3 Additional functionalities of seqme

3.1 Fold functionality

Metrics, such as Improved Precision and Improved Recall (Kynkäänniemi et al. 2019), can become biased when there is a discrepancy in the number of designed sequences and reference sequences. Furthermore, for any evaluated metric, it is often desirable to estimate its standard deviation across multiple subsets of the designed sequences. To address these matters, we introduce Fold, which splits the sequences into K groups of equal size and computes the metric K times. This approach mitigates metric-specific sample-size biases, e.g. FBD, reduces the number of discarded sequences, and enables the estimation of standard error.

3.2 Support of iterative designs

Recall, seqme enables evaluation of a given tuple of sequences using a set of metrics of interest. This is convenient for one-shot design, where the given tuple of sequences is evaluated only once. On top of that, seqme also supports iterative evaluation workflows. This functionality enables the same metrics to be applied iteratively, e.g. to evaluate design across training steps of generative models, or throughout the iterative sequence refinement process in genetic algorithms and Bayesian optimization.

3.3 Visualizations

seqme provides several convenient functions for visualizing metric results. It can display a customizable table of metrics for all evaluated sequence groups and plot a parallel-coordinates chart, where each axis corresponds to one metric. For inspecting a single metric, seqme offers a barplot with optional error or deviation bars. For iterative sequence design, it provides functionality to plot metric trajectories across design iterations for multiple sequence designs. These visualizations make sequence-design performance easy to interpret and help reveal potential failure modes.