Changelog • glysmith

glysmith (development version)

Implemented single-question flow in inquire_blueprint() and modify_blueprint() for a more streamlined user experience.
Q&A context is now preserved when modifying blueprints with modify_blueprint().
Improved error messages for steps in branches with missing dependencies.
Fun fact now displays only once at the start of AI interactions.
Enhanced LLM prompting with atomic question guidance for better clarification questions.
Changed default model from deepseek-reasoner to deepseek-chat.

step_preprocess() now does not generate QC plots anymore. The QC plots are now generated by step_plot_qc().

Add step_plot_qc() for generating Quality Control plots.
Add step_cox() for Cox proportional hazards survival analysis.
Add step_correlation() for correlation analysis.
Add step_sig_boxplot() for generating boxplots of significant variables.
Add step_plsda() and step_oplsda() for supervised dimensionality reduction.
inquire_blueprint() now includes sample_info and var_info summaries in LLM context for better suggestions.
modify_blueprint() now has clarification question capability.

Improved LLM prompt for vague inputs and minimal blueprints in inquire_blueprint().
Added Hmisc to Suggests for step_correlation() dependency.
inquire_blueprint() UI has been optimized.
Optimize the speed of inquire_blueprint() and modify_blueprint().
AI prompts of some steps have been updated.
Per-plot width and height are now customizable in step functions.

Add step_adjust_protein() for adjusting glycoform quantification by protein expression.
Add step_logo() for generating logo plots.
inquire_blueprint() now asks clarifying questions to better understand user requirements.
inquire_blueprint() now asks the user to confirm the blueprint before returning it.

inquire_blueprint() has a better UI/UX.
Updated the system prompt of inquire_blueprint() to improve its understanding of step usage.
Fixed a bug where inquire_blueprint() could not access step parameters.
Fixed a bug on Windows where file paths generated by the LLM could not be parsed.

LLM-powered report generation is updated largely.
- The report structure (sections to include, orders of figures and text chunks, etc.) is now determined by the LLM.
- Titles, subtitles, figure titles are now polished by the LLM.
- A text description of each figure is now generated by the LLM using multimodal reasoning.
step_preprocess() now also generates Quality Control (QC) plots.
Add cast_data() to fetch data from glysmith_result object.
When the name parameter is not specified in cast_plot(), cast_table(), or cast_data(), the functions return the names of all available objects. This makes it easier to know what objects are available.
Add modify_blueprint() to modify a blueprint using natural language.

inquire_blueprint() now prints a brief description of the blueprint to console, instead of an explanation of why the steps are chosen. There is a slight difference in the wording.
Optimize the prompt of inquire_blueprint() to let it group steps using br(). Also, it now understands all step arguments and can use them if needed.
“Tables” section is removed from the report. It is duplicated with the README.md file generated by quench_result().
Add a new vignette about AI features.

The dynamic argument mechanism was removed. Now you can directly pass arguments to the step functions.

Added br() for creating branches in a blueprint. inquire_blueprint() also supports blueprints with branches now.
step_dea_limma() now support covariates and paired comparison, via the covariate_cols and subject_col arguments.

Fix the issue that blank lines were printed to console when the package is first loaded in one session.
Update system prompt and the reflection strategy in inquire_blueprint() to make the LLM work better.
Fix a dependency issue of step_dea_kruskal().
Update the Get Started vignette to introduce br().

Added step_quantify_motifs() for quantifying glycan motifs. Updated relevant steps to support the on parameter for motif_exp experiments.
inquire_blueprint() now prints an LLM-generated explanation to the console before returning the blueprint.
inquire_blueprint() now supports an exp argument, accepting a glyexp::experiment() object, to provide more accurate, context-aware suggestions.

Refined step documentation to improve clarity and optimize inquire_blueprint() performance.
Enhanced error messages for failed steps.
Suppressed a harmless warning in quench_result() when used with step_pca().
Console messages from step_pca() now include the experiment name during forge_analysis().
Moved several dependencies (pROC, Rtsne, uwot, EnhancedVolcano, org.Hs.eg.db, clusterProfiler, ggplotify, pheatmap, and factoextra) to “Suggests”. These are now automatically checked and installed when required by blueprint steps.

Add inquire_blueprint() use a Large Language Model to generate a blueprint from user prompt.
Add step_tsne() for t-distributed Stochastic Neighbor Embedding.
Add step_umap() for Uniform Manifold Approximation and Projection.
Add step_roc() for Receiver Operating Characteristic analysis.
step_pca() now also generates loading plot and screeplot, besides the individual plot.
step_pca() (along with the newly added step_tsne() and step_umap()) now has an on parameter to control which experiment to perform the analysis on.

Update report content of step_ident_overview(). This is to adapt to the new glyexp::summarize_experiment() function introduced in glyexp 0.11.0.
Update blueprint_default(): step_ident_overview() is now the first step, before step_preprocess().
Fix the bug that report content of DEA step for derived traits has “NA vs NA” contrast.
Some plot names, table names, or step IDs have been changed to be more consistent.
Update documentations of some step functions to better reflect their usage.
glysmith_result now has a data field to store the data generated by the blueprint.