Plant breeding changes the inherited characteristics of plant populations to meet a defined purpose. A dependable breeding programme connects the intended users and growing conditions with parent choice, genetic variation, selection, testing and the way planting material will be produced. Its result is not simply an exceptional plant in one plot: it is material whose useful performance can be demonstrated and maintained in the intended setting.
The gold standard is to keep that whole chain visible. This guide explains why the best parent is not always the tallest plant, why a promising cross needs further evaluation, and why a variety that wins in one environment can lose in another. All lettered lines, measurements, costs and numerical examples are original fictional teaching cases, not crop recommendations or reported trial results.
THE GOLD STANDARD · PLANT BREEDING
Connect inherited variation with useful performance
Choose a route through parent choice, selection, comparative trials and practical breeding decisions.
ROUTE 1 · CHAPTERS 1–3
Define the breeding purpose
Connect users, inheritance and parental contributions.
ROUTE 2 · CHAPTERS 4–6
Create and maintain variation
Follow crosses, reproductive routes and backcross expectations.
ROUTE 3 · CHAPTERS 7–10
Test candidates fairly
Connect selection response, trials, environments and requirements.
ROUTE 4 · CHAPTERS 11–14
Evaluate the programme
Examine prediction, speed, stage decisions and repair.
ROUTE 5 · CHAPTERS 15–17
Transfer and continue
Change environmental weights, answer questions and practise.
Full chapter index · Try the transfer task · Gold Standard subject hub
Chapter index
Define the breeding purpose · Chapters 1–3
Create and maintain variation · Chapters 4–6
Test candidates fairly · Chapters 7–10
Evaluate the programme · Chapters 11–14
Transfer and continue · Chapters 15–17
CHAPTER 1 OF 17 · Define the breeding purpose
1. Define improvement before making a cross
Improvement always has a direction and a context. A farmer seeking reliable production under limited water may value a different combination of traits from a grower supplying a market with strict appearance requirements. A breeding goal should identify the crop, users, environment, management conditions and measurable properties that matter.
Imagine a fictional grain crop with three proposed goals: increase yield, mature before a seasonal dry period, and meet a required grain-quality threshold. These goals are not interchangeable. A late, high-yielding candidate may fail the timing requirement; an early candidate may be unacceptable to the intended processor. Calling either one simply superior hides the decision that needs to be made.
Write an acceptance statement before inspecting candidates. For this teaching case, suppose maturity must occur within 100 days, quality must score at least 7 on a defined test, and yield is then compared among candidates satisfying both requirements. The exact numbers are invented. Their purpose is to distinguish mandatory conditions from properties to optimise.
CGIAR’s target product profile example describes using quantified attributes and stakeholder input to guide breeding. Such a profile gives scientific measurements a practical destination. It also makes disagreements visible: people may agree about measured yield while valuing timing, processing quality or labour needs differently.
The aim is not to freeze every decision permanently. Conditions and needs can change. Preserve the original goal and explain revisions so that a later change in ranking is understandable. Without that record, a programme can appear to succeed by quietly redefining success after seeing its results.
CHAPTER 2 OF 17 · Define the breeding purpose
2. Separate inheritance, breeding and crop management
Plant Genetics asks how hereditary information is transmitted and expressed. Breeding uses that knowledge to choose and develop material for an objective. Crop management changes the conditions under which plants grow. All three affect outcomes, but evidence for one does not automatically establish the others.
Suppose plants from one seed lot yield more after better irrigation. That is evidence of a management response under the comparison used. It is not, by itself, evidence that the population has undergone inherited improvement. Conversely, a genetic difference can be useful only when the management and environment allow it to be expressed.
In a second fictional comparison, Candidate A receives extra fertiliser while Candidate B does not. If A yields more, the trial cannot isolate the candidate difference. The repair is a design that compares the candidates under comparable conditions, or explicitly tests candidate-by-management combinations when that is the intended question.
The distinction matters for reporting progress. “Our plots yielded more this year” and “the new material outperformed the reference under comparable conditions” are different statements. The first can be true while leaving the contribution of breeding unresolved.
A practical notebook should therefore keep material identity and management records separately, then connect them through plot identifiers. A name alone cannot explain a harvest, and a management record alone cannot establish which genetic material was tested. Breeding evidence depends on preserving both sides of the comparison.
CHAPTER 3 OF 17 · Define the breeding purpose
3. Choose parents for the variation they can contribute
Parent choice begins with the target traits and the evidence available about potential donors. An impressive individual may owe part of its appearance to favourable conditions. A parent useful for one trait may also bring unwanted characteristics. The question is what variation the parent can contribute to the breeding objective.
FAO’s pre-breeding course identifies parental selection, use of non-adapted material, management of segregating populations and tracking genes of interest as connected tasks. Access to diversity creates opportunities, but useful diversity must be connected to a workable breeding population.
Consider fictional Parent P with acceptable maturity and quality but weak performance under a target stress. Parent Q contributes a promising stress-related trait but matures too late. Crossing them is a way to create combinations for evaluation; it does not guarantee that the offspring combine only their desirable features.
A useful planning table records the evidence for each proposed contribution. Was it measured in comparable conditions? Is it repeatable? Does the donor carry the trait in a background relevant to the intended cross? Which other properties will need to be monitored? This is more informative than choosing parents from a list of superlatives.
The important distinction is between a parent as a source and a finished candidate as a product. A valuable donor need not itself be a suitable variety for users. Equally, a popular variety need not contain every source of variation required for the next objective. Parent choice connects these roles without confusing them.
CHAPTER 4 OF 17 · Create and maintain variation
4. Crossing creates combinations; selection changes which continue
Crossing brings parental genetic contributions together. Segregation and recombination can generate variation in later offspring. Selection determines which individuals or families contribute to subsequent stages. These are different operations: generating many combinations is not the same as recognising useful ones.
Use a deliberately simple two-locus model. Assume an individual is AaBb, the loci assort independently, the plant self-fertilises, and there are no viability or scoring differences among genotypes. The probability of aa is one-quarter and the probability of bb is one-quarter. The probability of aabb is therefore one-sixteenth.
In 160 offspring the expected number of aabb individuals is 10. This is an expectation, not a guarantee that every group of 160 contains exactly ten. The chance of finding at least one is 1 − (15/16)^160, approximately 99.997%. The probability calculation assumes independent offspring draws under the stated model.
Changing those assumptions changes the answer. Linkage can invalidate the independent-locus multiplication. Differential survival can change the distribution among plants actually scored. An assay that does not distinguish genotypes introduces a separate observation problem. A correct Mendelian calculation cannot repair a mismatch between model and material.
The practical teaching lesson is to state what the cross is expected to generate, what will be observed, and which plants will continue. A programme that loses the relevant class before measuring it has not tested the original genetic expectation simply by counting the survivors.
CHAPTER 5 OF 17 · Create and maintain variation
5. Match the breeding route to reproduction
The way a crop reproduces affects what breeders can maintain and distribute. A self-pollinating line, a population maintained through crossing, a hybrid produced from selected parents and a vegetatively propagated clone are not the same kind of product. Their identity and performance must be considered through the appropriate reproductive route.
IRRI’s account of developing rice varieties describes creating populations, selecting lines, replicated testing and later evaluation. It is a rice example, not a single timetable or procedure for every crop. The general connection is that selection and testing must fit the biological material being developed.
For a simple diploid selfing illustration, following the Iowa State population-genetics explanation, start with an Aa plant and assume ordinary Mendelian segregation without selection. Half its offspring are expected to remain heterozygous at this locus. Repeating the same assumptions across selfed generations gives expected heterozygosity fractions of one-half, one-quarter and one-eighth relative to the initial heterozygous state. This is a locus-level model; it does not certify whole-genome uniformity or practical performance.
Now change the task to maintaining a selected plant by vegetative propagation. The relevant questions include faithful multiplication and subsequent performance, rather than the same segregation sequence. Plant Biotechnology explains why propagation output, identity and establishment require separate evidence.
Before interpreting a seed label or breeding claim, ask what is being maintained: an individual genotype, an inbred line, a hybrid combination or a population. If that question is unanswered, the meaning of uniformity, stability and next-generation performance remains unclear.
CHAPTER 6 OF 17 · Create and maintain variation
6. Read a backcross as a targeted recovery process
A backcross returns offspring to a selected parental background. Wageningen University’s marker-assisted breeding material explains the expected recovery of the recurrent-parent background and variation among individuals. In a simplified model, a donor contributes a desired feature while repeated crossing to a recurrent parent increases the expected recurrent-parent contribution. This is a strategy for a particular objective, not a guarantee that all unwanted donor effects disappear.
Assume the initial cross has one-half recurrent-parent ancestry in expectation. Backcrossing to that parent gives an expected three-quarters in the next generation, then seven-eighths, then fifteen-sixteenths. The recurrence is p next = (1 + p current)/2. These are genome-wide expectations under a simple model, not exact percentages for each individual or every chromosomal region.
Selection for a donor feature deliberately retains donor material associated with that feature. Nearby linked material may also remain. Therefore a high expected recurrent-background percentage cannot establish that the desired difference is the only difference. The actual selected material and its performance still need examination.
An original interpretation task makes the limitation concrete. A student sees fifteen-sixteenths and concludes that 93.75% of every plant’s phenotype must match the recurrent parent. The fraction does not describe a partition of phenotype. It concerns an ancestry expectation under the model, while phenotype depends on particular alleles, interactions and conditions.
IRRI’s rice account describes marker-assisted backcrossing as a way to improve selected traits in an existing variety. Read that as a defined breeding strategy with subsequent checks. Do not turn it into a universal claim that a few backcrosses create an otherwise unchanged plant or eliminate the need for comparative trials.
CHAPTER 7 OF 17 · Test candidates fairly
7. Selection response is not the selected-parent difference
The selected parents can differ substantially from the original population without their offspring reproducing that entire phenotypic difference. The existing heritability and selection-response article derives this distinction and explains its assumptions in detail.
For a compact fictional example, let the original mean be 40 units and the selected-parent mean be 46. The selection differential is 6 units. If a justified simple model uses narrow-sense heritability of 0.25, its predicted response is 0.25 × 6 = 1.5 units. The predicted mean on the comparable environmental baseline is 41.5, not 46.
This calculation is not a universal predictor for every breeding design. The relevant genetic unit might be an individual, family mean, inbred line or clone. The variance definition, mating design and target conditions need to match the prediction. A value copied from a different population cannot simply be inserted because it has the name heritability.
Piepho and Möhring’s 2007 primary paper addresses the difficulty of calculating heritability and response from unbalanced plant-breeding trials. Its relevance is a warning against treating a convenient ratio as automatically adequate for the actual design.
Keep predicted and observed response separate. A prediction guides planning. A controlled comparison supplies evidence about realised performance. If they differ, investigate assumptions, uncertainty and environmental differences before deciding that the entire theory or the entire experiment must be wrong.
CHAPTER 8 OF 17 · Test candidates fairly
8. Replication and randomisation protect comparisons
A plant’s measured performance includes more than its genetic identity. Soil position, weather, management and measurement can contribute. Trials need a design capable of distinguishing candidate differences from those other sources of variation. Replication and randomisation address related but different problems.
In a fictional field, the northern half is more productive than the southern half. Placing every A plot in the north and every B plot in the south entangles candidate and position. Repeating measurements on plants within those same plots does not remove the confounding. More numbers are not automatically more independent comparisons.
Now imagine four blocks, each containing both A and B in randomised positions. The paired yield differences A minus B are 0.2, 0.4, −0.1 and 0.3 units. Their mean is 0.2. This descriptive result is compatible with variation among blocks; it does not establish the difference with unlimited certainty or demonstrate performance in other environments.
Preserve plot identity, block, candidate, sowing date, management and measurement definition. When a record is missing, record why. Quietly deleting failed plots can change the comparison if the failure is associated with candidate performance or a particular environmental condition.
A trial report should make it possible to reconstruct the design before interpreting the mean. The question is not whether the spreadsheet looks complete. It is whether the observations support the comparison being claimed. Trial quality begins before sowing, with the allocation of material to the conditions in which it will be measured.
A candidate’s relative performance can depend on the environment. This is especially important when a breeding programme serves several growing conditions. A single average can hide a reversal, while a trial network can be unrepresentative of the conditions users actually face.
Consider invented yields for A and B. In a favourable environment, A yields 8 and B yields 7. Under a specified stress, A yields 3 and B yields 5. A wins the first comparison and B the second. Neither candidate is best without a stated target or decision rule.
If the target contains favourable conditions with weight 0.8 and stressed conditions with weight 0.2, A’s weighted score is 0.8 × 8 + 0.2 × 3 = 7.0. B’s is 0.8 × 7 + 0.2 × 5 = 6.6. If the weights become 0.3 and 0.7, the scores become 4.5 for A and 5.6 for B. The preferred candidate reverses.
These are weighted averages of fictional inputs. The weights might represent a specified target distribution for the exercise; they are not forecasts. In practice, uncertainties in the performance estimates and in the target distribution both matter. The weight assigned to severe losses may also reflect a decision objective beyond average yield.
FAO’s adaptation-strategy discussion distinguishes broad from specific adaptation and explains the role of multi-environment evidence. Its central practical lesson is to connect the testing strategy to the intended region and material, rather than assume that the widest average is always the most useful answer.
CHAPTER 10 OF 17 · Test candidates fairly
10. Distinguish a threshold from a weighted preference
Some traits are requirements; others can be traded against one another. A weighted score can be useful only if it represents the actual decision. It should not allow a candidate to compensate for failure of an essential condition merely by excelling elsewhere.
Return to the fictional requirement of maturity within 100 days and quality at least 7. Candidate A matures in 98 days, scores 8 for quality and yields 6 units. B matures in 106 days, scores 9 and yields 8. C matures in 96 days, scores 7 and yields 6.5. Under the stated requirements, B is not eligible even though it has the largest yield.
Between A and C, yield favours C while quality favours A. The next decision requires a stated preference or additional evidence. It cannot be settled by repeating that all three properties matter. A transparent choice explains whether quality beyond the minimum is valuable and how that value is assessed.
Now change the maturity requirement to 110 days because the intended production system changes. B becomes eligible. Its earlier exclusion was not a claim that B was biologically worthless; it was a consequence of the original use case. The revised decision needs fresh evidence that the new production context is adequately represented.
This distinction protects both scientific clarity and practical usefulness. Measurements describe candidates. Requirements and preferences determine how those measurements enter a choice. Keeping the layers separate makes a change of purpose visible instead of making it look like inconsistent science.
CHAPTER 11 OF 17 · Evaluate the programme
11. Markers and predictions still need validation
A molecular marker can provide information relevant to selection, but a marker result and a useful field phenotype are not identical claims. The relationship depends on what the marker measures, its association with the target feature, the population and the intended use.
For a fictional diagnostic exercise, suppose 100 candidates include 40 that truly satisfy a defined target trait under the test conditions. A screening method identifies 36 of those 40 and also flags 12 of the other 60. It flags 48 candidates altogether, of which 36 satisfy the target. The positive predictive fraction in this example is 36/48 = 75%.
The method finds 90% of target-positive candidates, but that is a different fraction: 36/40. Confusing the two would overstate how many flagged candidates are suitable. The example describes a generic screening comparison, not measured accuracy of any molecular marker or commercial test.
A useful selection stage may tolerate some false positives if later evaluation can identify them. It may be much less acceptable to lose rare useful candidates early. The appropriate balance depends on the cost of follow-up, the frequency of the target and the consequences of each error.
Record the reference against which the screen was assessed. Validation on the same material used to develop a predictor may not establish how it performs on new families or environments. The practical question is whether the prediction remains useful when the material and decision move beyond the original development set.
CHAPTER 12 OF 17 · Evaluate the programme
12. Faster generations and faster progress are different
Shortening generation time can accelerate part of a breeding programme. It does not automatically shorten every stage, improve selection accuracy or establish that a candidate will perform well in users’ fields. Measure the time to the relevant outcome rather than only the speed of one biological transition.
Watson and colleagues’ 2018 primary study demonstrated accelerated generation cycling in several crops under specified conditions. The finding supports the possibility of speeding generation advancement in tested systems. It does not provide a universal generation count for all plants or remove later performance testing.
An original arithmetic example compares Programme A, expected to gain 6 trait units per six-year cycle, with Programme B, expected to gain 4 per three-year cycle. Their simple annualised gains are 1 and about 1.33 units per year. The smaller per-cycle gain has the larger annualised value under these assumptions.
Now suppose B requires two additional years before the result reaches its intended endpoint. Its relevant interval becomes five years, and the ratio becomes 4/5 = 0.8 units per year. The conclusion changes because the endpoint and time boundary were clarified.
Neither calculation predicts indefinite linear gain. Genetic variation, objectives, environments and operational constraints can change over cycles. Annualisation is a comparison tool. A responsible report states the gain definition, time boundary and uncertainty, and does not extrapolate one favourable cycle into a permanent promise.
CHAPTER 13 OF 17 · Evaluate the programme
13. A breeding programme is a sequence of decisions
A candidate must pass through decisions whose evidence requirements differ. Early stages may handle many candidates with limited measurements. Later stages can use more detailed comparisons on fewer candidates. The purpose is not simply to make the funnel narrower; it is to spend effort where it most improves the final choice.
Imagine a fictional programme starting with 800 candidates. It retains 200 after an initial screen, 40 after a second stage and 8 for detailed testing. The retention fractions are 25%, 20% and 20% respectively. The final eight are 1% of the original set. These conditional denominators should be reported explicitly.
A low final fraction does not prove strong selection. If early screening removed useful candidates for irrelevant reasons, the funnel can be narrow and ineffective. Conversely, a higher retention fraction can be sensible when early evidence is weak and useful diversity would otherwise be lost.
| Decision | Evidence needed | Common mistaken shortcut |
|---|---|---|
| Choose parents | Relevant contributions and limitations | Select the largest-looking individual |
| Advance a candidate | Evidence for the stage’s defined criteria | Treat a marker signal as complete performance |
| Compare environments | Comparable trials and a target distribution | Average unlike conditions without explanation |
| Choose final material | Requirements, performance and uncertainty | Rank only the most impressive trait |
| Deliver planting material | Identity and suitable multiplication | Assume a successful trial guarantees supply |
At each handover, preserve the candidate’s identity, the reason for advancement and the unresolved questions. This lets the next stage test what remains uncertain rather than rediscovering what an earlier team meant by promising.
CHAPTER 14 OF 17 · Evaluate the programme
14. Diagnose a disappointing result without rewriting history
Suppose a selected line fails to outperform a reference. Begin by checking whether the intended comparison was actually made. Were the materials correctly identified? Were the environments relevant? Were management and measurements comparable? Were the criteria defined before the results were examined?
If candidate identity is uncertain, resolve that before interpreting the outcome genetically. If one trial experienced a special stress, preserve it and explain its relevance rather than quietly discarding it. If the proposed advantage appears only after changing the outcome definition, distinguish an exploratory finding from a confirmed objective.
An original example shows the danger of moving baselines. A new line yields 6 units this year; an old reference yielded 5 last year. The naive improvement is 20%. But a reference grown alongside the new line this year yields 6.2. The contemporaneous difference is −0.2, approximately −3.23% relative to 6.2. The historical comparison had mixed candidate differences with season differences.
That does not prove the new line is inferior in every situation. It shows that the original 20% claim did not isolate its contribution. Further comparisons might reveal a useful advantage in another trait or environment. Report that as a separate question with its own evidence.
The repair process should leave an audit trail: original claim, observed problem, revised explanation and next discriminating test. A disappointing result can improve the programme if it reveals a mistaken assumption. Concealing it removes information that could prevent the same error in another cycle.
CHAPTER 15 OF 17 · Transfer and continue
15. Transfer task: a different season mix changes the answer
Use the invented A and B yields from Chapter 9. Let w be the weight of favourable conditions and 1 − w the weight of stressed conditions. A yields 8 and 3; B yields 7 and 5. Find the value of w at which their weighted means are equal. Then explain which candidate wins on either side.
Answer and reasoning
A’s score is 8w + 3(1 − w) = 3 + 5w. B’s is 7w + 5(1 − w) = 5 + 2w. Equality requires 3w = 2, so w = 2/3. A has the larger weighted mean when favourable conditions have weight greater than two-thirds. B has the larger mean below that threshold.
At equality, both scores are 19/3, approximately 6.33 units. Equal weighted means do not mean equal performance profiles: A still has the larger favourable-condition yield and the smaller stressed-condition yield. A decision concerned with avoiding low outcomes could therefore require additional reasoning even at the equality point.
Now suppose the four yields were estimates with substantial uncertainty. The neat threshold is conditional on treating them as known. A real recommendation should examine how uncertainty changes the comparison, and whether the two-environment model adequately represents the target system.
The transferable skill is to identify which assumption supports the ranking. When users, environments, requirements or evidence change, rebuild the comparison. Do not memorise that one candidate was once declared the winner and carry that label into a different problem.
Is the highest-yielding plant always the best parent?
No. Its phenotype reflects its conditions as well as its genetic contribution, and parent choice may target traits other than yield. The useful question is what relevant variation it can contribute and what evidence supports that expectation.
Does crossing two good parents guarantee better offspring?
No. Crossing generates combinations that require evaluation. Desirable properties do not necessarily travel together, and later segregation can produce a range of outcomes. A cross is part of a strategy, not a promise about every descendant.
Is breeding the same as genetic engineering?
No. Breeding includes selection and reproductive strategies that do not require genetic engineering. Biotechnology can support breeding, but the breeding objective and evaluation remain necessary regardless of the tools used.
Can a marker replace every field measurement?
No. A marker answers a defined genetic or predictive question. Whether it supports selection for a practical outcome depends on validation in the relevant material and context. Measurements of performance remain important where the claim concerns performance.
Why test at more than one location or season?
Because the environment can alter both absolute performance and candidate rankings. More locations are useful when they improve representation of the target conditions. A large but poorly chosen trial network is not automatically more informative than a smaller, well-designed one.
Does rapid generation advancement prove rapid delivery?
No. It can shorten one part of the programme. Evaluation, multiplication and the requirements of the intended production system still affect the time to useful delivery. Report the time boundary alongside the speed claim.
Is diversity only useful before selection?
No. Diversity also matters for future objectives and changing conditions. The immediate product and the resources retained for future breeding are different assets. Selecting a current candidate need not mean treating all other material as worthless.
CHAPTER 17 OF 17 · Transfer and continue
17. Practise the reasoning and follow the next route
Start with a paper exercise before collecting more terminology. Choose one fictional objective, define two mandatory requirements and one trait to compare among eligible candidates. Construct three candidates, including one that excels in the comparison trait but fails a requirement. Explain why the ranking differs from a simple sort by yield.
Next, reproduce the weighted-environment calculation without looking at the answer. Change the weights, find the equality point, and explain the biological meaning of each number. This tests whether the mathematics is connected to the breeding decision rather than being used as decoration.
Then read an actual breeding report. Record the material, target, parents or population, selection unit, trial design, environments and reference material. Separate observations from predictions and recommendations. If the report does not provide a needed detail, mark the gap rather than inventing a plausible value.
The Gold Standard of Education frames this as learning that transfers: use the same reasoning when the crop, objective or environment changes. Plant Genetics supplies the inheritance foundation; the heritability specialist develops selection-response calculations; Plant Biotechnology explains the route from cells and tissues to dependable plants.
For institutional learning, begin with FAO’s pre-breeding course and CIMMYT’s quantitative-genetics textbook entry. The linked IRRI account is a historical rice example, while the primary papers on trial inference and speed breeding address particular scientific questions. None is used here as a current legal release checklist or a guarantee of commercial performance.
Return to the Gold Standard master for related subjects or the How X Works hub for broader mechanisms. Keep the central question with you: what evidence connects this chosen material to the purpose it is expected to serve?