fix: build the database when a uniprot_N column of complex_input.csv is unused

[cindykrafft]

Fixes #224.

Cause. db_utils.create_db runs sanity_test_report_unknown_proteins, which checks that every subunit accession in complex_input.csv exists in protein_input.csv by merging each uniprot_N column against the uniprot column (cellphonedb/utils/db_utils.py:603). When no complex in the file fills one of the uniprot_N columns present in the header, pandas reads that column as an all-NaN float64 column, and the merge against the string column raises

ValueError: You are trying to merge on float64 and object columns for key 'uniprot_3'. If you wish to proceed you should use pd.concat

(float64 and str under pandas 3). That is the normal shape of a custom database built the way the T0_BuildDBfromFiles notebook describes: the released complex_input.csv header has uniprot_1..uniprot_5, only one released complex uses uniprot_5 and four use uniprot_4, so any subset that drops those rows fails to build. Converting the columns to strings before saving the CSV does not help, because create_db re-reads the files. #137 (2023, closed) shows the same traceback from the same merge for a custom database with an essentially empty complex_input.csv.

Fix. The sanity test now takes the non-empty accessions of each subunit column (dropna()) and reports those absent from protein_input.csv. That is exactly what the outer merge computed (uniprot null and uniprot_N not null), without depending on the column dtype. The warning text and content are unchanged: on the released v5.0.0 input files, master and this branch print identical output, both as-is (no warning) and with two proteins removed from protein_input.csv (both are reported by both versions). No result of any analysis method is affected.

Test. CreateDbUnitTests in method_tests.py (a new TestCase with its own setUp, so it needs no database download and runs in CI as-is):

• test_create_db_with_unused_subunit_columns: create_db on hand-written input files with the standard four-column header and a single dimer (uniprot_3/uniprot_4 empty), checking the zip is produced with the expected complex and its two-row composition — fails on master with the ValueError above;
• test_unknown_complex_proteins_still_reported: the sanity test on a frame with an all-NaN float64 column and one unknown accession still names that accession and nothing else — also fails on master.

Run. Python 3.12, pandas 2.3.3: the two new tests fail on unmodified master and pass with the change; flake8 . --count --select=E9,F63,F7,F82 reports 0 and flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 reports 0 before and after. The rest of method_tests.py is unchanged by this PR (the four tests that read ../../../example_data/test.h5ad fail on master in this checkout for the pre-existing path reason that #210 addresses).

Found while working through open issues in Mytochondria, a volunteer project that verifies fixes for the software behind published results (methods and harnesses: https://github.com/cindykrafft/mytochondria/tree/main/audits/cellphonedb)

---

Generated by Claude Code