Skip to content

mecfs_bio.build_system.task.ppp_ldsc.ppp_ldsc_context

Shared, protein-invariant context for LDSC over the UKB-PPP database.

Every per-protein slim file stores beta/se in the SAME variant-index row order, so the alignment of index variants to LD scores, the regression LD scores (ld), the total reference-SNP count M, and the genome-sorted order used for contiguous jackknife blocks are all identical across proteins. We therefore build them ONCE and reuse them for every protein (and, later, for cross-trait rg and LCV).

The context also carries each retained variant's chromosome and hg38 position so a per- protein cis mask (variants within +/- a window of the protein's gene) can be computed, and its row position in the slim files so a protein's beta/se can be gathered by positional index.

Built by intersecting the variant index (rsID) with the reference LD scores (SNP), then optionally dropping strand-ambiguous variants and the extended MHC region (on the index's primary hg38 POS), and sorting by (chromosome, position).

The reference LD scores arrive as a consolidated dataframe (ConsolidateLDScoresTask), which also carries the common-variant count M_5_50 the regression normalizes by, so this module reads dataframes rather than the LDSC authors' per-chromosome file layout.

Classes:

  • PppLdscContext –

    The shared regression SNP set, genome-sorted. All arrays are parallel (one entry per

Functions:

  • build_cis_mask –

    Boolean over the context SNP set, True for CIS variants (to be excluded): those on

  • build_ppp_ldsc_context –

    Intersect the variant index with the LD scores and assemble the shared context.

PppLdscContext

The shared regression SNP set, genome-sorted. All arrays are parallel (one entry per retained variant, in (chromosome, position) order).

We do not carry a separate regression-weight LD score (wLD): with sep_weights=False the weight LD score equals ld, and the batched kernel reuses ld directly. Reinstate a wld array here only if a future path (e.g. externally supplied regression weights) needs it.

Attributes:

chrom instance-attribute 

chrom: ndarray

ld instance-attribute 

ld: ndarray

m instance-attribute 

m: float

n_snps property 

n_snps: int

pos instance-attribute 

pos: ndarray

row_pos instance-attribute 

row_pos: ndarray

build_cis_mask 

build_cis_mask(
    context: PppLdscContext,
    gene_chrom: int,
    gene_start: int,
    gene_end: int,
    window_bp: int,
) -> np.ndarray

Boolean over the context SNP set, True for CIS variants (to be excluded): those on the gene's chromosome within window_bp of the gene body [start, end].

Source code in mecfs_bio/build_system/task/ppp_ldsc/ppp_ldsc_context.py 
137
138
139
140
141
142
143
144
145
146
147
148
149
150
def build_cis_mask(
    context: PppLdscContext,
    gene_chrom: int,
    gene_start: int,
    gene_end: int,
    window_bp: int,
) -> np.ndarray:
    """Boolean over the context SNP set, True for CIS variants (to be excluded): those on
    the gene's chromosome within window_bp of the gene body [start, end]."""
    return (
        (context.chrom == gene_chrom)
        & (context.pos >= gene_start - window_bp)
        & (context.pos <= gene_end + window_bp)
    )

build_ppp_ldsc_context 

build_ppp_ldsc_context(
    index_df: DataFrame,
    ld_df: DataFrame,
    *,
    drop_strand_ambiguous: bool = True,
    exclude_mhc: bool = True,
) -> PppLdscContext

Intersect the variant index with the LD scores and assemble the shared context.

index_df: the variant index, with columns CHR, POS (hg38), rsID, and is_strand_ambiguous, in its canonical row order (row i aligns to row i of every slim file). ld_df: consolidated reference LD scores with columns CHR, SNP (rsID), L2 and M_5_50.

M is taken from the LD scores BEFORE the join, since it counts reference variants and must not shrink to the subset the index happens to cover.

Source code in mecfs_bio/build_system/task/ppp_ldsc/ppp_ldsc_context.py 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
def build_ppp_ldsc_context(
    index_df: pl.DataFrame,
    ld_df: pl.DataFrame,
    *,
    drop_strand_ambiguous: bool = True,
    exclude_mhc: bool = True,
) -> PppLdscContext:
    """Intersect the variant index with the LD scores and assemble the shared context.

    index_df: the variant index, with columns CHR, POS (hg38), rsID, and
        is_strand_ambiguous, in its canonical row order (row i aligns to row i of every
        slim file).
    ld_df: consolidated reference LD scores with columns CHR, SNP (rsID), L2 and M_5_50.

    M is taken from the LD scores BEFORE the join, since it counts reference variants and must
    not shrink to the subset the index happens to cover.
    """
    m_total = total_m_5_50(narwhals.from_native(ld_df, eager_only=True))
    joined = (
        index_df.with_row_index(_ROW_POS_COL)
        .join(
            ld_df.select(LD_SCORE_RSID_COL, LD_SCORE_LD_SCORE_COL),
            left_on=GWASLAB_RSID_COL,
            right_on=LD_SCORE_RSID_COL,
            how="inner",
        )
        .sort([GWASLAB_CHROM_COL, GWASLAB_POS_COL])
    )
    if drop_strand_ambiguous:
        joined = joined.filter(~pl.col(PPP_INDEX_IS_STRAND_AMBIGUOUS_COL))
    if exclude_mhc:
        mhc = extended_mhc_interval("38")
        in_mhc = (
            (pl.col(GWASLAB_CHROM_COL) == mhc.chrom)
            & (pl.col(GWASLAB_POS_COL) >= mhc.start)
            & (pl.col(GWASLAB_POS_COL) <= mhc.end)
        )
        joined = joined.filter(~in_mhc)

    return PppLdscContext(
        row_pos=joined[_ROW_POS_COL].to_numpy().astype(np.int64),
        ld=joined[LD_SCORE_LD_SCORE_COL].to_numpy().astype(float),
        chrom=joined[GWASLAB_CHROM_COL].to_numpy().astype(np.int64),
        pos=joined[GWASLAB_POS_COL].to_numpy().astype(np.int64),
        m=m_total,
    )