The Same VCF Genotype Call Can Hide Different Likelihoods

Start with what the selected consumer needs

An export described as preserving every genotype call may satisfy a consumer that uses those calls as categorical inputs. The same promise leaves a likelihood-sensitive consumer's requirement unresolved. The relevant question is which distinctions the receiving input must retain, rather than whether both offers use the VCF name.

This explanation uses VCFv4.5 and BCFv2.2, printing e821e4f dated February 25, 2026. In an ordinary sequence-allele record, GT uses 0 for REF and positive indices for the ALT list. Thus diploid, unphased GT 0/0 calls the reference allele on both copies; it contains no vector of alternatives.

Suppose a consumer needs to distinguish a reference call with little relative likelihood for a heterozygote from a reference call with much more. Sending 0/0 for both gives that consumer identical call inputs. It cannot reconstruct the omitted distinction from that label. This is an information requirement, even before anyone evaluates a model.

References: [VCF]

Decode PL as a contrast between genotype states

VCF defines GL as log10-scaled genotype likelihoods. PL carries their Phred-scaled, nearest-integer representation, with the same genotype ordering. For a diploid record with one ALT allele, that order is 0/0, 0/1, 1/1.

For genotype g compared with genotype r, the likelihood ratio represented by the stored scale is 10^(-(PL_g - PL_r)/10). A difference of 20 therefore represents a ratio of 0.01. Adding the same offset to every entry leaves those differences unchanged; a comparison of contrasts does not require a universal minimum-zero rule.

HTSJDK 4.3.0 corroborates the conversion: GenotypeLikelihoods converts integer PL entries to log10 values by division by -10.0. Its reverse conversion subtracts the largest log10 value before scaling and rounding. That is a named implementation's behavior, rather than a claim about every producer.

Decoding the integers recovers their represented contrasts. It cannot undo rounding to recover the producer's original unrounded likelihoods, so numerical precision must remain tied to the supplied scale.

References: [VCF][HTSJDK]

VCF-L2: identical calls, different likelihood profiles

Stipulate two fictional exports for sample VCF_L2_SAMPLE on reference VCF_L2_REFERENCE, contig SYNTH1 at POS 101, with REF A and the single ALT G. Both records are explicitly diploid and unphased, and both supply GT 0/0. Export A supplies ordinary PL [0,20,40]; export B supplies PL [0,3,30]. The zero best entry is an explicit normalization premise of this example. These are invented profiles, not a variant-caller run or a measured specimen.

Using the stated genotype order and taking 0/0 as the relative baseline, A represents likelihoods proportional to [1,0.01,0.0001]. B represents [1,10^(-0.3),0.001], or approximately [1,0.501187,0.001]. The baseline 1 is a relative scale, not certainty that the reference call is correct.

In A, the heterozygote-to-reference likelihood ratio is 0.01. In B, it is approximately 0.501187. Dividing B's ratio by A's gives approximately 50.1187, or 50.12 times. The comparison concerns two within-export likelihood contrasts; it does not compare absolute data likelihoods across producers.

An export that retains only GT maps both profiles to 0/0. A recipient seeing that call cannot determine whether A's vector, B's vector or another compatible profile was supplied. The lost values are not encoded somewhere inside the call.

Both stipulated profiles give their largest represented likelihood to 0/0. That makes the example easy to follow without asserting that every pipeline selects GT by finding the smallest PL. How a particular producer made its call is a separate part of that producer's definition.

References: [VCF][HTSJDK]

Keep likelihood and posterior probability distinct

A likelihood asks how probable the supplied data are under a specified genotype model: P(data | genotype). A posterior asks how probable that genotype is given the data: P(genotype | data). VCF's naming conventions distinguish those directions, and its GP and PP fields concern posterior probabilities.

For a specified probabilistic model, posterior odds equal the likelihood ratio multiplied by the prior odds. Consequently, a likelihood ratio alone does not identify the posterior odds when the prior information is unspecified. Dividing the three relative likelihoods by their sum would introduce assumptions about the genotype model and priors; it would not reveal a posterior already present in GT.

B's value near 0.501187 is therefore a likelihood ratio against 0/0, not a 50.1187% probability that the call is wrong. Nor does the 50.12-fold comparison establish a corresponding change in error probability. The example supplies no calibration study, biological truth comparison or downstream AI accuracy result.

If the consumer actually requires posterior probabilities, obtaining PL instead of GT resolves only part of the input question. The offer must identify the required quantity and its model context, rather than quietly relabeling likelihoods as probabilities.

References: [VCF][BAYES]

Keep each value attached to its genotype state

A vector without its state labels is incomplete input. Preserve the association with the sample, reference identity, contig and position, REF, ordered ALT alleles and ploidy. The header's field definition matters too: PL has Number=G and integer values, with one entry per genotype state.

Adding a second ALT allele changes the ordinary diploid order to 0/0, 0/1, 1/1, 0/2, 1/2, 2/2. With REF A and ALT G,T, those states mean A/A, A/G, G/G, A/T, G/T, T/T. Six entries are required to express that full state set. Reordering ALT without a corresponding state mapping changes what the values describe.

That small extension explains why a promise to preserve the first three likelihoods is insufficient for an arbitrary record. The review must establish the actual allele and ploidy scope. This article covers ordinary full-allele GL/PL representations; local-allele LPL mappings and symbolic records require their own interpretation.

Missing likelihood information must also remain distinguishable from a supplied value. In particular, filling an absent entry with PL 0 would manufacture a represented likelihood. Neither a complete GT nor a vector-shaped container establishes that all required state values survived.

References: [VCF]

Compare exports against the retained contrast

Return to VCF-L2 and stipulate a consumer whose required feature is the 0/1-to-0/0 likelihood ratio for each identified sample and locus. An offer returning only the calls produces the same input for the two profiles. It cannot substantiate preservation of that feature.

Another offer retains the ordered PL values and their record context, and identifies how the receiver interprets them. If the delivered input supports that description, it preserves the contrasts represented on the stored integer scale. Evidence about a proposal must still be connected to what reaches the consumer.

A derived export could also satisfy the stated requirement. For example, an explicitly identified log10 ratio is -2 for profile A and -0.3 for B on this stored scale. Those numbers preserve this particular contrast even though the export is no longer the literal PL array. Its sample/locus association, state definition and numerical precision remain necessary.

That reduced representation would leave the 1/1 contrast unanswered. It is adequate only for the narrower requirement just stipulated. If the consumer uses the complete vector, the provider must account for every required state, including any omitted, rounded or capped values. The acceptance argument follows the consumer's actual input definition.

If the selected consumer instead needs only GT and its identifying context, retaining a likelihood vector is not automatically an additional requirement. Equally, preserving more fields does not demonstrate better model performance. The linked FASTQ and SAM articles address separate base-quality and alignment-coordinate decisions that should retain their own definitions.

Scope a separate physical installation need

If the chosen compute and storage equipment also needs physical installation, Leviathan's Rack-and-Stack & Colocation Installation service is a route to discuss agreed mounting, IT connections, labels and handoff records. Use the related service and project inquiry with a nonconfidential equipment, rack, site and connection outline to establish scope and availability.

That inquiry concerns the equipment installation. Bioinformatics, genotype interpretation, data preparation and AI input acceptance remain with the appointed data and software specialists.

References: [L1]