Methods

Four kinds of information

Genemap holds four kinds of information about an animal and keeps them apart, so the evaluation can tell them apart: what it inherited, what state it is in, what season it was in, and what each trait is worth where it lives.

The evaluation Live

Genemap solves a single-step genomic BLUP: pedigree relationships and genomic relationships blended into one H matrix, solved by preconditioned conjugate gradient, with the genomic part built from your SNPs after allele-frequency centring.

Traits are coupled through their genetic correlations and solved trait by trait in an outer loop, so a trait with few records borrows from the ones it is correlated with. It is not one joint solve of every trait at once, and we do not call it that.

Repeated records stay as records. An animal weighed six times contributes six observations with their contemporary groups, and the repeatability model uses them as such.

Pedigree runs to five generations. Where an ancestor exists in the pedigree but has no animal record, it enters the relationship matrix as a ghost so its descendants stay related. On the reference herd, 134 of 247 animals link through 378 such ancestors.

Composites and crossbreds can be evaluated on a breed-adjusted relationship matrix, centred within each breed, so a between-breed relationship reads as zero instead of as a spurious negative. It is a per-run option and the run records which matrix it used.

On request, Genemap can estimate a heritability from your data for a single-trait run, using a REML estimator that passed a recovery test on simulated data. Genetic correlations come from the published literature. Every run reports its basis per trait: estimated by the gated estimator, or supplied from the literature, and which.

How a run happens

• An evaluation runs when you ask for one (Advanced › Run an evaluation). Its results are written to your animals only when you choose to apply them.
• DNA results: sheep files can be uploaded directly; for cattle, send the lab's file to Genemap and it is loaded for you.
• Evaluations currently run up to about 1,190 animals in the pedigree, ancestors included. Larger populations need the large-population path, which is not switched on (see below).
• Your published breed-society figures are kept exactly as supplied, beside Genemap's own evaluation, and each figure on a page says which source it is.

The recovery gate, as measured

Simulated populations with known parameters, 20 replicates per cell, criteria fixed before the numbers were read. A cell passes only if the estimate is unbiased, its reported standard error is honest, every replicate converged, none sits on a boundary, and the replicates agree with one another.

Genemap therefore uses no genetic correlation estimated on fewer than 2,000 animals recorded on both traits, and no heritability or repeatability from fewer than 500. Below the floor a run uses the published parameter and says so per trait.

Crossbreds and major genes

Heterosis Partly live

The ranking pages credit crossbred and composite animals for heterosis from a published table, with a source for each figure. Inside the evaluation, heterosis is estimated as a Pollak-Quaas covariate for each trait where breed composition varies enough to estimate it, and reported as not estimable where it doesn't; maternal heterosis is fitted for sheep only. Those per-trait estimates are stored with the run but not yet shown on a page.

Major genes Sheep only, no data yet

For sheep, known major genes (Booroola FecB, FecX, FecG, FecL and Texel myostatin) are taken out before breeding values are solved, once at least five animals are genotyped for the gene. There is no cattle path yet, and no animal on the platform has major-gene results loaded today. The genotypes page shows a reference catalogue of 42 sheep variants and 26 cattle loci with their sources; that catalogue is for reading, not fitted.

The economic engine Checked daily

A national index prices every trait once, against an average enterprise, and hands the same weights to everyone. Genemap prices each trait for your enterprise, and shows every input behind every weight.

How the weights are made

The engine builds a steady-state model of your breeding herd or flock from your farm profile, solves a least-cost ration for each class of animal and reads the shadow prices off that ration, so a kilogram of extra mature cow weight is charged what it costs to feed on your feed base. It prices carcase traits against the grid one threshold at a time, weighted by how much of your drop sits at each threshold, and values fertility and survival by how far an improvement flows through the herd over the years, discounted back to today. Run on the same inputs it returns the same weights, and each set is stored with a fingerprint of the inputs that made it.

Market prices are read each night from the indicators Meat & Livestock Australia publishes. Every snapshot records the source, the time it was read and the sale date of each number. If an indicator can't be read, the read is retried; if it still fails, the last good value is carried forward, marked as such with its sale date, and nothing is estimated.

The engine prices a trait where the link from the breeding value to the farm's money runs inside the model. For cattle that is carcase weight, store weight, mature cow weight, P8 fat against the grid, days to calving through the length of your joining, heifer pregnancy, stayability as a provisional range, and net feed intake at what your finishing ration costs. For sheep it is sale weight, ewe weight, dressing percentage and lambs born and weaned. Growth is priced once: as a heavier carcase if you sell at a fixed age, or as fewer days on feed if you sell at a fixed weight. Marbling, eye muscle, yield and fat are priced when your grid pays on them and you have entered the grid. Calving and lambing ease, and milk in cattle, are priced when you enter the figures only you can supply. Where a trait has no channel, or pricing it would count the same money twice, it gets no weight, the row says why, and the ranking keeps its standard weight for that trait.

Today, with reference values standing in for what a farm hasn't entered, the engine puts a dollar value on 9 of 21 cattle traits and 6 of 24 sheep traits. Filling in your farm profile adds more. Anything left blank runs on a labelled reference value, and the rank page says, trait by trait, which weights rest on a reference value.

You can change your prices, rates and around 60 engine settings, and the index recalculates. Traits with an optimum in the middle, like mature cow weight and birth weight, are priced on a curve rather than a straight line; an animal whose growth potential exceeds what the pasture can feed is discounted for nutrition fit; and the breed class is shifted by the property's heat load. An AI-proposed set of weights is kept beside the engine's as a labelled comparison only.

A daily check confirms that the night's market snapshot is fresh and that every organisation with animals has a current set of weights, repairs a missing set by running the engine, and writes the result to a ledger. The tag on this section reads Live only when that check has passed within the last day and a half.

A worked example Illustrative

Breeding values in, your weights applied, one dollar index out. The figures below are made up to show the arithmetic; on a live account each weight is the engine's (or the ranking's standard weight where the engine has no channel), stored against the market snapshot it was computed from.

Bull A · self-replacing herd · heifers joined · marbling grid

On a growth-led national index this bull is moderate. On a farm that joins its own heifers and sells on a marbling grid, short days to calving and high marbling are worth more and mature cow weight costs money, so the same bull moves up the list.

Season Live

Each property with a location gets a season snapshot every day: rainfall and its 30-day and 90-day deciles against the 1991 to 2020 record for that spot, maximum and minimum temperature, humidity and a heat-load index, plus the feed-gap state, days of feed and pasture-quality path from the SeasonRisk feed-gap model where the property is mapped to a SeasonRisk enterprise.

When a weight, a scan or a score is saved, it is joined to its property's snapshot for that date. Records saved before the channel existed are back-filled after each daily run. On the animal's page the season is shaded under each series, so a drop in condition can be read against the season it happened in.

The reference property has a snapshot for every day since September 2025 (more than 400 days). Season is recorded beside the genetics today; it is not yet a term in the evaluation.

State: blood samples Research stage

This pipeline is built but no deliveries have been processed, so no scores exist on the platform today. When it runs: a blood sample goes to any laboratory that returns raw gene counts, and the delivery is uploaded as three files (the run's identity, one row per sample, and the count matrix). Each sample is matched to an animal and passes a quality gate on mapping rate, read depth, RNA integrity where supplied, and a sex check against the register, with thresholds that name their source.

Genemap normalises the counts itself, estimates a batch effect where the design allows it and refuses to where it doesn't, and scores each sample against the definitions registered for that tissue. A score is placed on a reference range built from Australian ruminant samples on the platform for the same tissue and age class; a range needs ten scorable results before it exists.

Expression is not used in breeding values. No genetic parameter has been estimated for any expression score, and the table that would hold one refuses it by constraint. Every score definition is labelled Illustrative.

Decision tools

How we check our own numbers

Provenance on every number

A breeding value reports its source: a breed-society file with its date, or a Genemap run with the model, the relationship matrix it used, and per trait whether the parameters were estimated or supplied and from where. A season snapshot reports the weather source, the climatology, the heat-load formula and the run that wrote it. A market price reports its sale date.

Recovery gates

An estimator has to get known answers right on simulated data before its output is shown. The variance-component estimator passed on 8 of 9 simulated cells and failed the ninth honestly. The BayesC marker prior passed its test; BayesR did not and is not used. The large-population path has a measured agreement gate against the direct solve and does not deliver where the gate has not been met.

Verified deployments

After each deploy of the evaluation, the live code is downloaded and compared with the version that was tested; the version, its fingerprint and the commit are written in a deployment ledger.

Refusals with reasons

A file that partly matches is shown as a refusal with a worksheet, not quietly dropped. A scored trait with no scale is refused at the database. A reference range with fewer than ten results is not built, and the page says so.

What Genemap does not claim

Saying this plainly is part of how the platform is meant to work.

• Dollar values are partly reference values. With nothing entered on the farm profile, the engine prices 9 of 21 cattle traits and 6 of 24 sheep traits on labelled reference values. Structure, disease, temperament and birth weight have no channel in it. There are no wool prices until the AWEX price guides are licensed, and the pricing of resilience in hard seasons is a framework that is not applied.
• No calibration against real returns yet. Calibrating the weights against what cohorts actually returned at sale or slaughter is planned, once kill-sheet and sale results build up. It does not run today.
• No shared regional benchmarks yet. Nothing is pooled across producers today; the shared reference is empty.
• Expression scores are not in any breeding value or index, no genetic parameter has been estimated for one, and no samples have been processed.
• Ultrasound reports no values yet. Scan capture is being built and validated.
• Reaction norms have not been estimated. An animal's genetic response to its season is not part of the evaluation; the season is recorded beside each measurement so that it can be.
• The large-population path is not demonstrated at scale. At 1,000 animals with 600 genotyped it meets its agreement gate only with a larger core than its default, and a 20,000-animal run did not finish in three hours. It is switched off and refuses to deliver where its gate is unmet, which is why evaluations stop at about 1,190 animals.
• Major genes are sheep only, and no animal has major-gene results loaded yet.
• Feed-efficiency devices, walk-over and connected scales, microbiome sequencing and wearables are not live feeds. Bring those measurements in as recorded traits or a scale file.
• Parameter floors. Genetic correlations are not estimated from fewer than 2,000 jointly recorded animals, and heritabilities from fewer than 500. The reference herd has not reached either floor, so its evaluation runs on published parameters and says so.