[] -- an array of pointers -- and the resulting garbage pointer crashed iniSubject(); the flat per-thread arrays were silently overrun in the same way. The id is now clamped to the last valid slot, matching what rx_get_thread() already did.

• RcppParallel is now a runtime import (added to Imports with an importFrom), so its shared library is loaded into the process before rxode2's. rxode2 links against RcppParallel (-lRcppParallel); with RcppParallel only in LinkingTo its DLL was not guaranteed to be loaded first, so on Windows library(rxode2) could fail with LoadLibrary failure: The specified module could not be found. This surfaced with RcppParallel 6.0.0, which statically links TBB and no longer ships the tbb.dll stub that previously happened to pull the library in.

• The symbolic derivatives of the relational operators (>, <, >=, <=) are now centered on the discontinuity a == b: the atanh(2*tol - 1) shift that placed the smoothed nascent-delta bump at a - b ~ +/-0.46 was removed. Since the forward pass evaluates relationals as hard booleans, the shifted bump gave sensitivity/exact-gradient consumers (e.g. FOCEI's analytic gradient paths) a spurious derivative in a band next to the threshold; the centered rule makes the derivative consistent with the forward value. This also makes the first derivatives of abs(), min(), and max() exact away from the boundary (#1159).

rxode2 5.1.5

New features

• rxControl(sigdig=) now derives the ODE solver tolerances with one solver-independent formula -- the same for stiff, non-stiff and auto-switching solvers. The rtol exponent IS sigdig and atol sits three orders below it: rtol = 10^(-sigdig) and atol = 10^(-sigdig-3). The sensitivity tolerances match the main solve (rtolSens = rtol, atolSens = atol), since gradients and covariances are built from them, and the steady-state tolerances run one order looser (ssRtol = ssRtolSens = 10*rtol, ssAtol = ssAtolSens = 10*atol). This matches how nlmixr2est derives solver tolerances from its optimization sigdig, so the same sigdig means the same thing whether it is used for estimation or for a plain rxSolve().

  sigdig remains NULL by default and continues to have no effect unless you pass it, so solves that do not name sigdig are unchanged.

  Two notes for callers who do pass it. First, the mapping is keyed to sigdig as a request for that many significant digits, which for small sigdig is looser than what the previous symmetric atol = rtol = 0.5*10^(-sigdig-2) gave: at sigdig = 4, rtol moves from 5e-7 to 1e-4, which is also looser than the 1e-6 default rtol (atol moves the other way, from 5e-7 to 1e-7). If you were using sigdig to tighten a solve, raise it or set atol/rtol directly. Second, each tolerance is resolved independently and only when you did not supply it, so an explicit atol/rtol overrides the main solve but does not propagate to the sensitivity or steady-state tolerances -- set those directly if they should change too.

• The SUNDIALS public headers are now vendored into the package (src/sundials_inc/) alongside the already-vendored SUNDIALS C sources, and the LinkingTo: sundialr dependency has been dropped. This guarantees the vendored sources always compile against headers from the same SUNDIALS release, instead of silently drifting when sundialr updates its bundled SUNDIALS (#1155). The vendored include is injected via PKG_CPPFLAGS so it precedes the LinkingTo include flags; otherwise the older SUNDIALS copy bundled inside StanHeaders would shadow it.

• Removed the dependency on qs2 (and hence stringfish). rxSerialize() now supports the base R types only ("xz", "bzip2", "base"); rxDeserialize() still reads qs2/qdata-serialized data and base91-encoded strings when the qs2 package is installed. Test data was converted from .qs2 to .rds.

Bug fixes

• The first and second derivatives of the Yeo-Johnson transform (rxTBSd() and rxTBSd2()) had the wrong sign for negative values when lambda was exactly 2. There yj(x) = -log(1 - x), so the derivatives are 1/(1 - x) and 1/(1 - x)^2, both positive; the special case returned them negated, which also contradicted the general formula's limit and made the derivative discontinuous in lambda at 2. Since Yeo-Johnson is monotone increasing, its first derivative must be positive everywhere.

• Review of the fix above found further errors in the same transform family (all pre-existing, none introduced by that fix):

  • rxTBSd2() returned a wrong second derivative for the logit transform (an algebra error in the closed form) and for the composed logit + yeoJohnson / probit + yeoJohnson transforms, where the chain rule used the first Yeo-Johnson derivative in place of the second and dropped the inner-transform curvature term.
  • rxTBSi() did not invert the composed logit + yeoJohnson / probit + yeoJohnson transforms: it applied the forward Yeo-Johnson transform (or skipped it entirely) instead of the inverse, so rxTBSi(rxTBS(x)) did not return x for lambda != 1. This affected simulation back-transforms of those error models.
  • The lambda gradient of the transform log-Jacobian (powerDL, used by estimation routines) was wrong on the negative Yeo-Johnson branch (-log1p(x) instead of -log1p(-x), NaN for x < -1), returned a spurious 0 at exactly lambda == 1 for boxCox/yeoJohnson, was missing the probit + yeoJohnson case (returned NA), and returned a spurious log(x) (instead of 0) for the lambda-free lnorm transform. The log-Jacobian itself (powerL) clamped the wrong term in its logit guard, giving an unprotected log(0) at the upper bound.
  • For boxCox/lnorm, rxTBSd() and rxTBSd2() returned the clamp constant sqrt(.Machine$double.eps) itself for x at or below the clamp instead of clamping x and evaluating the derivative formula, making the derivatives discontinuous (and ~15 orders of magnitude too small) at the boundary. The clamp now feeds the usual formula, matching how every other transform in the family handles the guard.

• On Windows with RcppParallel >= 6.0.0 (which statically links TBB through Rtools and no longer loads tbb.dll), the stale -ltbb/-ltbbmalloc flags and the -L path to RcppParallel's old dynamic TBB directory that StanHeaders::LdFlags() still emits are stripped at configure time, so the rxode2 DLL no longer records an unresolvable runtime dependency on tbb.dll. The strip is keyed to that stale -L<RcppParallel/lib> signature, so a future StanHeaders that emits corrected flags -- or a user-supplied TBB via TBB_LINK_LIB/TBB_LIB -- is left untouched (#1161).

• The parsed_md5 of a model no longer depends on how many models were built before it in the session. linCmtSens was folded into the hash but only assigned after the model was parsed, so the first build of a session hashed with an unset value and every later build hashed with the previous call's value. Because the compiled DLL is named from parsed_md5, the same model could get two different cache keys (and hence a redundant recompile) depending on build order. It is now set before the parse.

• ev$id on an event table now returns the per-row id column (matching as.data.frame(ev)$id) instead of the unique subject ids, so idiomatic subsets like ev[ev$id == 3, ] and per-subject assignments like ev$wt <- 50 + 20 * ev$id no longer silently recycle a short vector; the unique ids remain available via ev$env$ids. [.rxEt now errors on a logical row index whose length matches neither 1 nor the number of rows, and columns assigned with ev$col <- value (new covariates as well as previously hidden canonical columns such as cmt) now round-trip through as.data.frame(ev) (#1154).

• delay()/past() models containing an if/else block failed to solve with unexpected 'else': the DDE helpers parsed the rxNorm() text directly, which puts } and else on separate top-level lines; the normalized text is now parsed wrapped in a { } block. In addition, a past() history inside an if/else branch is now rejected with a clear error (it was invisible to validation), and delay-duration root-variable resolution now sees assignments made inside if/else branches (#1151).

• The vendored SUNDIALS *NewEmpty constructors now allocate with calloc instead of malloc, so any struct fields added by a newer SUNDIALS release are NULL (and safely ignored) rather than uninitialized (#1155).

• Fixed a cross-subject leak in batched multi-subject linCmt() solves: the per-thread inter-event amount buffer was never cleared between subjects, so with cores < nSub every subject after the first on a thread could start from the previous subject's compartment amounts (surfaced by a modeled alag()) (#1153; by Hidde van de Beek).

rxode2 5.1.4

Bug fixes

Model piping

• Model piping no longer shares the meta environment by reference between the original and the piped model. .newModelAdjust() assigned the previous model's meta env directly (to retain sticky items), so both models shared one env -- including the cached simulation model ($meta$.simModelBase). Whichever model was solved first cached its simulation model for both, so a piped model could silently drop an appended compartment/state (e.g. a nonmem2rx import: mod %>% model(d/dt(AUC) <- f, append=TRUE)) or the original model could silently gain the piped model's states/estimates. The meta env is now copied via .copyEnv() (which drops .simModelBase), so each model keeps its own cache.

Compilation

• Silenced the CRAN -Wlto-type-mismatch warnings seen with LTO/gcc builds. The rxSolveWarnPush() forward declaration in src/init.c was missing the variadic ... of its definition, and the ODEPACK /DLS001/ common block was declared with two inconsistent (but memory-equivalent) layouts across the LSODE/LSODA step routines. Both are now declared consistently; the fixes are layout-preserving and the Fortran solvers produce identical results.

rxode2 5.1.3

New features

• rxOptExpr() gains chunkLines and parallel, to optimize a large machine-generated model (a sensitivity- or Jacobian-augmented model) in contiguous cost-balanced chunks rather than in a single pass.

• Delay differential equations: delay(state, T) evaluates an ODE state at t - T (Monolix semantics), with past(state, T) <- expr defining the pre-history. Delayed states are interpolated from the solver's dense output; delay models default to the "dop853+ros4" composite and cap the step size to the smallest delay. The dense-output/history machinery is adapted from the 'dde' package (Rich FitzJohn, Wes Hinsley, Imperial College), whose authors are added as contributors.

• Forward sensitivities for delay models, so delay() models estimate with gradient-based methods such as FOCEi. Parameter-dependent delays are supported at first order (rxDelayD()); second/third order are generated for constant delays (rxDelayD2()/rxDelayD3()) and rejected for parameter-dependent delays (use a numeric or Gauss-Newton Hessian there).

• Many new ODE solver methods: a large suite of explicit Runge-Kutta tableaus (orders 3-14), stiff Rosenbrock and implicit Runge-Kutta methods ("ros43", "ros6", "radauiia5", "gauss6", "sdirk43", "backwardEuler", ...), symplectic steppers, SUNDIALS CVODE ("cvode", linear solver selectable via cvodeLinSolver=), and LSODE/BDF. Implicit methods auto-generate an analytic Jacobian. New helpers rxIsStiff(), rxIsNonStiff(), rxIsImplicit(), rxIsDense(), and rxIsAutoSwitch() classify methods; see the new "ODE solvers" article.

• AutoSwitch composite methods written "primary+secondary" (e.g. "dop853+ros4"): a non-stiff primary with reactive fallback to a stiff secondary, in both the standard and dense-output paths.

• Adjoint sensitivity solving: rxSolveAdjoint() and rxSolveAdjointRk4() return the same rx__sens_<state>_BY_<param>__ output as forward sensitivities via a backward sweep. Exact discrete adjoints exist for the one-step methods ("s" suffix, e.g. "dop853s"), "liblsodaadj", and "cvodesadj", including event jumps (dose/reset/replace/multiply), modeled alag/rate/dur, and steady state. Stiff adjoint and forward-sensitivity solvers integrate the augmented system with its analytic Jacobian.

• Jump sensitivities for dosing events (based on https://github.com/dkaschek/EventSensitivities), replacing finite differences as the default (rxode2.eventSens option: "jump", "fd", "both"). Hybrid jump sensitivities are used for matrix exponential and linCmt() models (up to 3rd order for the ODE/matrix exponential cases).

• Automatic conversion of linear ODE models to linCmt() at solve time (rxSolve(..., useLinCmt=TRUE), the default), passing detected PK parameters explicitly. Handles a central sub-system with an output-only peripheral observable; systems linCmt() cannot represent stay explicit ODEs, and a conversion that will not compile falls back to the ODE (only rxode2).

• Adaptive dosing helpers (bolus(), infuse(), replace(), etc.) now work inside linCmt() and mixed linCmt()+ODE models, with Jacobian handling of the dosing events; odeToLin() preserves and renames them when converting.

• linCmt() sensitivity (linCmtB) solves now run in parallel across subjects on the default forward-mode AD Jacobian path (linCmtSensType="AD"), which is stack-local with no shared Stan arena. The reverse-mode AD ("ADr") and finite-difference paths remain single threaded.

• Inductive linearization and matrix exponentials rewritten with a more NONMEM-like interface (automatic ODE->syntax translation retained) and symbolic-differentiation gradients.

• Added a forward automatic-derivative linear compartment model.

• ar(cor) residual term simulating continuous-time AR(1) residuals for normal, t, and cauchy error models, addable per endpoint alongside any transform; cor is in [0, 1) and the lag correlation decays as cor^(time gap) (Karlsson, Beal and Sheiner 1995). Estimation is supported in nlmixr2est (nlm and focei families).

• lag0()/lead0()/diff0(): like lag()/lead()/diff() but return 0 instead of NA when there is no prior/following record. A calculated variable may now reference itself through lag()/lag0()/diff() (a first-order recurrence); a non-lag self-reference is still a required input parameter.

• rxOmegaVarCovDeriv(): non-Cholesky Omega path returning Omega^{-1}, log|Omega|, and their first/second derivatives with respect to each free variance-covariance element.

• rxExpandSens3_() generates analytic third-order forward sensitivity equations; .rxSens() gained a vars3 argument.

• For downstream packages: rxSetSolveAtolRtol()/rxGetSolveAtolRtol() in the C function-pointer API, and setRxThreadId() so a package can drive the per-subject solve from its own OpenMP team.

• rxTest() test blocks now muffle stray progress messages (e.g. "calculate sensitivities"); set options(rxode2.test.verbose = TRUE) to see them. Messages asserted with expect_message() are unaffected.

• coef() methods for rxUi models (and model functions). By default coef() returns the fixed-effect (theta) estimates; coef(model, level = "omega") returns the random-effect variability matrix and coef(model, level = "all") returns both. nlme::fixef() continues to return the fixed effects.

Bug fixes

• The C accessors exposed through the function-pointer API (getRxNsub(), getSolvingOptions(), getSolvingOptionsInd(), and the other rx_solve* accessors) no longer segfault when handed a NULL or uninitialized solve. They fall back to the global solve; a scalar counter/flag accessor (nsub, nall, nobs, npars, ...) simply reports zero before any solve, exactly as before, so downstream code that probes those counts at load time keeps working (for example babelmixr2's PopED integration, which queries them from .onLoad). An accessor that must dereference a per-subject record (getSolvingOptionsInd()) instead raises a normal catchable R error stating that the solving environment is not set up, rather than dereferencing a NULL pointer and crashing the R process. This hardens downstream packages that call an accessor before their solve pointer has been populated (for example a cold first nls/nlm fit in nlmixr2est).

• A Jacobian entry df(state)/dy(THETA[n]) or df(state)/dy(ETA[n]) (a bracketed parameter reference, which the grammar accepts) no longer segfaults. The synthetic _THETA_n_/_ETA_n_ symbol was never registered, so its index stayed -2 and the model validator read tb.ss.line[-2] out of bounds. This crashed nlm/FOCEi fits that re-parse their generated calcJac model (whose parameters are THETA[n]) in the residual/table step -- notably for a delay-differential-equation model whose delay parameter appears in a product of delayed states.

• past(state, tau) on a state with no d/dt(state) now reports that cleanly instead of corrupting the heap. The error path appended nothing to the message buffer and then trimmed a trailing ', that was never written, moving the write offset before the start of the buffer; the damage surfaced as a double free or corruption abort on a later, unrelated parse rather than at the offending model. The message now names the property ('past(G)' present, but d/dt(G) not defined), and a property with no message branch can no longer underflow the buffer.

• rxOptExpr() no longer fails on a model that uses past(state, tau) and is long enough to be optimized in chunks. A past() line only parses in a chunk that also holds the matching d/dt(), and sensitivity augmentation appends past() after every d/dt() -- so it reliably landed in a chunk of its own. It is now disguised for the duration of the optimization like any other compartment-scoped left-hand side, and restored byte-exactly afterwards. Together with the fix above this unblocks estimating a non-constant-history DDE (e.g. the rheumatoid arthritis model of Koch et al. 2014, J Pharmacokinet Pharmacodyn 41:291-318, Example 6).

• rxAppendModel() now warns (instead of erroring) when the appended models have no variables in common, so the combined model is still returned; use common=FALSE to suppress the warning (#520).

• rxFixPop() no longer tries to literally substitute a fixed mixture proportion (mix()). A mixture proportion must stay a named model-block variable, so substituting its value made the re-parse throw from mix() ("the probabilities in a mixture must be in the model block ..."); a downstream caller wrapping rxFixPop() in try() leaked that error to the console during otherwise-successful mixture fits. Fixed mixture proportions are now excluded from the substitution.

• Tests that use datasets from the suggested nlmixr2data package (theo_sd, warfarin, nmtest) now guard their use with skip_if_not_installed("nlmixr2data"), so the test suite runs cleanly when nlmixr2data is not installed (#95).

Estimation / symengine translation (rxFromSE())

• Convert raw R comparison/logical operators (>, ==, &, ...), not only their rxGt()/rxEq() symengine forms; fixes "user function '>' requires 0 arguments" in FOCEi models with inter-occasion variability (nlmixr2/nlmixr2#390).

• Recognize bare relationals on the second conversion pass of a Subs() over a Derivative(); unblocks FOCEi IOV models that also have a between-subject eta on a parameter without IOV.

• The numeric-constant canonicalization now evaluates operands in baseenv() only and guards zero-length results, fixing an "argument is of length zero" error and silent substitution of user-workspace variables (#1109).

• A trig function (sin/cos/tan) whose argument is a compound expression divided by something (for example sin(2 * 3.14 * (time - mtime1) / period)) no longer drops its whole argument. The division branch fell through without returning when the numerator was not a single token, so the argument became NULL and the emitted C code was sin() -- which failed to compile with "too few arguments to function 'sin'". Such models (for example an enterohepatic gallbladder model with a sinusoidal release) now build and fit (nlmixr2/nlmixr2est#513).

Model parsing / mu-referencing

• Summing two or more population parameters in an expression that has no random effect (for example a combined residual error W <- sqrt(sigma.1. + sigma.2.)) is no longer misreported as "2+ single population parameters in a single mu-referenced expression". That check now fires only for a genuine mu-referenced expression (one that also contains an eta), and the message names the parameters that were actually summed instead of the first parameters in the model (#471).

Delay models

• calcJac=TRUE rewriting (also used by the stiff ros4/dop853+ros4 path) no longer breaks delay models declaring literal THETA_n_/ETA_n_ parameters: constant ~ intermediates stay bound, the literal names are restored, and past() history lines are re-emitted.

• A state read by delay() is always kept as a real ODE, so delayed states named like sensitivities (rx__sens_*) keep their defining d/dt() and can use the stiff/dense composite directly.

• Delay models whose analytic Jacobian cannot be generated now fall back to dop853 (dense) instead of liblsoda, which recorded no delay history and silently returned pre-history values.

• An lhs reading delay() is now reported correctly in the output data frame (#1140). The dense delay history was freed at the end of each subject's solve, so the post-solve lhs recalculation returned the constant pre-history (0) at every record even though the delayed value drove the ODE. The history is now kept until rxSolveFree() releases the subject, which also plugs a leak on the discrete-adjoint (rk4s) path where it was never freed.

linCmt() models

• Fixed a compartment-indexing bug where a model with both an error model and an in-equation compartment reference (e.g. Cp <- peripheral1 / vp) read an unwritten slot.

• tad(<state>)/tlast(<state>) no longer return NA or the wrong value when the model also declares an extra cmt() for an algebraic observable (nlmixr2est#685).

• The automatic linCmt() conversion no longer fires on a nonlinear model whose nonlinearity is written through a state-derived observable (e.g. Michaelis-Menten via Cc <- central / vc).

• The automatic linCmt() conversion no longer changes results when the event data addresses a compartment (in a dose or an observation record) by the name of an ODE compartment the conversion renames (e.g. an ODE centre compartment addressed as CMT = "centre", which the conversion renames to central). Such a solve now falls back to the original ODE model instead of routing the record nowhere and returning all-zero predictions.

• Fixed the automatic linCmt() conversion cache reusing the first model's initial estimates for a later model that shares the same model({}) equations but has a different ini({}) block, which made structurally identical models with different parameters return identical predictions.

• Fixed the string form of the compartment argument in the adaptive dosing helpers (e.g. bolus(50, cmt = "depot")).

Solving

• Zero the LSODA solver work memory on allocation (alloc_mem, calloc instead of malloc). The shared work block (Nordsieck history yh, Jacobian workspace wm, acor/savf, ...) was left uninitialised and parts are read before the integrator writes them on some paths (e.g. a first stiff/BDF step at an extreme point), making a solve non-deterministic. Surfaced by valgrind as reads of uninitialised LSODA memory inside FOCEi/impmap inner solves, and downstream as an occasional blown-up importance-sampling fit run after a prior (parallel) fit. Solving is otherwise unchanged.

• lag()/diff() (and first()/last()) previously returned a constant instead of the prior record's value for calculated variables and time-varying covariates; they now read the prior record (NA on each individual's first record) and work through the estimation/symengine path. Only lag(x, 1)/diff(x, 1) are supported for calculated variables.

• Bug fix for mix() models and iCov models.

• The rxMemoryEstimate() RAM detection no longer calls the defunct utils::memory.limit() (which warned on every Windows solve); total RAM is now queried natively in C (GlobalMemoryStatusEx on Windows, sysctl on macOS, sysconf on Linux) and available memory reuses the allocator preflight estimate (rxAvailableMemoryBytes()). This also drops the memuse suggested dependency and the shell-command fallbacks.

• Fixed out-of-bounds heap reads (AddressSanitizer-confirmed; results unchanged): rxSolve() parameter setup when subjects share one event table in an nsim > 1 sorted solve; syncIdx() dose-index lookup; cvPost() with a 1x1 omega; linCmt.h linCmtStan2ssInf8; etTran() combineDvid; rxDerived() derived1.

rxode2 5.1.2

• geom_cens() / stat_cens() no longer emit "Ignoring unknown aesthetics" warnings when censoring aesthetics are mapped. Documentation corrected to describe the two supported lowercase forms: lower/upper (both required) or cens (with optional limit). The two forms cannot be mixed, lower and upper are now required together, and limit without cens is rejected rather than silently ignored.

• Checks for is.loaded() before loading a rxode2 model. This helps fix the m1 ODR issue shown in nlmixr2est.

• Moved dim.rxEt() here instead of in nlmixr2est

rxode2 5.1.1

• Various low level fixes to allow nlmixr2est to have parallelized focei.

• Parallelized the rxode2 data.frame creation.

• Added parallel solving mirai for clusters and HPC support.

• Added out of memory solve using arrow/duckdb. These out-of-memory (rxSolveOom) solves behave like a standard solved object: it prints the $params and $inits (mirroring the rxSolve console output), supports $, head(), nrow(), ncol()/dim() and the usual as.data.frame()/as_tibble()/ as.data.table() coercions, and exposes the per-subject parameter table and initial conditions that are now persisted alongside the chunked data. A DuckDB query layer over the parquet chunks is used for lazy access (head(), single-column extraction, schema) when available. The chunks can also be queried lazily with dplyr (via as.arrow() or arrow::to_duckdb()) so that filtering and aggregation are pushed down to the on-disk chunks and a possibly out-of-memory result never has to be fully materialized. The storage/query engine can be pinned with the rxode2.oom.backend option ("auto", "duckdb", "arrow" or "rds"); the option is also forwarded to parallel (mirai) workers.

• Use ALTREP for id, sim.id, repeated simulation event columns (evid, cmt, ss, amt, rate, dur, ii, time), covariates and kept variables when blocks are identical across simulations; falls back to filled out columns when runtime event mutation is detected (evid_() push growth / per-individual event reallocation). Also factors cannot currently be represented by altrep, so they are forced to be fully represented.

• Change compile flags and compiler directives for rxode2 models to speed up how they run.

• Have a pre-allocated context pool for lsoda in both liblsoda and lsoda (faster because memory doesn't need to allocated and deallocated so often)

• Change OMP scheduling to dynamic to try to help load-balance the ode solving per subject.

• Simulation normal random numbers before integrating them into your solve.

• Add evid_() function to allow arbitrary doses and observations in a rxode2 model.

• Add splitBolus() function to split or relocate doses in the final output. This is done at translation time (but is respected by evid_()) so in general is a bit faster then arbitrary doses in an estimation step for nlmixr2

• Add %% operator to valid rxode2 syntax

• Create per-individual ODE solving tolerances for use in focei.

• Fix potential security and memory-management issues that could lead to crashes or undefined behavior including integer overflow

• Change dop853 to allow per state tolerances and parallel solving like liblsoda.

• Change dop853 to be able to use dense=TRUE for the 8th order dense polynomial interpolation between dosing events.

• Now dop853 can be parallelized per thread.

• Change mtime state-based dosing to use less memory.

• Add plogis() translation inside rxode2 to it's c-based expit() functions

• Refactored et() to be mostly in R, fixing many issues (#722 , #725, #858, #732, #723, #721, and #724) and allowing dosing/sampling windows to use ii, addl and until (realized immediately)

• Add linToOde() convert linCmt() models to ODEs.

• Fix IOV simulation issue observed in #982.

• Fix sticky variable calculation (#1013, #1025)

• More easily identify initial conditions (#948)

• Fix sensitivities in the linCmt() that did not match the ODE (#1018, #1012)

• Added in-solve addition of observations (obs()), bolus doses bolus(), infusion doses infuse() or infuseDur(), system resets reset(), compartment replacement replace(), multiplicaton events multiply(), and phantom/transit compartment events phantom(). For more granular control you can also use evid_().

• Refactor string comparison in rxode2 so that it is actually doing an integer comparison when running the ODE solving routine (simulation and estimation) instead of using a string comparison. It makes using strings like (sex == "male") run faster.

• Add rxMemoryEstimate() and rxMemSummary() to estimate the amount of memory that is required for a rxode2 solve.

• Add tolFactor, a per individual change of the tolerances to be used in solving. This is used have individualized tolerances from nlmixr2est.

• Add serializeFile as an option to save the rxode2 C fitting data and then restore as needed.

• Add out of memory solve capabilities

rxode2 5.0.2

• Allow state-dependent dur(), rate(), alag(), mtime() now allow states to modify their behavior. The state value at the time of the event is used to calculate any changes.

• Fix: all six ODE solve loops now use precomputed timeThread values for event times instead of recomputing via getTime_() with ypNA, preventing NA propagation for any state-dependent lag scenario.

• Export the internal .rxGetSeed() and .rxSetSeed() for use in the nlmixr2save package.

• Bug fix for .copyUi() with the new format (5.0+) of rxode2 ui models

• With new versions of R, getOption() is no longer a bottleneck, so syncing to local variables is no longer done internally

• Allow transforms to return NA.

• Drop magrittr and use |> instead of %>% in the examples (requires R 4.1)

• Change default model serialization to bzip2 and move binary code generation inside of C.

• Fix where getting seed saves/modifies the RNG scope, as well as a bug fix for restoring the random seed state

rxode2 5.0.1

• Change random number generation to always return doubles internally as well as no longer take a rxode2 individual structure, this is inferred by the thread number.

• Change string representation of model variables to internal binary C code (to avoid macOS M1 sanitizer issues with strings).

• Allow user to change the internal serialization type with options("rxode2.serialize.type"); Currently can be one of "qs2", "qdata", "base", "bzip2" and "xz". This option must be set before rxode2 is loaded, once loaded it keeps the option initially set. This is set to xz which is from base R, but could be sped up with either "qs2" (more future proof) or "qdata" (a bit faster).

• Removed lsoda CDIR$ IVDEP directive, as requested by CRAN.

rxode2 5.0.0

• Better error for tad(depot) when linCmt() doesn't include a depot compartment.

• Remove qs dependency; For rxode2 ui objects, use lists instead of serialized objects. The internal C++ code still generates qs2 sterilization objects (#950)

• Fixed translation for censoring/limit to account for a possible CMT variable before the CENS / LIMIT column (#951, #952)

• Added dmexpit() for getting the diagonal Jacobian.

• Added special handling of mixest and mixunif.

rxode2 4.1.1

• Stacking for multiple-endpoint ipredSim now matches multiple-endpoint sim; Issue #929

• Fix occasional $props that threw an error with empty properties (when using properties like tad0()); Issue #924

• Allow mixture models mix() to be loaded with rxS() as a step to support mixtures in nlmixr2's focei; Issue #933.

• Identify the correct transformation type for iov variables (#936)

• Fix multiple compartment simulation edge cases where simulations were not being performed (#939)

• When referencing cmt in models, the variable is forced to be CMT (related to #939)

• Added ability to use mixest or mixunif to preserve the selected mixture estimates when performing a table step for a nlmixr2 mixture model (#942)

rxode2 4.1.0

• Change rxui $ evaluation when completing in rstudio, fixes strange calculations popping up in rstudio (#909)

• Add orphan rxode2 model unloading when using rxUnloadAll(), and change the return type to always be a boolean.

• Add assertRxUiIovNoCor to assert IOVs have no correlations in them.

• Handle the levels for inter-occasion variability in the ui better (#614)

• Create a new function mix() that will allow mixture models to be simulated in preparation of mixture support in nlmixr2. This allows mixture models to be specified as: v = mix(v1, p1, v2, p2, v3) where the probability of having v=v1 is modeled by p1, v=v2 is modeled by p2, and v=v3 is modeled by probability 1-p1-p2.

• Created new functions mlogit() and mexpit() to convert probabilities used in mixture models to log-scaled values. mlogit() converts the probabilities to log-scaled values (using root-finding) and mexpit() converts the log values into probabilities. The equation for the conversion of log to probabilities is $p_i = \frac{exp(x_i)}{1+\sum_{j=1}^{N-1}exp(x_j)}$

• Added new assertion assertRxUiNoMix which throws an error when a mixture model is present (ie mix())

• Fix for label processing when calling rxode2(uiModel)

rxode2 4.0.3

• For CRAN's m1 ASAN checks of nlmixr2est, loading and unloading the same dll or by deleting the dll and recreating the exact same code, and then loading the dll will cause the ASAN check to flag an odr violation. Because of this, a mechanism to not unload dlls has been added. This allows the next version of nlmixr2est to not have issues with Mac m1 san checks.

rxode2 4.0.2

• At the request of CRAN, be a bit more careful so that names are not duplicated. Now include the md5 hash, a global counter and random 4 digit and number combination. In addition add the name of the original function so it will be easier to debug in the future.

• Fall back to data.frame rbind when rbind.rxSolve() fails

rxode2 4.0.1

• Add the ability to use rbind for solved rxode2 frames.

• Fix LTO issue for _rxode2_calcDerived

rxode2 4.0.0

• Add more information errors about NAs during solving.

• Fix rxDerived() for mixed vector and non-vector input.

• Fix model variables for alag(cmt) when they are defined before d/dt() or linCmt()

• Just in time use of state.ignore in the model variables, fixes negative length error observed in #857.

• Fix steady state bug with time-varying covariates. Now the covariates are inferred at the time of the steady state (instead of searching through the subject based on the projected time).

• Rework the linear solved systems to use the wnl solutions, and threaded linear systems solve (for non-gradient solutions). This new method closes a variety of linear compartment model bugs (#261, #272, #441, #504, #564, #717, #728, #827, and #855)

• Added new types of bounds for event tables:

  • 3 point bounds et(list(c(low, mid, high))) when specified this way, they will not change. Perfect for use with babelmixr2's PopED (#862, #863, #854)

  • Intervals simulated by normal values instead of uniform. In this case the first seen interval will be 3 elements with NA at the end et(list(c(mean, sd, NA), c(mean, sd))), and the other elements can simply be 2 declaring the c(mean, sd)

  • Of course the uniform windows of et(list(c(low, high))) still work

  • Currently these different types of windows cannot be mixed.

• Add ability to pipe a list or named numeric as an eta with %>% ini(~eta)

• Added a fix for event tables where expanding IDs in non-sequential order. In particular if the first ID is not the minimum ID when expanding the first event table, the smallest ID was not in the output table. Now the smallest ID is in the event table. (Fixes #878, #869, #870)

• Added ability to pipe ini() or lotri(), or any other expression that can be converted to an ini with as.ini(). Also allows ini() expressions to be converted to lotri with as.lotri(). Fixes #871

• Added new type of variability expression for simulation and estimation with focei and likelihood related methods: +var(). This changes standard deviation parameters to variance parameters.

• Added new type of endpoint expression for focei estimation +dv(). This only transforms the data and not the predictions. I can only see it being useful in model linearization.

• Bug fix for parameters that are in both input ($params) and output ($lhs) that respects the order of the $lhs declaration (Fixes #876)

• Add rxFixRes to literally fix the residual estimates in a model (#889)

• Now modeled duration of 0 is treated as a bolus dose (#892)

rxode2 3.0.4

• Add stable hashes for rxUi objects (#838, #689)

• Fix for iov simulation (#842)

• Fix for rxnbinom() called directly from R (#847) and expand it to match more close with R's rnbinom() including allowing named mu= calls. In rxode2 ui, these are also now allowed.

rxode2 3.0.3

• Add logit/expit named expressions, that is logit(x, high=20) becomes logit(x, 0, 20) in ui models.

• Updated random ui models like rxnorm(sd=10) to accept complex numeric expressions like rxnorm(sd=10+1).

• Updated random ui models to accept complex non-numeric expressions like rxnorm(sd=a+b)

• Rework the tad() and related functions so they use the same interface as compartments (this way they do not depend on the order of compartments); See #815. For mu-referencing, Also allow dummy variables to ignore state requirements (ie podo(depot) in a single line will not error when parsing mu-referenced equations).

• Add getRxNpars to api. This allows the development version of babelmixr2 to better check what model is loaded and unload/reload as necessary.

• Add rxUdfUiControl() to rxode2 user function to get control information from something like nlmixr2

• Bug fix for tracking time after dose when dosing to 2 compartments occur at the exact same time (#804, #819)

• Change transit() model so that it uses tad0(), podo0() and related functions for a bit more stable simulation and estimation

• Fix compile flags to work with BH 1.87 (#826)

rxode2 3.0.2

• Bug fix for api, the censoring function pointer has been updated (#801).

• Query rxode2.verbose.pipe at run time instead of requiring it to be set before loading rxode2.

• Have correct values at boundaries for logit, expit, probit, and probitInv (instead of NA). For most cases this does not break anything.

• Add a new style of user function that modifies the ui while parsing or just before using the function (in the presence of data).

• Used the new user function interface to allow all random functions in rxode2 ui functions to be named. For example, you can use rxnorm(sd=3) instead of having to use rxnorm(0, 3), although rxnorm() still works.

rxode2 3.0.1

• Explicitly initialize the order vector to stop valgrind warning (requested from CRAN)

rxode2 3.0.0

Breaking Changes

• The model properties was moved from $params to $props so it does not conflict with the low level rxode2 model $params

• Error when specifying wd without modName

• With Linear and midpoint of a time between two points, how rxode2 handles missing values has changed. When the missing value is lower than the requested time, it will look backward until it finds the first non-missing value (or if all are missing start looking forward). When the missing value is higher than the requested time, the algorithm will look forward until it finds the first non-missing value (or if all are missing, start looking backward).

• The order of ODEs is now only determined by the order of cmt() and d/dt(). Compartment properties, tad() and other compartment related variables no no longer affect compartment sorting. The option rxode2.syntax.require.ode.first no longer does anything.

• The handling of zeros "safely" has changed (see #775)

  • when safeZero=TRUE and the denominator of a division expression is zero, use the Machine's small number/eps (you can see this value with .Machine$double.eps)

  • when saveLog=TRUE and the x in the log(x) is less than or equal to zero, change this to log(eps)

  • when safePow=TRUE and the expression x^y has a zero for x and a negative number for y replace x with eps.

  Since the protection for divide by zero has changed, the results will also change. This is a more conservative protection mechanism than was applied previously.

• Random numbers from rxode2 are different when using dop853, lsoda or indLin methods. These now seed the random numbers in the same way as liblsoda, so the random number provided will be the same with different solving methods.

• The arguments saved in the rxSolve for items like thetaMat will be the reduced matrices used in solving, not the full matrices (this will likely not break very many items)

Possible breaking changes (though unlikely)

• iCov is no longer merged to the event dataset. This makes solving with iCov slightly faster (#743)

New features

• You can remove covariances for every omega by piping with %>% ini(diag()) you can be a bit more granular by removing all covariances that have either eta.ka or eta.cl by: %>% ini(diag(eta.ka, eta.cl)) or anything with correlations with eta.cl with %>% ini(diag(eta.cl))

• You can also remove individual covariances by %>% ini(-cov(a, b)) or %>% ini(-cor(a,b)).

• You can specify the type of interpolation applied for added dosing records (or other added records) for columns that are kept with the keep= option in rxSolve(). This new option is keepInterpolation and can be locf for last observation carried forward, nocb which is the next observation carried backward, as well as NA which puts a NA in all imputed data rows. See #756.

  • Note: when interpolation is linear/midpoint for factors/characters it changes to locf with a warning (#759)

  • Also note, that the default keep interpolation is na

• Now you can specify the interpolation method per covariate in the model:

  • linear(var1, var2) says both var1 and var2 would use linear interpolation when they are a time-varying covariate. You could also use linear(var1)

  • locf() declares variables using last observation carried forward

  • nocb() declares variables using next observation carried backward

  • midpoint() declares variables using midpoint interpolation

• linear(), locf(), locb(), midpoint(), params(), cmt() and dvid() declarations are now ignored when loading a rxode2 model with rxS()

• Strings can be assigned to variables in rxode2.

• Strings can now be enclosed with a single quote as well as a double quote. This limitation was only in the rxode2 using string since the R-parser changes single quotes to double quotes. (This has no impact with rxode2({}) and ui/function form).

• More robust string encoding for symengine (adapted from utils::URLencode() and utils::URLdecode())

• Empty arguments to rxRename() give a warning (#688)

• Promoting from covariates to parameters with model piping (via ini()) now allows setting bounds (#692)

• Added assertCompartmentName(), assertCompartmentExists(), assertCompartmentNew(), testCompartmentExists(), assertVariableExists() testVariableExists(), assertVariableNew(), assertVariableName(), and assertParameterValue() to verify that a value is a valid nlmixr2 compartment name, nlmixr2 compartment/variable exists in the model, variable name, or parameter value (#726; #733)

• Added assertRxUnbounded(), testRxUnbounded(), warnRxBounded() to allow nlmixr2 warn about methods that ignore boundaries #760

• Added functions tad0(), tafd0(), tlast0() and tfirst0() that will give 0 instead of NA when the dose has not been administered yet. This is useful for use in ODEs since NAs will break the solving (so can be used a bit more robustly with models like Weibull absorption).

• rxode2 is has no more binary link to lotri, which means that changes in the lotri package will not require rxode2 to be recompiled (in most cases) and will not crash the system.

• rxode2 also has no more binary linkage to PreciseSums

• The binary linkage for dparser is reduced to C structures only, making changes in dparser less likely to cause segmentation faults in rxode2 if it wasn't recompiled.

• A new model property has been added to $props$cmtProp and $statePropDf. Both are data-frames showing which compartment has properties (currently ini, f, alag, rate and dur) in the rxode2 ui model. This comes from the lower level model variable $stateProp which has this information encoded in integers for each state.

• A new generic method rxUiDeparse can be used to deparse meta information into more readable expressions; This currently by default supports lower triangular matrices by lotri, but can be extended to support other types of objects like 'nlmixr2's foceiControl() for instance.

Bug fixes

• Fix ui$props$endpoint when the ui endpoint is defined in terms of the ode instead of lhs. See #754

• Fix ui$props when the ui is a linear compartment model without ka defined.

• Model extraction modelExtract() will now extract model properties. Note that the model property of alag(cmt) and lag(cmt) will give the same value. See #745

• When assigning reserved variables, the parser will error. See #744

• Linear interpolation will now adjust the times as well as the values when NA values are observed.

• Fix when keeping data has NA values that it will not crash R; Also fixed some incorrect NA interpolations. See #756

• When using cmt() sometimes the next statement would be corrupted in the normalized syntax (like for instance locf); This bug was fixed (#763)

• keep will now error when trying to keep items that are in the rxode2 output data-frame and will be calculated (#764)

Big change

• At the request of CRAN, combine rxode2parse, rxode2random, and rxode2et into this package; The changes in each of the packages are now placed here:

rxode2et (no changes before merge)

rxode2et 2.0.13

• Fix import of data where there are NA times

rxode2et 2.0.12

• Fix formatting issues identified by m1mac, as requested by CRAN

rxode2et 2.0.11

• Make the stacking more flexible to help rxode2 have more types of plots

• Add toTrialDuration by Omar Elashkar to convert event data to trial duration data

• Fix Issue #23 and prefer variable values over NSE values

rxode2et 2.0.10

• Fix dollar sign accessing of objects (like data frames), as pointed out by @frbrz (issue #16)

• Use rxode2parse functions for internal event table creation (where they were moved to).

• Dropped C++14 and let the system decide.

rxode2et 2.0.9

• Split off et(), eventTable() and related functions.

• Also split off rxStack() and rxCbindStudyIndividual() in this package.

• Added a NEWS.md file to track changes to the package.

rxode2random (before merge)

• Fix a bug when simulating nested variables (#25)

rxode2random 2.1.0

• Breaking Change changed distributions from the standard C++ <random> to boost::random. Since this is not dependent on the compiler, it makes the random numbers generated from Mac, Windows and Linux the same for every distribution. Unfortunately with a new random number transformation, the simulation results will likely be different than they were before. The exception to this is the uniform number, which was always the same between platforms.

rxode2random 2.0.13

• Fixed formatting issues (as requested by CRAN and identified on m1mac)

rxode2random 2.0.12

• Added function dfWishart which gives (by simulation) an approximation of the degrees of freedom of a Wishart to match a rse value.

• Added function swapMatListWithCube which swaps omegaList with omegaCube values

• Ensure that the outputs are integers (instead of long integers) as requested by CRAN for some checking functions.

rxode2random 2.0.11

• Fix qassert LTO

rxode2random 2.0.10

• Moved fast factor to rxode2parse to allow etTrans to be moved there

rxode2random 2.0.9

• Initial release of rxode2random, which separates the parallel safe, random number generation from 'rxode2' into a separate package to reduce 'rxode2' compilation time. This should make CRAN maintenance a bit easier.

• Added a NEWS.md file to track changes to the package.

rxode2parse (fixed before merging)

• As requested by CRAN remove the C code SET_TYPEOF which is no longer part of the C R API.

rxode2parse 2.0.19

• Added a evid suffix of 60 for cases where evid=2 adds an on event (fixes tad() calculation in certain edge cases)

• Initialize all variables to NA

rxode2parse 2.0.18

• Removed linear compartment solutions with gradients from rxode2parse (and rxode2) when compiled with intel c++ compiler (since it crashes while compiling).

• Fixed m1mac string issues as requested by CRAN

rxode2parse 2.0.17

• Added ability to query R user functions in a rxode2 model (will force single threaded solve)

• Moved core rxFunParse and rxRmFunParse here so that C and R user function clashes can be handled

• Model variables now tracks which compartments have a lag-time defined

• For compartment with steady state doses (NONMEM equivalent SS=1, SS=2), an additional tracking time-point is added at to track the time when the lagged dose is given. As an upshot, the lagged dose will start at the steady state concentration shifted by + ii - lag in rxode2 (currently for ode systems only)

• This release calculates non bio-availability adjusted duration for all rates instead of trying to figure the rate duration during solving.

• Make double assignment an error, ie a <- b <-

• NA times are ignored (with warning)

• Steady state bolus doses with addl are treated as non steady state events (like what is observed in NONMEM)

• Timsort was upgraded; drop radix support in rxode2 structure

• etTrans now supports keeping logical vectors (with the appropriate version of rxode2).

• Security fixes were applied as requested by CRAN

rxode2parse 2.0.16

• Import data.table explicitly in the R code (before was imported only in C/C++ code)

rxode2parse 2.0.15

• Updates the make flags to support CXX17.

rxode2parse 2.0.14

• 'linCmt()' translations of 'alpha', 'beta', 'gamma', 'k21', 'k31', 'vc' now error instead of ignoring 'gamma' and 'k31' to give 2 cmt solution

• transit compartment internal code now changes dose to 0.0 when no dose has been administered to the depot compartment. This way dosing to the central compartment (without dosing to the transit compartment) will not give a NA for the depot compartment (and consequently for the central compartment)

• Moved rxDerived here and added tests for it here as well.

• Moved etTransParse here and added tests for it here as well (makes up most of etTrans). In addition the following changes were made to etTransParse()/etTrans():

  • The internal translation (etTrans()) will not drop times when infusions stop. Before, if the infusion stopped after the last observation the time when the infusion stopped would be dropped. This interferes with linCmt() models.

  • Breaking change/bug fix evid=2 are considered observations when translating data to internal rxode2 event structure

  • Fix edge case to find infusion duration when it is the first item of the dosing record at time 0.

• Fixed a bug for certain infusions where the rate, ii and/or ss data items were dropped from the output when addDosing=TRUE

• Also have internal functions to convert between classic NONMEM events and rxode2 events

• Have an internal function that gives information on the linear compartmental model translation type, which could be useful for babelmixr2

• 'time' in model is now case insensitive

• Use function declaration in rxode2parseGetTranslation() to determine thread safety of functions available to rxode2

• Add check for correct number of function arguments to parser.

• Like R, known functions can be assigned as a variable and the function can still be called (while not changing the variable value). For example you can have a variable gamma as well as a function gamma().

• Fix garbled error messages that occur with certain messages.

• Fixed errors that occurred when using capitalized AMT variables in the model.

rxode2parse 2.0.13

• Version bump for dparser (so binaries will be built correctly)

rxode2parse 2.0.12

• Bug fix for strict prototypes

• Removed sprintf as noted by CRAN

• Made rxode2parse dll binary independent of rxode2()

rxode2parse 2.0.11

• Bug fix for strict aliasing as requested by CRAN

rxode2parse 2.0.10

• Use strict aliasing as requested by CRAN

rxode2parse 2.0.9

• Initial release to split of rxode2parse from rxode2 to reduce compilation time of 'rxode2'

rxode2 2.1.3

Bug fixes

• Make sure that the object is a uncompressed rxode2 ui for solving with rxSolve (See #661)

• Fix #670 by using the last simulated observation residual when there are trailing doses.

New features

• Create a function to see if a rxode2 solve is loaded in memory (rxode2::rxSolveSetup())

• Create a new function that fixes the rxode2 population values in the model (and drops them in the initial estimates); rxFixPop()

Other changes

• Pendantic no-remap (as requested by CRAN)

• gcc USBAN fix (as requested by CRAN)

rxode2 2.1.2

Other changes

• rxUi compression now defaults to fast compression

• Fixes String literal formatting issues as identified by CRAN (#643)

• Removes linear compartment solutions with gradients for intel c++ compiler (since they crash the compiler).

rxode2 2.1.0

Breaking changes

• Steady state with lag times are no longer shifted by the lag time and then solved to steady state by default. In addition the steady state at the original time of dosing is also back-calculated. If you want the old behavior you can bring back the option with ssAtDoseTime=FALSE.

• "dop853" now uses the hmax/h0 values from the rxControl() or rxSolve(). This may change some ODE solving using "dop853"

• When not specified (and xgxr is available), the x axis is no longer assumed to be in hours

New features

• User defined functions can now be R functions. For many of these R functions they can be converted to C with rxFun() (you can see the C code afterwards with rxC("funName"))

• Parallel solving of models that require sorting (like modeled lag times, modeled duration etc) now solve in parallel instead of downgrading to single threaded solving

• Steady state infusions with a duration of infusions greater than the inter-dose interval are now supported.

• Added $symengineModelNoPrune and $symengineModelPrune for loading models into rxode2 with rxS()

• When plotting and creating confidence intervals for multiple endpoint models simulated from a rxode2 ui model, you can plot/summarize each endpoint with sim. (ie. confint(model, "sim") or plot(model, sim)).

  If you only want to summarize a subset of endpoints, you can focus on the endpoint by pre-pending the endpoint with sim. For example if you wanted to plot/summarize only the endpoint eff you would use sim.eff. (ie confint(model, "sim.eff") or plot(model, sim.eff))

• Added model$simulationIniModel which prepend the initial conditions in the ini({}) block to the classic rxode2({}) model.

• Now model$simulationModel and model$simulationIniModel will save and use the initialization values from the compiled model, and will solve as if it was the original ui model.

• Allow ini(model) <- NULL to drop ini block and as.ini(NULL) gives ini({}) (Issue #523)

• Add a function modelExtract() to extract model lines to allow modifying them and then changing the model by piping or simply assigning the modified lines with model(ui) <- newModifiedLines

• Add Algebraic mu-referencing detection (mu2) that allows you to express mu-referenced covariates as:

cl <- exp(tcl + eta.cl + wt_cl * log(WT/70.5))

Instead of the

cl <- exp(tcl + eta.cl + wt_cl * log.WT.div.70.5)

That was previously required (where log.WT.div.70.5 was calculated in the data) for mu expressions. The ui now has more information to allow transformation of data internally and transformation to the old mu-referencing style to run the optimization.

• Allow steady state infusions with a duration of infusion greater than the inter-dose interval to be solved.

• Solves will now possibly print more information when issuing a "could not solve the system" error

• The function rxSetPipingAuto() is now exported to change the way you affect piping in your individual setup

• Allow covariates to be specified in the model piping, that is mod %>% model(a=var+3, cov="var") will add "var" as a covariate.

• When calculating confidence intervals for rxode2 simulated objects you can now use by to stratify the simulation summary. For example you can now stratify by gender and race by: confint(sim, "sim", by=c("race", "gender"))

• When calculating the intervals for rxode2 simulated objects you can now use ci=FALSE so that it only calculates the default intervals without bands on each of the percentiles; You can also choose not to match the secondary bands limits with levels but use your own ci=0.99 for instance

• A new function was introduced meanProbs() which calculates the mean and expected confidence bands under either the normal or t distribution

• A related new function was introduced that calculates the mean and confidence bands under the Bernoulli/Binomial distribution (binomProbs())

• When calculating the intervals for rxode2 simulated objects you can also use mean=TRUE to use the mean for the first level of confidence using meanProbs(). For this confidence interval you can override the n used in the confidence interval by using n=#. You can also change this to a prediction interval instead using pred=TRUE.

• Also when calculating the intervals for rxode2 simulated object you can also use mean="binom" to use the binomial distributional information (and ci) for the first level of confidence using binomProbs(). For this confidence interval you can override the n used in the confidence interval by using n=#. You can also change this to a prediction interval instead using pred=TRUE. With pred=TRUE you can override the number of predicted samples with m=#

• When plotting the confint derived intervals from an rxode2 simulation, you can now subset based on a simulated value like plot(ci, Cc) which will only plot the variable Cc that you summarized even if you also summarized eff (for instance).

• When the rxode2 ui is a compressed ui object, you can modify the ini block with $ini <- or modify the model block with $model <-. These are equivalent to ini(model) <- and model(model) <-, respectively. Otherwise, the object is added to the user defined components in the function (ie $meta). When the object is uncompressed, it simply assigns it to the environment instead (just like before).

• When printing meta information that happens to be a lotri compatible matrix, use lotri to express it instead of the default R expression.

• Allow character vectors to be converted to expressions for piping (#552)

• rxAppendModel() will now take an arbitrary number of models and append them together; It also has better handling of models with duplicate parameters and models without ini() blocks (#617 / #573 / #575).

• keep will now also keep attributes of the input data (with special handling for levels); This means a broader variety of classes will be kept carrying more information with it (for example ordered factors, data frame columns with unit information, etc)

• Piping arguments append for ini() and model() have been aligned to perform similarly. Therefore ini(append=) now can take expressions instead of simply strings and model(append=) can also take strings. Also model piping now can specify the integer line number to be modified just like the ini() could. Also model(append=FALSE) has been changed to model(append=NULL). While the behavior is the same when you don't specify the argument, the behavior has changed to align with ini() when piping. Hence model(append=TRUE) will append and model(append=FALSE) will now pre-pend to the model. model(append=NULL) will modify lines like the behavior of ini(append=NULL). The default of model(line) modifying a line in-place still applies. While this is a breaking change, most code will perform the same.

• Labels can now be dropped by ini(param=label(NULL)). Also parameters can be dropped with the idiom model(param=NULL) or ini(param=NULL) changes the parameter to a covariate to align with this idiom of dropping parameters

• rxRename has been refactored to run faster

Internal new features

• Add as.model() for list expressions, which implies model(ui) <- ui$lstExpr will assign model components. It will also more robustly work with character vectors

• Simulated objects from rxSolve now can access the model variables with $rxModelVars

• Simulation models from the UI now use rxerr.endpoint instead of err.endpoint for the sigma residual error. This is to align with the convention that internally generated variables start with rx or nlmixr

• Sorting only uses timsort now, and was upgraded to the latest version from Morwenn

Bug fixes

• Simulating/solving from functions/ui now prefers params over omega and sigma in the model (#632)

• Piping does not add constants to the initial estimates

• When constants are specified in the model({}) block (like k <- 1), they will not be to the ini block

• Bug fix for geom_amt() when the aes transformation has x

• Bug fix for some covariate updates that may affect multiple compartment models (like issue #581)

Maintenance fixes

• Modify plot code to work with development xgxr

rxode2 2.0.14

• CRAN requested that FORTRAN kind be changed as it was not portable; This was commented code, and simply removed the comment.

• Bug-fix for geom_amt(); also now uses linewidth and at least ggplot2 3.4.0

• Some documentation was cleaned up from rxode2 2.0.13

rxode2 2.0.13

Bug fixes

• A bug was fixed so that the zeroRe() function works with correlated omega values.

• A bug was fixed so that the rename() function works with initial conditions for compartments (cmt(0))

New features

• A new function zeroRe() allows simple setting of omega and/or sigma values to zero for a model (#456)

• Diagonal zeros in the omega and sigma matrices are treated as zeros in the model. The corresponding omega and sigma matrices drop columns/rows where the diagonals are zero to create a new omega and sigma matrix for simulation. This is the same idiom that NONMEM uses for simulation from these matrices.

• Add the ability to pipe model estimates from another model by parentModel %>% ini(modelWithNewEsts)

• Add the ability to append model statements with piping using %>% model(x=3, append=d/dt(depot)), still supports appending with append=TRUE and pre-pending with append=NA (the default is to replace lines with append=FALSE)

• rxSolve's keep argument will now maintain character and factor classes from input data with the same class (#190)

• Parameter labels may now be modified via ini(param = label("text")) (#351).

• Parameter order may be modified via the append argument to ini() when piping a model. For example, ini(param = 1, append = 0) or ini(param = label("text"), append = "param2") (#352).

Internal changes

• If lower/upper bounds are outside the required bounds, the adjustment is displayed.

• When initial values are piped that break the model's boundary condition reset the boundary to unbounded and message which boundary was reset.

• Added as.rxUi() function to convert the following objects to rxUi objects: rxode2, rxModelVars, function. Converting nlmixr2 fits to rxUi will be placed in the s3 method in the corresponding package.

• assertRxUi(x) now uses as.rxUi() so that it can be extended outside of rxode2/nlmixr2.

• rxode2 now supports addl with ss doses

• Moved rxDerived to rxode2parse (and re-exported it here).

• Added test for transit compartment solving in absence of dosing to the transit compartment (fixed in rxode2parse but solving tested here)

• Using ini() without any arguments on a rxode2 type function will return the ini() block. Also added a method ini(mod) <- iniBlock to modify the ini block is you wish. iniBlock should be an expression.

• Using model() without any arguments on a rxode2 type function will return the model() block. Also added a new method model(mod) <- modelBlock

• Added a new method rxode2(mod) <- modFunction which allows replacing the function with a new function while maintaining the meta information about the ui (like information that comes from nonmem2rx models). The modFunction should be the body of the new function, the new function, or a new rxode2 ui.

• rxode2 ui objects now have a $sticky item inside the internal (compressed) environment. This $sticky tells what variables to keep if there is a "significant" change in the ui during piping or other sort of model change. This is respected during model piping, or modifying the model with ini(mod)<-, model(mod)<-, rxode2(mod)<-. A significant change is a change in the model block, a change in the number of estimates, or a change to the value of the estimates. Estimate bounds, weather an estimate is fixed or estimate label changes are not considered significant.

• Added as.ini() method to convert various formats to an ini expression. It is used internally with ini(mod)<-. If you want to assign something new that you can convert to an ini expression, add a method for as.ini().

• Added as.model() method to convert various formats to a model expression. It is used internally with model(mod)<-. If you want to assign something new that you can convert to a model expression, add a method for as.model().

rxode2 2.0.11

• Give a more meaningful error for 'rxode2' ui models with only error expressions

• Break the ABI requirement between roxde2() and rxode2parse()

• The new rxode2parse will fix the sprintf exclusion shown on CRAN.

rxode2 2.0.10

• Time invariant covariates can now contain 'NA' values.

• When a column has 'NA' for the entire id, now 'rxode2' warns about both the id and column instead of just the id.

• To fix some CRAN issues in 'nlmixr2est', make the version dependency explicit.

rxode2 2.0.9

• Remove log likelihoods from 'rxode2' to reduce compilation time and increase maintainability of 'rxode2'. They were transferred to 'rxode2ll' (requested by CRAN).

• Remove the parsing from 'rxode2' and solved linear compartment code and move to 'rxode2parse' to reduce the compilation time (as requested by CRAN).

• Remove the random number generation from 'rxode2' and move to 'rxode2random' to reduce the compilation time (as requested by CRAN).

• Remove the event table translation and generation from 'rxode2' and move to 'rxode2et' to reduce the compilation time (as requested by CRAN).

• Change the rxode2 ui object so it is a compressed, serialized object by default. This could reduce the C stack size problem that occurs with too many environments in R.

• Warn when ignoring items during simulations

• Export a method to change 'rxode2' solve methods into internal integers

• Bug fix for time invariant covariates identified as time variant covariate when the individual's time starts after 0.

rxode2 2.0.8

Breaking changes

• rxgamma now only allows a rate input. This aligns with the internal rxode2 version of rxgamma and clarifies how this will be used. It is also aligned with the llikGamma function used for generalized likelihood estimation.

• ui cauchy simulations now follow the ui for normal and t distributions, which means you can combine with transformations. This is because the cauchy is a t distribution with one degree of freedom.

• ui dnorm() and norm() are no longer equivalent to add(). Now it allows you to use the loglik llikNorm() instead of the standard nlmixr2 style focei likelihood. This is done by adding dnorm() at the end of the line. It also means dnorm() now doesn't take any arguments.

• Vandercorput normal removed (non-random number generator)

New features

• Allow models in the nlmixr2 form without an ini({}) block

• Allow model piping of an omega matrix by f %>% ini(omegaMatrix)

• Standard models created with rxode2() can no be piped into a model function

• Families of log-likelihood were added to rxode2 so that mixed likelihood nonlinear mixed effects models may be specified and run.

• The memory footprint of a rxode2 solving has been reduced

• Piping now allow named strings (issue #249)

Bug fixes

• rxode2's symengine would convert sqrt(2) to M_SQRT_2 when it should be M_SQRT2. This has been fixed; it was most noticeable in nlmixr2 log-likelihood estimation methods

• rxode2 treats DV as a non-covariate with etTran (last time it would duplicate if it is in the model). This is most noticeable in the nlmixr2 log-likelihood estimation methods.

New features

• A new flag (rxFlag) has been created to tell you where in the rxode2 solving process you are. This is useful for debugging. If outputting this variable it will always be 11 or calculating the left handed equations. If you are using in conjunction with the printf() methods, it is a double variable and should be formatted with "%f".

• An additional option of fullPrint has been added to rxode2() which allows rprintf() to be used in almost all of rxode2() steps (inductive linearization and matrix exponential are the exception here) instead of just the integration ddt step. It defaults to FALSE.

rxode2 2.0.7

• Removed accidental ^S from news as requested by CRAN.

• Bug fix for more complicated mu-referencing.

• Change rxode2 md5 to only depend on the C/C++/Fortran code and headers not the R files. That way if there is binary compatibility between nlmixr2est and rxode2, a new version of nlmixr2est will not need to be submitted to CRAN.

rxode2 2.0.6

Breaking changes

Solving controls

• The options for rxControl and rxSolve are more strict. camelCase is now always used. Old options like add.cov and transit_abs are no longer supported, only addCov is supported.

• A new option, sigdig has been added to rxControl(), which controls some of the more common significant figure options like atol, rtol, ssAtol, ssRtol, with a single option.

Simulations

• For simulations, $simulationSigma now assumes a diagonal matrix. The sigma values are assumed to be standard normal, and uncorrelated between endpoints. Simulation with uncertainty will still draw from this identity diagonal matrix

• Parallel solving now seeds each simulation per each individual based on the initial seed plus the simulation id. This makes the simulation reproducible regardless of the number of cores running the simulation.

Other breaking changes

• Solved objects now access the underlying rxode model with $rxode2 instead of $rxode

• Since this change names, rxode2, rxode and RxODE all perform the same function.

• Options were changed from RxODE.syntax to rxode2.syntax.

• Assigning states with rxode2.syntax.assign.state (was RxODE.syntax.assign.state) is no longer supported.

• Enforcing "pure" assignment syntax with = syntax is no longer supported so rxode2.syntax.assign is no longer supported (was RxODE.syntax.assign).

• Since R supports ** as an exponentiation operator, the pure syntax without ** can no longer be enabled. Hence rxode2.syntax.star.pow (was RxODE.syntax.star.pow) no longer has any effect.

• The "pure" syntax that requires a semicolon can no longer be enabled. Therefore rxode2.syntax.require.semicolon (was RxODE.syntax.require.semicolon) no longer has any effect.

• The syntax state(0) can no longer be turned off. rxode2.syntax.allow.ini0 (was RxODE.syntax.allow.ini0) has been removed.

• Variable with dots in variable and state names like state.name works in R. Therefore, "pure" syntax of excluding . values from variables cannot be enforced with rxode2.syntax.allow.dots (was RxODE.syntax.allow.dots).

• The mnemonic et(rate=model) and et(dur=model) mnemonics have been removed. rate needs to be set to -1 and -2 manually instead.

• The function rxode2Test() has been removed in favor of using testthat directly.

• Transit compartments need to use a new evid, evid=7. That being said, the transitAbs option is no longer supported.

• ID columns in input parameter data frames are not sorted or merged with original dataset any more; The underlying assumption of ID order should now be checked outside of rxode2(). Note that the event data frame is still sorted.

Additional features

• The UI functions of nlmixr have been ported to work in rxode2 directly.

• rxModelVars({}) is now supported.

• You may now combine 2 models in rxode2 with rxAppendModel(). In fact, as long as the first value is a rxode2 evaluated ui model, you can use c/rbind to bind 2 or more models together.

• You may now append model lines with piping using %>% model(lines, append=TRUE) you can also pre-pend lines by %>% model(lines, append=NA)

• You may now rename model variables, states and defined parameters with %>% rxRename(new=old) or if dplyr is loaded: %>% rename(new=old)

• You can fix parameters with %>% ini(tcl=fix) or %>% ini(fix(tcl)) as well as unfix parameters with %>% ini(tcl=unfix) or %>% ini(unfix(tcl))

Internal changes

• Strict R headers are enforced more places

• Since there are many changes that could be incompatible, this version has been renamed to rxode2

• rxode2() printout no longer uses rules and centered headings to make it display better on a larger variety of systems.

Bug fixes

• tad() and related time features only reset at the start of an infusion (as opposed to starting at the beginning and end of an infusion)

RxODE 1.1.3

• Change handling of missing covariates while interpolating "nocb" so that the time-varying covariates use "nocb" interpolation (#469)

RxODE 1.1.2

• Fix subject initialization of focei problem (#464)

• Fix LHS offset to allow internal threading and more parallel processing in the future.

• Remove warnings for duration and rate

• Don't export pillar methods any more (simply register at load if present)

• As requested by CRAN, change fortran and C binding for BLAS an LINPACK

RxODE 1.1.1

• Fix the LTO issue that CRAN identified.

• Move the omp files so they come first to support clang13, as identified by CRAN.

• For now, be a little more conservative in dur() and rate() warnings because linCmt() models in nlmixr currently produce irrelevant warnings.

RxODE 1.1.0

• Always calculate "nolhs" for using numeric differences when the inner problem. This allows the inner problem to fallback to a finite difference approximation to the focei objective function.

• Updated the parser C code grammar using latest dparser CRAN package

• Added a new cbind function that is used to mix data frame input with simulated individual parameters and residual parameters, rxCbindStudyIndividual().

• Now data frame input can be mixed with simulating from omega and sigma matrices (though not yet in nested simulations)

• Race conditions when simulating random numbers is solved by chunking each simulation into groups that will always be performed per each thread. This way the simulation is now reproducible regardless of load. Because of the chunking, simulations with random numbers generated inside of it are now threaded by default (though a warning is produced about the simulation only be reproducible when run with the same number of threads)

• Simulations were double checked and made sure to use the engine reserved for each core run in parallel; Some of the random generators were not taking random numbers from the correct engine, which was corrected. Therefore, simulations from this version are expected to be different (in parallel) than previous versions.

• Added function rxSetSeed() to set the internal RxODE seed instead of grabbing it from a uniform random number tied to the original R seed. This will avoid the possibility of duplicate seeds and is the best practice.

• Updating parameter pointers is done once per ID and locked based on ID to remove the recursion in #399, but still have the correct behavior see #430

• Parsing updated to retain "param()" in normalized model, #432.

• Handle edge case of interpolation at first index correctly, fixes #433

• Instead of storing each dose information sequentially, store dose information at the same index of the evid defining the dose. This memory rewrite is to fix the issue #435.

• Start using strict headers as it is required for the forthcoming release of Rcpp. Thanks to Dirk Eddelbuettel for some of the fixes and alerting us to this change.

• Check arguments for add.dosing() more strictly. See Issue #441

• Issue a warning when either dur() or rate() is in the model but the modeled rate and duration is not included in the event table.

• When the data requires a modeled rate and modeled duration but it is not in the model, warn about the mismatch in data

• Added a back-door for debugging. If you specify options(RxODE.debug=TRUE) then each solve saves the solving information to the file "last-rxode.qs" before actually solving the system.

• Only will try to solve RxODE problems on compatible models; If the model is not supported it will throw an error instead of crashing (See #449)

• Turn off parallel ODE solving whenever the system needs to sort times based on model dosing. Currently this type of solving is not thread safe.

• Update timsort headers to latest version.

RxODE 1.0.9

• At the request of CRAN, stripping the debugging symbols for the CRAN release is no longer performed. This means a larger binary size for RxODE in this release.

• At the request of CRAN the liblsoda code has been changed so that the memory in C defined by _C() is now defined by _rxC(). This will be seen in some of the error messages, which will no longer match the error messages of unmodified liblsoda.

• iCov behavior has shifted to merge on the input event dataset. See Issue #409; This is more in line with expectations of iCov behavior, and reduces the amount of code needed to maintain iCov.

  The iCov in the pipeline is no longer supported because it simply is a merge with the event dataset.

  This can be a breaking change depending on the code you use. Note that clinical trial simulations, resampling is likely better than trying to fill out iCov for every individual which was the prior use.

• Bug fix for crashes with string covariates or factor covariates, issue #410. Also factor column names are compared with case insensitivity just like the rest of the column names for event tables or data sets in RxODE.

RxODE 1.0.8

• Fix issue #399

RxODE 1.0.7

• Change syntax vignette to use markdown option screenshot.force=FALSE. This should get rid of the webshot error

• Change to depend on dparser 1.3.0, which has some memory fixes

RxODE 1.0.6

• RxODE imports but does not link to checkmate any longer. This change should make recompilation of RxODE to work with different releases of checkmate unnecessary.

• Default Solaris solver changed back to "lsoda"

• Fix Bug #393, where in certain circumstances rxSolve(...,theta=) did not solve for all subjects.

• Will not ignore NEWS and README when building the package so that they will show up on CRAN. You can also access the news by news(package="RxODE")

• Changed ODR model names from time id to _rx followed by the md5 hash id and a per-session counter id; For packages the id is _rxp followed by the md5 hash and a per-session counter id.

• Changed qs to be more conservative in hash creation. Add a check hash as well as NOT using altrep stringfish representation.

RxODE 1.0.5

• Maintenance release -- use std::floor and cast variables to double for internal C functions. This should allow a successful compile on Solaris CRAN.

• Changed units from an Imports to a Suggests to allow testing on Solaris rhub

• Changed ODR model names from time id to _rx followed by the md5 hash id; For packages the id is _rxp followed by the md5 hash.

• Removed AD linear compartment solutions for Windows R 3.6, though they still work for Windows R 4.0 (You can get them back for Windows R 3.6 if you install BH 1.66.0-1 and then recompile from source).

  • This will cause nlmixr to fail with solved systems on Windows 3.6. Currently the Stan Headers do not compile on this system so they are disabled at this time.

• RxODE imports but does not link to qs any longer; This change should make recompilation of RxODE to work with different releases of qs unnecessary.

• RxODE now checks for binary compatibility for Rcpp, dparser, checkmate, and PreciseSums

RxODE 1.0.4

Breaking changes

• RxODE can only use supported functions (could be breaking); You may add your own functions with rxFun and their derivatives with rxD

• RxODE now uses its own internal truncated multivariate normal simulations based on the threefry sitmo library. Therefore random numbers generated within RxODE like providing rxSolve(...,omega=) will have different results with this new random number generator. This was done to allow internal re-sampling of sigmas/etas with thread-safe random number generators (calling R through mvnfast or R's simulation engines are not thread safe).

• RxODE now moved the precise sum/product type options for sum() and prod() to rxSolve or rxControl

• cvPost now will returned a named list of matrices if the input matrix was named

• rxSolve will now return an integer id instead of a factor id when id is integer or integerish (as defined by checkmate). Otherwise a factor will be returned.

• When mixing ODEs and linCmt() models, the linCmt() compartments are 1 and possibly 2 instead of right after the last internal ODE. This is more aligned with how PK/PD models are typically defined.

• EVID=3 and EVID=4 now (possibly) reset time as well. This occurs when the input dataset is sorted before solving.

• When EVID=2 is present, an evid column is output to distinguish evid=0 and evid=2

New features

• Add the ability to order input parameters with the param() pseudo-function

• Add the ability to resample covariates with resample=TRUE or resample=c("SEX", "CRCL"). You can resample all the covariates by ID with resampleID=TRUE or resample the covariates without respect to ID with resampleID=FALSE

• Comparison of factors/strings is now supported in RxODE; Therefore ID=="Study-1" is now allowed.

• Completion for elements of rxSolve() objects, and et() objects have been added (accessed through $)

• Completion of rxSolve() arguments are now included since they are part of the main method

• Allow simulation with zero matrices, that provide the simulation without variability. This affects rxSolve as well as rxMvnrnd and cvPost (which will give a zero matrix whenever one is specified)

• et() can dose with length(amt) > 1 as long as the other arguments can create a event table.

• Rstudio notebook output makes more sense

• Printing upgraded to cli 2.0

• Caching of internal C data setup is now supported increasing speed of optim code when:

  • Event Table doesn't change
  • The size of the parameters doesn't change
  • inits do not change (though you can specify them as cmt(0)=... in the model and change them by parameters)
  • See Issue #109

• Allow while(logical) statements with ability to break out if them by break. The while has an escape valve controlled by maxwhere which by default is 10000 iterations. It can be change with rxSolve(..., maxwhere = NNN)

• Allow accessing different time-varying components of an input dataset for each individual with:

  • lag(var, #)
  • lead(var, #)
  • first(var)
  • last(var)
  • diff(var)

Each of these are similar to the R lag, lead, first, last and diff. However when undefined, it returns NA

• Allow sticky left-handed side of the equation; This means for an observation the left handed values are saved for the next observations and then reassigned to the last calculated value.

  This allows NONMEM-style of calculating parameters like tad:

mod1 <-RxODE({
    KA=2.94E-01;
    CL=1.86E+01;
    V2=4.02E+01;
    Q=1.05E+01;
    V3=2.97E+02;
    Kin=1;
    Kout=1;
    EC50=200;
    C2 = centr/V2;
    C3 = peri/V3;
    d/dt(depot) =-KA*depot;
    d/dt(centr) = KA*depot - CL*C2 - Q*C2 + Q*C3;
    d/dt(peri)  =                    Q*C2 - Q*C3;
    d/dt(eff)  = Kin - Kout*(1-C2/(EC50+C2))*eff;
    if (!is.na(amt)){
        tdose <- time
    } else {
        tad <- time - tdose
    }
})

It is still simpler to use:

mod1 <-RxODE({
    KA=2.94E-01;
    CL=1.86E+01;
    V2=4.02E+01;
    Q=1.05E+01;
    V3=2.97E+02;
    Kin=1;
    Kout=1;
    EC50=200;
    C2 = centr/V2;
    C3 = peri/V3;
    d/dt(depot) =-KA*depot;
    d/dt(centr) = KA*depot - CL*C2 - Q*C2 + Q*C3;
    d/dt(peri)  =                    Q*C2 - Q*C3;
    d/dt(eff)  = Kin - Kout*(1-C2/(EC50+C2))*eff;
    tad <- time - tlast
})

If the lhs parameters haven't been defined yet, they are NA

• Now the NONMEM-style newind flag can be used to initialize lhs parameters.

• Added tad(), tad(cmt) functions for time since last dose and time since last dose for a compartment; Also added time after first dose and time after first dose for a compartment tafd(), tafd(cmt); time of last dose tlast(), tlast(cmt) and dose number dosenum() (currently not for each compartment)

• Changed linear solved systems to use "advan" style linCmt() solutions, to allow correct solutions of time-varying covariates values with solved systems; As such, the solutions may be slightly different. Infusions to the depot compartment are now supported.

• Added sensitivity auto-differentiation of linCmt() solutions. This allows sensitivities of linCmt() solutions and enables nlmixr focei to support solved systems.

  • One solution is to use Stan's auto-differentiation which requires C++14

• When calculating the empirical Bayesian estimates for with rxInner (used for nlmixr's 'focei') ignore any variable beginning with rx_ and nlmixr_ to hide internal variables from table output. This also added tad=tad() and dosenum=dosenum() to the ebe output allowing grouping by id, dose number and use TAD for individual plot stratification.

• Added ability to prune branching with rxPrune. This converts if/else or ifelse to single line statements without any if/then branching within them.

• Added ability to take more complex conditional expressions, including:

  • ifelse(expr, yes, no)
  • x = (x==1)*1 + (!(x==1))*2
  • if (logic){ expr} else if (logic) {expr} else {}. The preferred syntax is still only if/else and the corresponding parsed code reflects this preference.
    • Note ifelse is not allowed as an ODE compartment or a variable.

• Switched to symengine instead of using sympy

  • Remove dependence on python.
  • Since symengine is C-based and doesn't require the python interface it is much faster than sympy, though some functions in sympy are no longer accessible.
  • Also symengine requires R 3.6, so now RxODE requires R 3.6

• Added new ODE solving method "indLin", or inductive linearization. When the full model is a linear ODE system this becomes simply the matrix exponential solution. Currently this requires a different setup.

• Added arbitrary function definition to RxODE using rxFun

  • Requires function, arguments and corresponding C-code
  • Derivatives (if required) can be added to the derivative table rxD. When taking deviates without a derivative function, RxODE will use numerical differences.

• Will error if RxODE does not know of a function that you are trying to use; This could be a breaking change. Currently:

  • C's functions from math.h are supported
  • R's function returning and taking doubles are supported
  • Other functions can be added using rxFun and rxD

• Added NA, NaN, Inf and +Inf handling to a RxODE model. Can be useful to diagnose problems in models and provide alternate solutions. In addition, added R-like functions is.nan, is.na, is.finite and is.infinite which can be called within the RxODE block.

• Allowed the following data variables can be accessed (but not assigned or used as a state):

  • cmt
  • dvid
  • addl
  • ss
  • amt
  • rate
  • id which requires calling the id as factor ID=="1" for instance.

• Kept evid and ii as restricted items since they are not part of the covariate table and are restricted in use.

• Added the following random number generators; They are thread safe (based on threefry sitmo and c++11) and your simulations with them will depend on the number of cores used in your simulation (Be careful about reproducibility with large number of threads; Also use parallel-solve type of RxODE simulations to avoid the birthday problem).

  During ODE solving, the values of these are 0, but while calculating the final output the variable is randomized at least for every output. These are:

  • rxnorm() and rxnormV() (low discrepancy normal)
  • rxcauchy()
  • rxchisq()
  • rxexp()
  • rxf()
  • rxgamma()
  • rxbeta()
  • rxgeom()
  • rxpois()
  • rxt()
  • rxunif()
  • rxweibull()

  In addition, while initializing the system, the following values are simulated and retained for each individual:

  • rinorm() and rinormV() (low discrepancy normal)
  • ricauchy()
  • richisq()
  • riexp()
  • rif()
  • rigamma()
  • ribeta()
  • rigeom()
  • ripois()
  • rit()
  • riunif()
  • riweibull()

• Added simeta() which simulates a new eta when called based on the possibly truncated normal omega specified by the original simulation. This simulation occurs at the same time as the ODE is initialized or when an ODE is missing, before calculating the final output values. The omega will reflect whatever study is being simulated.

• Added simeps() which simulates a new eps from the possibly truncated normal sigma at the same time as calculating the final output values. Before this time, the sigma variables are zero.

All these change the solving to single thread by default to make sure the simulation is reproducible. With high loads/difficult problems the random number generator may be on a different thread and give a different number than another computer/try.

Also please note that the clang and gcc compiler use different methods to create the more complex random numbers. Therefore MacOS random numbers will be different than Linux/Windows at this time (with the exception of uniform numbers).

These numbers are still non-correlated random numbers (based on the sitmo test) with the exception of the vandercorput distributions, so if you increase the number of threads (cores=...) the results should still be valid, though maybe harder to reproduce. The faster the random number generation, the more likely these results will be reproduced across platforms.

• Added the ability to integrate standard deviations/errors of omega diagonals and sigma diagonals. This is done by specifying the omega diagonals in the theta matrix and having them represent the variabilities or standard deviations. Then these standard deviations are simulated along with the correlations using the IJK correlation matrix (omega dimension < 10) or a correlation matrix or Inverse Wishart-based correlation matrix (omega dimension > 10). The information about how to simulate this is in the variability simulation vignette.

• Now have a method to use lotri to simulate between occasion variability and other levels of nesting.

• Added lower gamma functions See Issue #185

• Upgraded comparison sort to timsort 2.0.1

• Changed in-place sort to a modified radix sort from data.table. The radix search was modified to:

• Work directly with RxODE internal solved structures
• Assume no infinite values or NA/NaN values of time
• Always sort time in ascending order
• Changed sorting to run in a single thread instead of taking over all the threads like data.table

• Changed method for setting/getting number of threads based on data.table's method

• Added function rxDerived which will calculate derived parameters for 1, 2, and 3 compartment models

• More descriptive errors when types of input are different than expected

Engine changes

• Moved many C functions to C++. CRAN OpenMP support requires C++ only when C and C++ are mixed. See:

  https://stackoverflow.com/questions/54056594/cran-acceptable-way-of-linking-to-openmp-some-c-code-called-from-rcpp

• No longer produces C code that create the model variables. Instead, use qs to serialize, compress and encode in base91 and then write the string into the C file. The qs package then decodes all of that into the model variables. This also increases the compilation speed for models in RxODE.

• Pre-compile RxODE headers once (if cache is enabled), which increases compilation speed for models in RxODE

• RxODE's translation from the mini-language to C has been refactored

Bug fixes:

• Occasionally RxODE misidentified dual lhs/param values. An additional check is performed so that this does not happen.

• For solved matrices with similar names (like "tadd" and "tad") RxODE will now prefer exact matches instead of the first match found when accessing the items with $tad.

• A fix where all ID information is kept with keep=c(""..."")

• Transit compartment models using the transit ODE or variable are now allowed. Also check for more internally parsed items (see Issue #145).

• Bug fix for etSeq and etRep where greater than 2 items were mis-calculated

RxODE v0.9.2-0

• New plotting engine
• Various bug fixes for upcoming R 4.0 release:
  • Dropped some imports for 21 imports restriction
  • Fixed incompatibility with new ggplot2 3.3.0
  • Fixed allowing NAs in RxODE dataset
  • Fixed setting all compartment default values for bioavailability, rate, etc.
  • Added additional protection against floating point -> NaN for power functions

RxODE v0.9.1-9

• Minor namespace/documentation changes for R 4.0 compatibility

RxODE v0.9.1-8

• Added the ability to have an input parameter to be assigned to a new value (Issue #135)
• Added LINPACK authors as contributors
• Added a NEWS.md file to track changes to the package