FastRet

• Added semi-transparent background to legends for better readability
• Changed the point character from unfilled circles with colored borders to filled circles with black borders for better visibility.
• Constant predictions (causing correlation to be NA) do no longer cause a warning.

3. preprocess_data():

• Added argument add_cds to control whether chemical descriptors should be added to the input data using getCDs().
• Added argument rm_ucs to control whether unsupported columns (i.e. columns that are neither mandatory nor optional) should be removed from the input data.
• Added argument rt_terms to control whether transformations of the RT column (square, cube, log, exp, sqrt) should be added to the input data.
• In case of missing mandatory columns (SMILES, RT, NAME) an error is now raised.
• INCHIKEY and Chemical Descriptors listed in CDFeatures are allowed as optional columns.
• Columns that are neither mandatory nor optional are automatically removed.
• Improved runtime for generation of polynomial features and/or interaction terms.
• Removal of NA and/or near-zero-variance predictors is now done after adding polynomial features and/or interaction terms.
• Improved progress output.

4. train_frm():

• Added argument do_cv to control whether cross-validation should be performed for performance estimation. Default is TRUE.
• Removal of near-zero-variance predictors and/or removal of NA values is now done as part of the internal model training, i.e. it happens separately for each fold during cross-validation. This prevents data leakage from the training set to the validation set. The corresponding hint about "overoptimistic cross-validation results" has consequently been removed from the documentation.
• Argument method now accepts two values for training models with xgbtree base: "gbtreeDefault" (train xgboost with default params) and "gbtreeRP" (train xgboost with parameters optimized for the RP dataset). The old value "gbtree" still works and is now an alias for "gbtreeDefault".
• Improved documentation of the return value (i.e. frm objects are fully specified now).
• Added type checking for each user input.
• Performance estimation via cross-validation now uses the new clipping mechanism provided by clip_predictions(). Of course, the clipping is always based on the RT range of training folds, not the whole original training data.
• Calculating proportion of variance explained (R²) no longer throws a warning for constant predictions (causing correlation to be NA). Instead, R² is set to 0 in such cases.

5. predict.frm():

• Data transformations applied to the training data (adding polynomial features and/or adding interaction terms) are now automatically applied to new data as well. This was not the case before, leading to errors when trying to predict RTs using models trained with degree_polynomial>1 and/or interaction_terms=TRUE, unless the transformations were manually applied to the new data beforehand.
• Added argument clip to allow clipping of predictions to be within the RT range of the training data. Works for both adjusted and unadjusted models.
• Predictions are now clipped to be within a sensible range by default. To produce unclipped predictions, set clip=FALSE. See clip_predictions() for details.
• If a chemical descriptor is NA in the new data, but required by the model, the NA value is now replaced by the mean of that descriptor in the training data. Previously, the prediction for such entries was NA. The old behavior can be restored by setting impute=FALSE in predict.frm().

6. selective_measuring():

• Added argument rt_coef, allowing user to control the influence of RT on the clustering. A value of 0 means that RT is ignored, a value of "max_ridge_coefficient" means that RT has the same weight as the most important chemical descriptor and a value of 1 means no scaling at all (except standardization to z-scores, which is applied before to the whole dataset before the ridge regression is trained).

7. adjust_frm():

• Added argument seed to allow reproducible results.
• Added argument do_cv to control whether cross-validation should be performed for performance estimation. Default is TRUE.
• Added argument adj_type to control which model should be trained for adjustment: supported options are "lm", "lasso", "ridge", or "gbtree". Previously, only "lm" was supported. To stay backwards compatible, the default is "lm".
• Added argument add_cds to control whether chemical descriptors should be added to the input data using getCDs(). Only recommended for adj_type other than "lm".
• Added support for mapping by SMILES+INCHIKEY in addition to SMILES+NAME. SMILES+INCHIKEY is used by default if both columns are present in the adjusted and the original training data. Otherwise SMILES+NAME is used as before.
• Improved error handling. Previously, unmappable entries in the new data had been ignored silently. Now, an error is raised in such cases.
• Function arguments are now stored in the returned frm object for better reproducibility.
• Mapping is now performed by matching each new entry to the average RT of all original training entries with the same key (SMILES+INCHIKEY or SMILES+NAME). Example: if the new dataset contains a key twice, and the original training data contains the key three times, both new entries are mapped to the average RT of the three original entries.
• Performance estimation via cross-validation now uses the new clipping mechanism provided by clip_predictions(). Of course, the clipping is always based on the RT range of training folds, not the whole original training data.
• Calculating proportion of variance explained (R²) no longer throws a warning for constant predictions (causing correlation to be NA). Instead, R² is set to 0 in such cases.

8. print.frm():

• frm objects can now be printed directly to the console in a user-friendly format.

9. clip_predictions():

• New utility function for clipping predicted RTs to be within a sensible range. Used internally by train_frm(), predict.frm() and adjust_frm().

10. get_predictors():

• Added arguments base and adjust to control whether predictors for the base model, the adjustment model or both should be returned.

Bugfixes:

1. Interaction terms generated by preprocess_data() are now generated

correctly as product of the involved features instead of a division. This

follows common practice in regression modeling and avoids division by zero

issues. Passing older models, trained with division-based interaction terms,

to downstream functions like predict.frm() or adjust_frm() will now lead

to an error. (This is not a breaking change, as predict.frm() and friends

have in fact never been able to handle such models).

2. plot_frm() with type "scatter.cv.adj" or "scatter.train.adj" now correctly

shows retention times from the new data (used for model adjustment) as x-axis

values instead of the original training retention times.

3. catf() now only emits escape codes (i.e. colored output), it the output is

directed to a terminal. If the output is redirected to a file or a pipe, no

escape codes are emitted anymore. Since catf() is used throughout the

package for logging, this fixes the output for the whole package.

Internal Improvements:

1. Added or improved unit tests for:

• adjust_frm()
• fit_gbtree()
• fit_glmnet()
• get_param_grid()
• get_predictors()
• getCDs()
• plot_frm()
• predict_frm()
• preprocess_data()
• selective_measuring()
• train_frm()
• validate_inputdata()

2. Removed caret dependency by adding custom implementations for:

• createFolds()
• nearZeroVar()

3. Extract mapping and merging part of adjust_frm() into a private function

merge_dfs().

4. Replaced fit_glmnet(), fit_lasso() and fit_ridge() with a single

function fit_glmnet(), that takes the method ("lasso" or "ridge") as

parameter. Instead of a dataframe df that has to contain only predictors

plus the RT column (as reponse), the function now takes a matrix of

predictors X and a vector of responses y. This makes the function more

flexible and easier to test.

5. Replaced fit_gbtree_grid() with a much simpler function

find_params_best(). Instead of allowing the specification of every grid

parameter, the new function instead accepts a keyword searchspace for

specifying predefined grids to choose from.

6. Improved fit_gbtree by exposing lots of hardcoded internal xgboost

parameters as function parameters with sensible defaults. In particular, the

user can now set xpar to "default", "rpopt" or a predefined grid-size to

train the model with different hyperparameter settings. Furthermore, the

function is now written in a way that works with both, version 1.7.9.1 and

the new 3.1.2.1 version published on 2025/12/03 (yes, version 2.x was skipped

completely).

7. Added helper function get_param_grid() for returning predefined

hyperparameter grids for xgboost model training based on keywords like

"tiny", "small" or "large".

8. Added function benchmark_find_params() to benchmark runtime of

find_params_best() for different numbers of cores and/or threads. As it

turns out, choosing a higher number of cores is usually more efficient (at

the cost of worse progress output).

9. Added utility functions named(), as_str(), is_valid_smiles() and

as_canonical()