Process output files

All four guides use the same Python entry point, porefm.outputs.open_output. It reads saved predictions without loading model weights or using a GPU. Use a completed run directory and keep data/meta pairs and manifests together.

completed run directory
        |
        v
   open_output
        |
        +--> iter_tables ------> preview/filter individual read observations
        |
        `--> summarize_loci ---> one row per chromosome, position, strand
                                  |
                                  `--> summary Parquet for plots/analysis

Preview a few observations

from porefm.outputs import open_output

calls = open_output("outputs/zero-shot")
print(calls.format)
print(calls.schema.names)
for table in calls.iter_tables(batch_size=5):
    print(table.to_pandas())
    break

Change the path to outputs/supervised, outputs/pairwise, or outputs/residual to inspect another guide's result. Opening a pairwise run selects read_scores/; its companion site tables are read separately.

Choose the right site summary

Workflow	What the summary measures	Example
Supervised	Hard class counts/fractions, optionally mean class probabilities	Supervised summary
Zero-shot	Fraction of valid calls with a small canonical-null p-value	Zero-shot summary
Pairwise	Fraction of valid test calls that deviate from the control	Pairwise summary
Residual prototypes	Hard class counts, including within-family fractions	Residual summary

The snippets in those guides write directly to a summary Parquet instead of loading all read observations into RAM. Use a new summary filename for each analysis. The default memory budget is 8 GB; memory_limit and temp_directory can be set for larger runs.

To preview a saved summary:

import pyarrow.parquet as pq

summary = pq.ParquetFile("summaries/zero-shot.parquet")
for batch in summary.iter_batches(batch_size=5):
    print(batch.to_pandas())
    break

Coordinates and denominators

Reference position is zero-based. strand is 0 for forward, 1 for reverse. read_pos follows the producer's oriented read sequence; it is not a genomic coordinate. Site summaries exclude observations without valid reference coordinates. They summarize retained predictions, so coverage can be lower than BAM depth after input filtering, context checks, or edge trimming.

Filter by region when needed, but retain the calls that form the denominator:

sites = calls.summarize_loci(
    filters=[("chrom", "==", "chr3"),
             ("position", ">=", 1_000_000),
             ("position", "<", 1_010_000)],
    modified_pvalue_cutoff=0.01,
)
print(sites.head())

This small-region example applies to the zero-shot run opened above. For anomaly fractions, retain both significant and nonsignificant p-values. For class fractions, retain canonical and modified calls. Filtering to positive calls first makes the resulting fraction misleading.

The name fraction_modified in anomaly summaries means the fraction passing the chosen p-value cutoff, not a calibrated biological modification fraction. Likewise, effect scores and supervised probabilities have different meanings and should not be compared as if they were the same quantity.

For selected columns, read filters, DuckDB queries, and detailed schemas, see the output API reference.