We applied the continuous improvement methodology known as six-sigma to approach breeding as a process and enable breeding scheme improvements. Six-sigma is a five-step method: define, measure, analyze, improve, and control [DMAIC, (də.′meı.ık)]. The steps are undertaken iteratively to create a cyclical method for continuous improvements. Six-sigma was originally proposed by Bill Smith at Motorola in 1986 to improve industrial processes, in which 99.99966% of all opportunities to produce some features of a part are statistically expected to be free of defects (Tennant, 2017). In the case of plant breeding, this would imply producing better varieties than the ones existing in the market and steady genetic gains with higher probability. The six-sigma method reflects the scientific method, but it is used for process management rather than hypothesis testing. To increase the efficiency and ease of use of six sigma tools (e.g., value stream mapping, correlation analysis, etc.) by breeding teams, we developed a software named Breeding Pipeline Manager (BPM) to document, describe, and visualize market segments, product profiles and breeding schemes. BPM is available at http://bpm.excellenceinbreeding.org? In collaboration with multiple breeding programs for a wide range of crops within the CGIAR, such as line, hybrid, and clonally propagated species, we identified breeding decisions which fell into three categories: crossing, evaluation and selection decisions. We then summarized breeding schemes and decisions as a table containing the breeding stages in rows (e.g., seedling nursery, stage 1 yield testing, etc.) and the CES tasks and decisions in columns. BPM provides a graphical user interface for capturing breeding schemes in a standard format. BPM also allows users to create visualizations (flowcharts) of their breeding scheme. In addition, the BPM allows market segment and product profile definition and users can link breeding schemes to market segments.

The BPM back-end was developed in Node JS, an open-source, cross-platform, and scalable JavaScript runtime environment. The front-end graphical user interface was developed in React JS. The source code is available at https://gitlab.com/excellenceinbreeding/module2. The platform leverages NodeJS asynchronous technology to perform intensive calculations without affecting the performance of other functionalities of the system. In addition, the platform uses Docker containerization technology for continuous development and integration (Merkel, 2014; Boettiger, 2015). This will not only enable automation of the deployment process but also horizontal scaling on any cloud infrastructure (depending on traffic). An online manual showing the details of the available features of the software and their use can be found once connected in the tool under the question mark bar at the bottom menu.

We selected the CGIAR IITA-Cassava program to showcase the importance of using enabling tools and simulations to continuously optimize breeding programs. The IITA cassava program is situated in different parts of sub-Saharan Africa to serve region-specific challenges and market segments. The IITA-Cassava east-Africa pipeline, situated in Uganda, was chosen to showcase the use of SixSigma and the BPM tool to improve their breeding scheme. The five six-sigma steps were applied as follows in the cassava program. The problem was defined as lower than achievable genetic gain for traits of interest under the current scheme. The breeding targets and scheme were measured (documented) by capturing all CES decisions across all stages using the BPM tool (as described in the next section) through several interactions with the breeders. The analysis of the measured decisions and the genetic gain indicators revealed many possible improvements. We first chose to analyze the crossing decisions in the breeding scheme to identify possible improvements. The number of parents (nParents), number of crosses (nCrosses), number of progeny per cross (nProgeny), and recycling strategy were prioritized for evaluation via stochastic simulation in AlphaSimR (Gaynor et al., 2021). We proposed an improvement plan based on the close-to-optimal number of parents, number of crosses, number of progeny, and recycling strategy identified via simulation. The improvement plan used the A3 format (referring to the size of an A3 sheet that describes a project briefly) common in project management (Anderson et al., 2011). We then controlled the improvement by monitoring how key performance indicators (a set of quantifiable measurements used to gauge an institution’s overall long-term performance) stated in the improvement plan changed as the improvements proceeded.

To keep the resources constant with the baseline, we restricted the number of individuals (nIndividuals) at the F1 stage to 1,632 in all experimental simulation treatments. We then developed a grid to evaluate different numbers of parents in the crossing block using the following possible numbers of parents (nParents): 4, 8, 16, 32, and 64. The number of possible crosses for each level of number of parents was constrained to a maximum of nParents * (nParents − 1)/2, which is equivalent to all possible combinations of parents or a half-diallel, while considering the initial restriction that the number of individual progeny (nIndividuals) must be equal to 1,632. This resulted in the following possible numbers of crosses: 6, 12, 24, 48, 96, 204, 408, or 816. To keep the number of individual progeny constant at 1,632, the number of progeny per cross was set to 272, 136, 68, 34, 17, 8, 4, 2 for numbers of crosses equal to 6, 12, 24, 48, 96, 204, 408, and 816 respectively. The number of individual progeny is always equal to the number of crosses multiplied by the number of progeny per cross.

As such, a total of 24 simulation treatments were defined (Table 2). To identify the optimal number of parents, number of crosses, number of progeny per cross, and the best recycling strategy, a stochastic genetic simulation was conducted in the R package AlphaSimR (Gaynor et al., 2021).

Following the paradigm of approaching breeding as an industrial process (Bernardo, 2002), we adapted the six-sigma methodology to continuously improve the breeding schemes of breeding programs (Figure 3). Under this framework, we follow the project methodology Plan-Do-Study-Act inspired by William Edwards Deming named DMAIC, an acronym standing for Define, Measure, Analyze, Improve and Control steps that are cyclically repeated to reflect the continuous or cyclical approach (Aguayo, 1991; Figure 4). To demonstrate the use of the continuous improvement methodologies to optimize breeding processes leveraging from measuring tools like the breeding pipeline manager (BPM) and stochastic simulation, we engaged in discussions with the IITA Cassava program. First, the cassava team registered their breeding pipelines and the market segments targeted per pipeline. We found the IITA-cassava program to be composed of five breeding pipelines and on average tackling six market segments. The market segments and accompanying pipelines are stratified by a combination of regional consumption preferences and prevailing biotic and abiotic stresses. For example, most of the produced cassava in West Africa goes to processed (granulated and paste) products while in east Africa, the predominant preference is for fresh consumption with minimal processing (boiling, roasting and flour from dried roots). Subsequently, we focused on the IITA-Cassava east-Africa pipeline targeting market segments listed by breeders and generally described as fresh market and high-quality flour. Even though we proposed six-sigma for improving breeding schemes, the reader should keep in mind that continuous improvement applies to all components of the breeding process including the management roles which are responsible of the encouraging and incentivizing improvements.

The step of defining the problem was adapted to breeding scheme improvement by defining the problem as a suboptimal rate of response to selection (genetic gain) but pointing to one of the many crossing-evaluation-selection (CES) tasks and decisions at a given stage as the possible root of the problem. We found tools such as Project charter useful to define the problem (McKeever, 2006). We proceeded to use the project charter to define or state the problem in the IITA-Cassava east-Africa pipeline as having “potential for greater response to selection without increased expenditures.” Details in the business case, goal statement, timeline, scope, and team members can be found in Table 4. Unfortunately, we found that estimates of realized genetic gain were not available in the program to justify the definition of the problem. However, given the lack of an efficient recurrent selection strategy, we assumed the definition of the problem to be relevant to the program.

To facilitate the measuring step of the continuous improvement approach proposed, in which CES decisions are recorded for further analysis (Bhuiyan and Baghel, 2005), we developed the BPM software. The breeding pipeline manager tool (BPM) is equipped with a module to define breeding pipelines as the sum of efforts to deliver a product. Breeding pipeline definition is the highest-level unit of information clustering in the BPM tool. The pipeline can be linked to market segments defined by the user. The market segment is defined in the BPM tool as the sum of the client, the target population of environments (TPE), and final product characteristics displayed in Table 5. These aim to capture the characteristics that can make breeding a more targeted effort according to Ragot et al. (2018) (Figure 5A). The BPM module can be incorporated to any enterprise breeding system (database) to properly link the generated phenotypic data to clear target segments and pipelines. Market segments for the cassava pipeline were captured using the BPM tool and are shown in Figure 5A. The major focus is on lowland high-rainfall, late maturity, long, hard cassava for fresh and flour consumption.

On top of defining the market segments, breeding programs must describe specifics of the product to be released in the market. Here, the concept of product profile (sometimes referred to as product concept) applies (Ragot et al., 2018; Carey et al., 2021). The existence of these profiles can make the difference between success and failure (Carey et al., 2021; Mwanga et al., 2021). The BPM tool has a module to define product profiles and link them to specific market segments, and the cassava breeders used the tool to formalize such profiles (Figure 5B). One of the product profiles for example is focused on achieving defined levels of fresh yield, plant height, dry matter and cassava mosaic disease resistance.

Part of the design of breeding pipelines is the creation of a blueprint or a breeding scheme that will allow the breeder to achieve the product profile for the market segment while maximizing the genetic gain of the breeding population per dollar invested (Henryon et al., 2014). The blueprint should specify all the crossing, evaluation and selection tasks and decisions occurring at the different stages (e.g., recombination, multiplication and testing stages) for the purposes of population improvement and product development. Most breeding programs have these two purposes coupled in a way that advancement decisions influence the recycling decisions. Others have proposed and shown that decoupling the population improvement from product development by moving the recycling decision to very early stages (e.g., F2, nursery or multiplication stages) will increase the rates of genetic gain. Better products can be expected when the product development process is regarded as separate from a rapid cycling population improvement strategy (Gaynor et al., 2017).

Crossing, evaluation and selection (CES) decisions comprising the breeding scheme can and should be recorded at the highest level of detail and safeguarded for the benefit of the breeding organization in case of any adverse circumstances. In Table 6 we show the CES decisions that should be considered to capture the level of resolution and detail necessary to avoid loss of valuable information; these can be recorded by the BPM tool in the breeding scheme module. The software allows for breeding pipelines to manage multiple breeding schemes, as may happen when a program has a principal breeding scheme, but one or more parallel experimental breeding schemes, to accelerate genetic gains.

The measuring step of the six-sigma process was adapted by recording numerically and categorically all the different CES decisions across the various stages directly impacting genetic gain (e.g., number of parents, # crosses, # progeny, coupling method, etc.). The lack of available tools to measure/record breeding schemes was the motivation to develop the BPM tool presented above, although the BPM tool is inspired by the Value Stream Mapping approaches commonly used in process management (Singh et al., 2011). We measured or recorded the breeding scheme of the IITA-Cassava program using the BPM tool to capture all crossing, evaluation and selection decisions across the different breeding stages and a portion can be observed in Figure 6. We captured seven stages (crossing block, multiplication and five stages of yield and agronomic evaluation) across 52 different CES decisions for the East-Africa cassava pipeline that informed the analysis step to identify areas for improvement (Figures 4, 5). These decisions comprise the crossing, evaluation and selection strategies for both population improvement and product development.

The analysis step was adapted to breeding scheme improvement by replacing approaches such as correlation analysis. Correlation analysis is a method that links a response variable or key performance indicator (KPI) to another variable in the production process to understand relationships that could indicate the part of the process that needs to be refined. We instead conducted an analysis based on known quantitative genetic relationships between the various CES decisions and genetic gain (e.g., program size affects genetic gain depending on how effectively genetic variance is utilized and also linked to selection intensity). Additional tools like Fishbone (diagram to articulate the root causes of the problem) are not discouraged but we limited this exercise to one-to-one meetings with the breeding team to discuss the possible gaps while analyzing the current scheme together in the light of quantitative genetic principles (Ishii and Lee, 1996). We initially found several possible areas of improvement, including the small size of the program, the experimental design used at yield and agronomic evaluation stages, the coverage of the target population of environments (TPE), the opportunity to use molecular information, the potential improvement of analytical methods for genetic evaluation, a possibility to select the best families at earlier stages, the possibility to reduce the cycle length, and other decisions such as an improved crossing plan. Since it is well-known from classic quantitative genetics theory that using resources properly to maximize the genetic variance observed among and within families can maximize response to selection (Lynch and Walsh, 1998; Hallauer et al., 2010), we chose to optimize the decisions of number of parents, number of crosses and number of progeny per cross given the low number of parents used by the program in the crossing block and very likely limiting the rate or sustainability of genetic gains. Although we first focused on improving the resource allocation for the number of parents, crosses and progeny, the reader should remember that as a continuous improvement process, the other areas of opportunity identified should also be improved right after or at the same time depending on the resources available. This is just an example of how to implement breeding scheme improvement.

Prior to recommending an improvement plan, we used genetic simulation (Gaynor et al., 2021) to identify optimal use of resources (plots available) by defining a grid of possible treatments that contained different combinations of number of parents, crosses and progeny subject to the constraint of 1,632 individuals at the F1 stage assuming other factors constant (e.g., properly resourced, properly tested, etc.). Regarding recycling strategy, using overlapping cohorts to recycle (i.e., a mixed crossing block composed half of parents from the PYT and half of parents from the AYT) lead to higher genetic gain regardless of the number of parents (Figure 7A). Based on this observation, we evaluated the effect on genetic gain of the number of parents, crosses and progeny while recycling from the mixed PYT and AYT.

For the single complex trait which represented an index of multiple traits, the decision of the number of parents provided the greatest opportunity to increase genetic gain. An excessive number of parents -here, more than 30- always resulted in decreased genetic gain compared to use of fewer than 30 parents at both the 20 and 60-year time horizons (Figure 7C). At the 20-year time horizon, the optimal number of parents was ∼8–16. However, at the 60-year time, the optimal number of parents increased to between 16 and 32.

Increasing the number of crosses generally increased gain, but with diminishing returns to additional crosses at a given number of parents. At low numbers of parents, not enough possible unique crosses were available to take advantage of gains possible by increasing the number of crosses. Interestingly, the optimal number of crosses also differed in the short (20-year) and long (60-year) terms. At the 20-year timepoint, schemes with fewer crosses and more progeny per cross tended to have higher gain across numbers of parents, but at the 60-year timepoint schemes with relatively more crosses and fewer progeny had higher gain. However, even at the 60-year timepoint, the optimal number of crosses was much less than the possible half-diallel of unique crosses.

Given the genetic parameters specified for the cassava program, the use of ∼8–16 parents, ∼24 crosses and ∼68 progeny per cross in each crossing block per year was the optimal distribution to maximize genetic gain at the 20-year time horizon (Figure 7A). At the 60-year time horizon, the optimal distribution was 16–32 parents, 60 crosses, and ∼30 progeny per cross (Figure 7B). To consider both short- and long-term interests of the breeding program, we chose to recommend use of 15–30 parents recycled from the combined PYT and AYT stages with 40 crosses and 40 progeny, given the constraint of the program to handle ∼1,632 materials to start.

The improvement step in the six-sigma method was adapted to breeding scheme improvement by using management tools like the A3 format to reflect the current and future state of the CES decision (subprocess) together with an action plan laying with detail the actions required to achieve the future state (Supplementary File 1; Anderson et al., 2011). We included a RACI chart (responsible, accountable, consulted and informed people in the improvement plan) to formalize the process to achieve the desired improvement. Is important to notice that a RACI chart can and should be employed during the management of the different tasks of the breeding process and not only for the continuous improvement of breeding schemes. We propose that the future state and actions included in the improvement plan should be guided by sound quantitative genetics principles and recommendations coming from state-of-the-art tools, such as evaluation of new strategies by genetic simulation (Mi et al., 2014; Gaynor et al., 2017; Pook et al., 2020, 2021). We expect that results obtained through simulation can identify close-to-optimal solutions and changes to the breeding CES tasks and decisions.

Based on the simulation findings, a meeting with the IITA-cassava breeding team was held to discuss the optimal scenarios revealed by simulations and the next steps. The recommendation to use between 15 and 30 parents in the crossing block depending on the target breeding-time-horizon was and to use a mixed crossing block of parents from both the PYT and the AYT was accepted by the team. The improvement plan developed by the IITA-cassava program included detailing the current and future state can be found in the Supplementary File 1. The improvement plan developed included actions like team agreement on the modification of the number of parents, number of crosses and number of progeny per cross used in the crossing block, the development of a new SOPs for the crossing block stage, training the technicians to execute the new SOPs, monitoring the genetic gain across years to confirm the positive change, among others.

The control step was adapted to breeding program improvement by adding a monitoring section to the improvement plan that keeps track of the progress of the action plan through the inclusion of key performance indicators (KPIs), deadlines, and risks, as it does in other industrial processes. To monitor or control the progress of the improvement plan in the IITA-cassava, deadlines and key performance indicators for the different actions were defined and monitored to ensure that changes occur. Once the new process was adopted, we moved to the next possible crossing evaluation or selection decision that could be causing low rates of genetic gain. This process is still undergoing together with other improvements identified.