Epistasis Effects of Chalkiness and Application Using Marker-Assisted Recurrent Selection in Indica Rice

Abstract

Chalkiness is a complex quantitative trait regulated by both genetic and environmental factors. Reducing chalkiness has long been a research focus in rice genetics and breeding. A total of 108 markers on/closely linked to starch biosynthesizing genes, grain shape and chalkiness QTLs were used to detect interactions affecting chalkiness. A total of 30 and 39 marker pairs with significant bigenic epistasis were identified for percentage of grain with chalkiness (PGWC) and degree of endosperm chalkiness (DEC), respectively, of which 16 were commonly found in both traits. Using markers associated with chalkiness and marker pairs with significant epistatic effects as candidate predictors increased the coefficient of determination (R²) of the best multiple regression models for predicting both traits. GBSSI, SSIIa and the interaction between GBSSI and GBSSII were consistently identified in optimal models, indicating their critical roles in regulating rice chalkiness. R² for DEC and PGWC ranged from 36.5% to 42.7% and from 52.9% to 73.8% in two environments, respectively. PGWC decreased significantly from 38.9% to 15.10% after three cycles using marker-assisted recurrent selection (MARS). This study suggests that epistasis contributes substantially to the regulation of chalkiness, and demonstrates that MARS can effectively improve chalkiness without imposing obvious negative impacts on eating quality.

1. Introduction

Chalkiness, the white opaque areas in the rice grain, is an undesirable trait in almost all of the rice industries, which not only reduces head rice recovery because of starch granules loosely packed with large space in the chalky areas, but also lowers the value of milled rice because of poor appearance quality [1,2]. Chalkiness exhibited a highly significant negative correlation with head rice yield (HRY). A 1% decrease in chalkiness was accompanied by approximately a 1% increase in HRY in indica rice [1]. Therefore, chalkiness is often used as an indicator to evaluate rice grain quality and commercial value [3]. The arrangement and fine structure of starch granules was mainly regulated by chalkiness, affecting the physicochemical properties [4,5]. Chalkiness was negatively correlated with the amylose content, peak and breakdown viscosity but positively correlated with the branching degree, the enthalpy of gelatinization, setback viscosity and pasting time, which further impair eating and cooking quality [6,7]. Therefore, eliminating chalkiness would produce more head rice in the same land area, which would benefit farmers, millers, traders and consumers, thereby contributing to food security and potential income.

Chalkiness is a complex trait affected by environmental factors, cultivation and agronomic practices, as well as physiological and metabolic processes during grain filling [8,9,10]. Numerous quantitative trait loci (QTLs) or genes related to chalkiness have been identified, which were distributed across all 12 chromosomes [11,12,13,14,15]. Up to now, more than one hundred genes were identified by map-based cloning, reverse genetics or mutation studies. Generally speaking, these genes can be divided into seven categories, that is, starch synthesis-related genes, grain shape-related genes, storage protein-related genes, transcriptional regulators, organelle development-related genes, high temperature- and reactive oxygen species-related genes, and genes encoding key enzymes in some specific metabolic pathways affecting chalkiness formation [16].

Studies using association mapping have revealed that some starch biosynthesizing genes were associated with chalkiness. For example, SSIIa, ISA2, PUL and SSIIIa were associated with chalkiness in a population with 233 Australian rice breeding lines [17]. We also reported that SSIIa, GBSSI, BEIIb, PUL, SSI and SSIVb were strongly associated with degree of endosperm chalkiness (DEC) or the percentage of grains with chalkiness (PGWC) using an indica population of 375 diverse advanced lines from irrigated rice breeding programs in different countries [18]. For example, three SNPs of SSIIa explained 5.52–18.96% of the phenotypic variance in the percentage of grains with chalkiness (PGWC) and the degree of endosperm chalkiness (DEC) across the six environments, and two SNPs on Wx explained 5.45–8.90% of the phenotypic variance of the above traits across four or more environments. In addition, the PGWC and DEC of the subset with low amylose content (AC) were 50% lower than those of others with intermediate or high AC [10]. Wx encodes GBSSI, the primary enzyme amylose synthesis, which has three major alleles, Wx^a, Wx^b and Wxⁱⁿ. Near-isogenic lines containing the Wx^b allele had lower AC, PGWC and DEC than those with other alleles [19,20]. Decreased expression of OsSSIIa and OsSSIIIa resulted in a higher DEC [21]. Slender grains can significantly reduce chalkiness, since substances are transported by the dorsal vascular bundles of the glumes to the core and belly during grain filling; therefore, higher width grains are prone to insufficient grain filling because a longer nutrient transportation route results in a chalky area. Lots of evidence has been found indicating that some genes-related grain shape or size are involved chalkiness formation [22,23,24]. The gs3 mutant created by CRISPR/Cas9 produced long, slender grains, with lower chalkiness than that of the wild type; however, gs3gl3.1 double mutation had significant positive effects on the regulation of grain length and width, resulting in wider and longer grains with higher PGWC and DEC, while the combination of slender-grain hybrid rice caused by gs3 and gw8 mutations promoted endosperm development and improved appearance quality [23]. gl9, as a novel allele, with a 7-kb insertion in the second exon of GS9, increased grain yield and appearance quality simultaneously [24].

It is well-known that epistasis (the interaction between genes) plays an important role in the genetic determination of many quantitative traits in rice, such as grain weight, panicle size, chalkiness and so on [18,25,26,27]. Our previous study reported that allelic combinations between SSIIa and RMw513 revealed more variations in DEC [18]. OsSPL14 regulates grain size, and its downregulated plants showed a chalky endosperm, since the OsSPL14 and OsNF-YB9/YC8-12 subunits cooperate to enhance grain appearance quality by promoting Wx and PDIL1-1 expression in rice [28]. The above studies suggest that it is possible to improve grain yield and quality synergistically. Our previous study identified 29 marker–trait associations (MTA) for chalkiness, including 13 SNPs on starch synthesis genes (GBSSI, SSI, SSIIa, SSIV, BEIIb and PUL), 8, 3, 3 and 2 MTAs located on gene/QTLs for chalkiness, grain yield, grain shape and others, respectively [18]. More recently, studies investigating the effects and identification of different allele combinations of SSIIa and Wx have revealed significant variation in rice eating and cooking quality [29,30] and chalkiness [31]. Therefore, further investigation on the epistatic effects on chalkiness formation among those genes/QTLs might be essential to the design of efficient and effective gene pyramiding strategies for eliminating chalkiness in rice breeding.

The objectives of the present study were to investigate the effects of epistasis between markers for a set of known genes/QTLs for chalkiness; to study whether it is feasible to predict chalkiness using model equation generated by multiple regression analysis using significant markers and marker-pairs; and to test potential of model equation on low chalky rice breeding using marker assisted recurrent selection (MARS).

2. Materials and Methods

2.1. Plant Materials

The original MARS population for low chalkiness was developed using 16 parental lines, including 6461, HHZ 5-SAL 9-Y 3-Y 1, Weed tolerance rice, Yangdao 6, SAGC-6, FFZ1, IR 02A127, IR04A421, IR06A119, IR-BB57, PR 33282-B-8-1-1-1-1-1, OM6600, B11598C-TB-2-1-B-7, CT 16658-5-2-2SR-2-3-6MP, PAU-201 and ADRON 125. To avoid huge loads on manual crossing for population (next cycle) generation, all 16 lines as recurrent parents were backcrossed with IR36ms three times individually, and then BC₃F₁ (BC stands for backcross) plants with MSms were selected for selfing. 200 BC₃F₂ seeds of each backcross were collected and mixed completely. About 3000 seedlings were transplanted with a spacing distance of 0.2 × 0.2 m on 2019 at Lingshui, Hainan (LS). This area was isolated surround using 3 m high polyethylene film before heading. Seeds on male sterile plants were harvested individually and then mixed equally for next random pollination on 2019 at Hangzhou, Zhejiang (HZ). After three seasons of random pollination, 20 seeds on each male sterile plant with 80–95 cm in height were collected and mixed as original population (cycle 0, slefing 0, C0S0) All field experiments were conducted at experimental farm of Zhejiang A & F University, HZ and LS. Seedlings at 21 and 28 days after sowing were transplanted by 0.2 × 0.2 m space at HZ and LS, respectively. The nitrogen fertilizer dose was set at 225 kg/ha, which was applied at a ratio of 5:3:2 as basal, tillering, and panicle fertilizer, respectively. Phosphorus (80 kg/ha as P₂O₅) and potassium (50 kg/ha as KCl) were applied with nitrogen before transplanting as basal fertilizer. Other field management was performed according to local farming practices.

2.2. Phenotyping, Genotyping and Statistical Analysis

To avoid chalky grains breaking during milling, PGWC was determined manually using more than 100 dehulled grains with three replications. Dehulled grains and quartz sand were added to 10 mL centrifuge tubes and shaken until the aleurone layer was completely removed. DEC of polished grains were measured using a Cervitec Grain Inspector 1625 (Foss, Hilleroed, Denmark). Sequences, physical position, related gene or QTLs of markers used in study were shown on Tables S2–S5. PCR amplification was conducted in a 10 μL reaction mixture containing 50 ng template DNA, 0.5 μM of each primer, 200 μM of each dNTP, 1.5 μM MgCl₂, 0.1% Triton X-100 and 1 U Taq polymerase and 1.0 μL of 10× PCR buffer under the following program: 5 min at 94 °C, followed by 30 cycles of 15 s at 94 °C, 15 s at 55 °C, and 30 s at 72 °C with a final extension of 5 min at 72 °C. PCR products were separated on 6% nondenaturing polyacrylamide gels and observed by the SYBR^® Safe staining method. In the present study, rare marker alleles, occurring at a frequency of less than 5% among all the lines for any one marker were excluded from digenic epistatic analysis. Finally, 108 markers remained. To detect the presence of epistasis, a fixed effect model analysis was conducted for all the marker pairs. The model includes marker 1, marker 2 and the interaction between marker 1 and marker 2 as fixed effects. An interaction at the p < 0.01 level was regarded as significant and presented with “×” between marker pairs. Pairwise interaction analysis among 108 markers was conducted using R programming language. Stepwise multiple regression analysis was performed using SPSS 19.0 (SPSS Inc. Chicago, IL, USA) to develop prediction models for PGWC and DEC. The analysis utilized 29 molecular markers (marked with “*” on Tables S2–S5) that showed significant associations with chalkiness, as well as the marker pairs with significant interactions detected in Table 1. All SNPs of the alleles presented were marked with an underline to distinguish them from traits or genes.

Table 1. Significant epistasis for DEC and PGWC detected using pairwise analysis.

Traits	Interaction	Chr a	Pos (Mb) a	Trait a	Chr b	Pos (Mb) b	Trait b	p Value	PVE
DEC	GBSSI4IF × GBSSIIF	6	1.77	SS	7	12.92	SS	1.33 × 10−6	5.91
	GBSSI4IF × RM5499	6	1.77	SS	7	9.99	GY	9.55 × 10−5	4.14
	GBSSI4IF × RMw513	6	1.77	SS	5	5.36	CK GS GW	1.77 × 10−4	3.37
	GBSSIIF × RI02341	7	12.92	SS	1	2.77	CK	5.84 × 10−3	2.13
	GBSSIIF × RM12827	7	12.92	SS	2	8.11	CK GS GW	1.75 × 10−4	3.84
	GBSSIIF × RM15206	7	12.92	SS	3	16.77	GS	1.94 × 10−5	4.58
	GBSSIIF × RM18360	7	12.92	SS	5	14.09	CK	2.11 × 10−5	4.42
	GBSSIIF × RM20373	7	12.92	SS	6	23.93	CK	5.62 × 10−4	3.12
	GBSSIIF × RM228	7	12.92	SS	1	21.98		6.20 × 10−3	2.06
	GBSSIIF × RM27792	7	12.92	SS	12	7.10	CK	4.13 × 10−3	2.12
	GBSSIIF × RMw513	7	12.92	SS	5	5.36	CK GS GW	1.15 × 10−4	3.46
	nkssr04 × RM22483	4	30.72	GY	7	25.29	CK	4.52 × 10−3	2.72
	PUL2F × RMw513	4	4.39	SS	5	5.36	CK GS GW	1.02 × 10−3	2.67
	RI02341 × RM5499	1	2.77	CK	7	9.99	GY	5.28 × 10−3	2.25
	RI04841 × RM13308	2	24.19	CK	2	18.90	CK	1.89 × 10−4	3.66
	RI04841 × RM15937	2	24.19	CK	8	30.94	GY	2.37 × 10−3	2.79
	RI04841 × RM18068	2	24.19	CK	5	6.11	CK GS GW	8.88 × 10−3	1.94
	RI04841 × RM18360	2	24.19	CK	5	14.09	CK	6.77 × 10−3	1.98
	RM12827 × RM12878	2	8.11	CK GS GW	2	8.82	CK GS GW	2.47 × 10−3	2.79
	RM12827 × RM15206	2	8.11	CK GS GW	3	16.77	GS	1.92 × 10−4	3.77
	RM12827 × RM5436	2	8.11	CK GS GW	7	9.07	GY	1.20 × 10−4	4.06
	RM12827 × RM5499	2	8.11	CK GS GW	7	9.99	GY	1.96 × 10−4	3.89
	RM12827 × RMw513	2	8.11	CK GS GW	5	5.36	CK GS GW	8.54 × 10−3	1.83
	RM13308 × RM5499	2	18.9	CK	7	9.99	GY	2.59 × 10−3	2.46
	RM13308 × SSIIIbIF	2	18.9	CK	4	31.54	SS	6.98 × 10−6	5.18
	RM13308 × YN27	2	18.9	CK	1	6.52	GS GW	3.73 × 10−3	2.45
	RM15206 × RM18360	3	16.77	GS	5	14.09	CK	1.44 × 10−5	4.61
	RM15206 × RM5499	3	16.77	GS	7	9.99	GY	2.46 × 10−4	3.55
	RM18068 × RM5499	5	6.11	CK GS GW	7	9.99	GY	1.88 × 10−4	3.84
	RM20373 × RM5436	6	23.93	CK	7	9.07	GY	1.00 × 10−3	2.91
	RM20373 × RM5499	6	23.93	CK	7	9.99	GY	3.81 × 10−3	2.32
	RM20373 × RMw513	6	23.93	CK	5	5.36	CK GS GW	7.72 × 10−3	1.78
	RM26985 × RM5499	11	21.38		7	9.99	GY	1.71 × 10−3	2.73
	RM27792 × RM5436	12	7.10	CK	7	9.07	GY	1.79 × 10−3	2.53
	RM27792 × RM5499	12	7.10	CK	7	9.99	GY	5.12 × 10−3	2.08
	RM3414 × RMw513	6	2.88	GY	5	5.36	CK GS GW	1.82 × 10−6	5.07
	RM5499 × RMw513	7	9.99	GY	5	5.36	CK GS GW	3.46 × 10−4	3.09
	RM5586 × ssr0649	4	19.71	CK	8	25.30	GY	3.11 × 10−3	2.61
	SSI1IF × ssr0649	6	3.08	SS	8	25.30	GY	2.34 × 10−4	3.89
PGWC	AGPL1F × RM13574	5	28.79	SS	2	23.64	CK	4.33 × 10−3	2.63
	AGPL1IF × RI05333	5	28.79	SS	2	21.59	CK	5.04 × 10−3	2.15
	AGPS2F × RM12878	8	15.54	SS	2	8.82	CK GS GW	3.86 × 10−4	3.99
	AGPS2F × RM13308	8	15.54	SS	2	18.90	CK	1.26 × 10−3	3.11
	BEIIa1F × RID711	4	20.05	SS	7	28.48	GS	6.92 × 10−3	2.07
	BeIIb3F × RM22483	2	19.36	SS	7	25.29	CK	3.53 × 10−3	2.43
	BEIIbIF × RM22483	2	19.36	SS	7	25.29	CK	6.50 × 10−3	2.16
	GBSSI4IF × GBSSIIF	6	1.77	SS	7	12.92	SS	1.10 × 10−3	2.92
	GBSSIIF × RM15206	7	12.92	SS	3	16.77	GS	3.25 × 10−5	4.39
	GBSSIIF × RM18360	7	12.92	SS	5	14.09	CK	1.11 × 10−4	3.76
	nkssr04 × RM22483	4	30.72	GY	7	25.29	CK	4.76 × 10−3	2.68
	RI04841 × RM13308	2	24.19	CK	2	18.90	CK	5.18 × 10−4	3.19
	RI04841 × RM15937	2	24.19	CK	8	30.94	GY	1.63 × 10−3	2.96
	RI04841 × SSI2F	2	24.19	CK	6	3.08	SS	3.13 × 10−3	2.35
	RI04841 × STS5803	2	24.19	CK	2	9.04	CK GS GW	8.95 × 10−3	1.96
	RM111 × RM16868	6	5.10	CK GY	4	19.28	CK	1.28 × 10−3	3.02
	RM12827 × RM5499	2	8.11	CK GS GW	7	9.99	GY	3.59 × 10−3	2.51
	RM13308 × SSIIIbIF	2	18.90	CK	4	31.54	SS	6.32 × 10−5	4.21
	RM13308 × YN27	2	18.90	CK	1	6.52	GS GW	1.71 × 10−3	2.84
	RM15206 × RM18360	3	16.77	GS	5	14.09	CK	4.12 × 10−5	4.18
	RM15206 × RM5499	3	16.77	GS	7	9.99	GY	6.17 × 10−4	3.16
	RM16 × RM18068	3	23.08		5	6.11	CK GS GW	4.28 × 10−3	2.51
	RM163 × RM3575	5	19.17	CK	5	21.31	CK	3.85 × 10−3	2.31
	RM16868 × SSI1F	4	19.28	CK	6	3.08	SS	9.24 × 10−3	1.98
	RM18068 × RM5499	5	6.11	CK GS GW	7	9.99	GY	7.97 × 10−4	3.15
	RM23270 × RM252	8	25.25	CK	4	25.34		1.06 × 10−3	3.13
	RM23422 × RM24718	8	25.25	CK	9	20.90	GY	7.76 × 10−4	3.42
	RM26985 × RM5499	11	21.38		7	9.99	GY	5.35 × 10−4	3.33
	RM5586 × ssr0649	4	19.71	CK	8	25.30	GY	5.95 × 10−3	2.28
	SSI1IF × ssr0649	6	3.08	SS	8	25.30	GY	2.13 × 10−4	3.95

Note: ^a and ^b represent the former and latter markers, respectively. SS, CK, GS, GW and GY indicate that the marker was on or closely linked to genes/QTLs of starch synthesis, chalkiness, grain shape, grain width and grain yield, respectively. The blank in the trait column represents an unknown value. Chr and Pos indicate the physical location of each marker on the chromosome. PVE denotes phenotypic variation explained.

2.3. Recurrent Selection Program

The procedure of recurrent selection applied in this study is illustrated in Figure S1. There were two RS programs conducted for low chalkiness germplasm enhancement, that is, phenotypic selection (PS) and MARS with prediction models, respectively. Each cycle has two growing seasons. In the first season, the C0S0 population was grown with around 10,000 plants, and then about 300 plants were selected based on breeder’s experience in plant architecture and yield traits. All selected plants were genotyped using markers in the prediction models, and PGWC was measured as well. For phenotypic selection, all plants were ranked from lowest to highest based on PGWC, and then 60 of 300 (20%) plants with relatively lower on PGWC were selected. About 200 seeds randomly sampled from each selected plant were mixed for random pollination to generate next cycle on next season. For prediction models selection, 60 of 300 (20%) plants with relatively lower predicted value of PGWC were selected using molecular models, and 200 seeds of each plant were mixed and planted for random pollination on next season. The same steps were conducted on following cycles. Random pollination season (C0S0, C1S0, C2S0 and C3S0) was grown at LS. Generation (C0S1, C1S1, C2S1 and C3S1) for PS and MARS was planted at HZ. Finaly, four cycles were carried out during 2020–2023. All selected plants were generation advanced using single seed descent. Advanced lines from C0, C1, C2 and C3 were grown in 2025 at Hangzhou; days to maturity (DTM), amylose content (AC) and PGWC were measured, respectively.

3. Results

3.1. Number of Marker-Pairs Involved in Bigenic Epistasis

A total of 30 and 39 marker pairs with significant bigenic epistasis were identified for PGWC and DEC respectively, 16 were commonly found in both traits (Table 1). Among them, only 13 and 11 interactions for PGWC and DEC were detected between markers with significant main effects at least in one environment [18]. In addition, 5 and 12 interactions for PGWC and DEC were found between markers without main effects. For 30 significant interactions of PGWC, 10 were between markers on/tightly linked to genes/QTLs for chalkiness and starch synthesis genes; seven were between markers related to chalkiness and grain yield; four were between markers both related to chalkiness; three were between markers for chalkiness and grain shapes; three were between markers related to starch synthesis genes and grain shapes; two were between grain shape and grain yield; one was between both markers on starch synthesis genes. For 39 significant interactions of DEC, 15 were between markers related to chalkiness and grain yield; nine were between markers on/for starch synthesis genes and chalkiness; five were between both markers related to chalkiness; five were between makers for chalkiness and grain shapes; five were between markers related to starch synthesis genes and grain shapes; four were between markers for grain shapes and grain yield; two were between both markers related to grain shapes; and one was between both markers on starch synthesis genes. A single pair of interaction explained only 1.78–5.91% of the phenotypic variation, which was relatively small.

A total of 28 markers with significant interactions for DEC were detected, distributed on 10 chromosomes, except Chr 9 and Chr 10 (Figure 1). RM5499, a marker on Chr 7 closely linked to the grain yield trait, interacted significantly with 10 markers. GBSSIIF was involved in digenic epistasis with other nine markers. RMw513, a marker on Chr 5 related to chalkiness, grain width and grain weight, was involved in seven significant interactions. RM12827, a marker on Chr 2 related to chalkiness, grain width and grain weight, showed significant interactions with six markers. The above 4 markers contributed to 27 of the 39 major digenic epistasis interactions of DEC.

Figure 1. Significant interactions detected among markers for DEC. The markers in blue are located on starch synthesis genes; the markers in red, purple and black are tightly linked to QTLs for chalkiness, grain yield and grain shape, respectively; the markers in bold are located on genes of grain shape or grain yield; the markers in yellow are tightly linked to QTLs associated with both grain shape and chalkiness. The numbers represent the length of chromosomes (Mb).

Unlike DEC, where most interactions were mainly concentrated on few markers, 40 markers were involved in 30 significant interactions for PGWC, distributed on 10 chromosomes except Chr 10 and Chr 12 (Figure 2). Four markers (RI04841, RM13308, RM16868 and RM5499) and four markers (SSIF, GBSSIIF, RM15206, RM22483) were involved in major epistatic interactions with other four and three markers, respectively. The remaining markers were involved only in 1 or 2 major interactions.

Figure 2. Significant interactions detected among markers for PGWC. The markers in blue are located on starch synthesis genes; the markers in red, purple and black are tightly linked to the QTLs for chalkiness, grain yield and grain shape, respectively; the markers in bold are located on genes of grain shape or grain yield; the markers in yellow font are tightly linked to QTLs associated with both grain shape and chalkiness. The numbers represent the length of chromosomes (Mb).

3.2. Digenic Epistatic Effects

The effects of all significant interactions for PGWC and DEC are presented in Table S1. Generally speaking, the effects of each interaction were quite small, ranging from 1.78% to 5.91% for DEC and from 1.96% to 4.39% for PGWC, respectively. As a typical case of significant interactions between two starch synthesis genes, GBSSI4IF × GBSSIIF was detected for both DEC and PGWC and explained up to 5.91% of the phenotypic variance (Table 1). Both markers had two major genotypes. The DEC of four subsets based on allele combinations of these two markers were quite different, ranging from 5.63% to 22.91% and from 12.51% to 25.37% at LS and HZ, respectively (Table S1, Figure 3A). The subset of genotypes with GBSSI4IF Allele2 + GBSSIIF Allele2 had the lowest DEC in both environments, while GBSSI4IF Allele1 + GBSSIIF Allele2 had the highest DEC. The same trend as DEC was observed for PGWC (Table 2, Figure 3B).

Figure 3. Effects of four allelic combinations between two makers. (A,C,E,G,I,K) represent interaction effects of six marker pairs on DEC (Y-axis %), namely GBSSI4IF × GBSSIIF, RI04841 × RM15937, RI04841 × RM13308, GBSSIIF × RM15206, GBSSIIF × RM18360, RM12827 × RM5499, respectively; (B,D,F,H,J,L) represent interaction effects of the same set of marker pairs on PGWC (Y-axis %). Error bars indicate the variance in PGWC or DEC of each subset according to allele combinations. LS and HZ represent Lingshui, Hainan and Hangzhou, Zhejiang, respectively (X-axis).

Table 2. Multiple regression models using significant markers and marker pairs, and the associated adjusted R² for three quality traits measured in six environments.

Trait	Env	Equation	R2
DEC	HZ	Y = 15.24 − 9.577(SSIIaSNP2) + 13.999(GBSSI1IF) + 8.342(RMw513.3) − 4.684(RM18068.4)	0.262
	LS	Y = 20.167 − 7.689(SSIIaIF) − 7.803(RMw513.4) − 7.72(PUL2F.1) + 8.197(GBSSI1IF)	0.212
PGWC	HZ	Y = 76.94 − 32.629(SSIIaSNP2) − 19.079(RM28502.3) − 14.916(RMw513.4) + 29.278(GBSSI1IF) − 13.555(GBSSI3IF) − 10.215(RM18068.4)	0.340
	LS	Y = 48.813 + 24.301(SSIIaF.1) − 9.963(RI02451) + 15.65(GBSSI1IF)	0.258
DEC	HZ	Y = 14.076 − 14.688(SSIIaSNP2) + 20.71(GBSSI1IF) − 7.017(RM18068.4) − 15.498(GBSSI3IF) + 10.743(GBSSI4IF × GBSSIIIF.4)	0.365
	LS	Y = 31.034 − 20.239(SSIIaIF) + 9.392(RM28502.2) + 7.599(GBSSI4IF × GBSSIIIF.4)	0.427
PGWC	HZ	Y = 59.323 − 15.275(SSIIaIF) + 32.186(GBSSI1IF) + 56.216(GBSSI4IF × GBSSIIIF.4) − 17.211(YN27 × RM13308.4) + 17.583(RMw513.3)	0.738
	LS	Y = 67.14 − 44.2(SSIIaIF) + 18.98(GBSSI1IF) + 58.08(GBSSI4IF × GBSSIIIF.4) − 8.562(YN27 × RM13308.4)	0.529

Note: the binary coding used for the markers in multiple regression analysis was: GBSSI1IF: G (1), T (0) (the SNPs G and T indicate binary codes of 1 and 0, respectively; the same applies to the following); GBSSI3IF: C (1), A (0); SSIIaIF: TT (1), GC (0); SSIIaSNP2: A (1), G (0); GBSSI4IF × GBSSIIIF.4: C/C(1), others(0); SSIIaF: 81bp (1), others (0) (if the size of the PCR product is 81bp, it is assigned a binary code of 1; otherwise, (0); PUL2F: PSBRC94 (1), others (0) (if the size of PCR product is as the same as variety PSBRC94 given binary code 1, if not given binary code 0, the same applies to the following); RI02451: PSBRC82 (1), others (0); RM18068.4: PSBRC82 (1), others (0); RM28502.3: IR64 (1), others (0); RMw513.3: Teqing (1), others (0); RMw513.4: IR64 (1), others (0); YN27 × RM13308.4: IR64 (1)/IR60 (1), others (0). All SNP alleles presented were underlined to distinguish them from traits or genes.

RI04841 is a marker associated with chalkiness, while RM15937 is associated with grain yield. The significant interaction between them explained 2.67% and 2.79% of the variance in DEC and PGWC, respectively (Table 1). On average, the DEC of the set based on RI04841 Allele2 +RM15937 Allele1 was 3.88%, while that of the other 3 sets was 3 times higher at LS. In another environment, the DEC of this set was 7.82% only; however that of the other three sets was higher than 18.46% on average. The PGWC of this set was 17.41% and 38.23% at LS and HZ respectively, while that of the others was higher than 47% and 67% on average at LS and HZ, respectively (Table S1, Figure 3C,D).

RM04841 and RM13308 were both linked to QTL or genes involved in chalkiness formation. Epistatic effects between them accounted for 3.66% and 3.29% of the phenotypic variance in DEC and PGWC, respectively (Table 1). The DEC of the set with RI04841 Allele2 and RM13308 Allele1 was 5.75% and 10.15% at LS and HZ, respectively, while that of the other sets was almost double. The effects on PGWC were quite similar to those on DEC. The PGWC of the set with RI04841 Allele2 and RM13308 Allele1 was 25.29% and 40.57% at LS and HZ, respectively, while that of the other combinations ranged from 44.75% to 54.99% and from 69.55% to 85.64% at LS and HZ, respectively (Table S1, Figure 3E,F).

The interaction between GBSSIIF and RM15206 explained 4.71% and 4.49% of the variation in DEC and PGWC. The former and latter markers were close to or on genes related to starch synthesis and grain shape, respectively (Table 1). The DEC of the set with GBSSIIF Allele2 and RM15206 Allele2 was 5.95% and 10.96% at LS and HZ, respectively. On the other hand, that of the set with GBSSIIF Allele2 and RM15206 Allele1 was the highest among the four subsets, 21.91% and 26.65% at LS and HZ, respectively. The PGWC of the set with GBSSIIF Allele2 and RM15206 Allele2 24.18% and 39.51% at LS and HZ, respectively. The set with GBSSIIF Allele2 and RM15206 Allele1 showed the highest PGWC in both environments, 78.60% and 87.23%, respectively (Table S1, Figure 3G,H).

A typical case of significant interaction between markers on/related starch synthesis genes and chalkiness formation was GBSSIIF and RM18360, which explained about 4.42% and 3.76% of variation in DEC and PGWC, respectively (Table 1). Among four subsets, the DEC of the set with GBSSIIF Allele2 and RM118360 Allele1 was lowest in both environments, 7.01% and 12.42%, respectively, while the set with GBSSIIF Allele2 and RM118360 Allele2 was the highest. A similar trend was observed in PGWC (Table S1, Figure 3I,J).

RM12827 and RM5499, markers related to grain shape and yield, respectively, explained about 3.89% and 2.51% of the variation in DEC and PGWC (Table 1). The set of RM12827 Allele2 and RM5499 Allele1 showed the lowest DEC and PGWC in both environments, while the set with RM12827 Allele2 and RM5499 Allele2 showed the highest (Table S1, Figure 3K,L).

3.3. Molecular Marker Prediction Models of PGWC and DEC

To avoid too many independent variables, only markers strongly associated with a trait detected by Zhao et al. and marker pair with significant interaction found in the present study were used in model prediction [18]. When pairwise interactions were not considered, the adjusted R² of the best prediction models for DEC varied from 0.212 to 0.262 across two environments (Table 2). Genes SSIIa and GBSSI were common in the models. When digenic epistasis was considered, the adjusted R² of the best prediction models for DEC varied from 0.365 to 0.427 (Table 2); that is, by taking digenic epistasis into account, R² increased in both testing environments. Only markers on SSIIa were involved in the prediction models. The adjusted R² of the best models for PGWC ranged from 0.258 to 0.340 when digenic epistasis was not considered, whereas it ranged from 0.529 to 0.738 when digenic epistasis was considered (Table 2). Markers on SSIIa, marker pairs of GBSSI and GBSSII were in the final models of both environments (Table 2). Markers on SSIIa were present in all eight models, although different markers were chosen in different environments or traits. GBSSI was included in seven models for chalkiness prediction, except for DEC at LS.

3.4. Effects of Marker-Assisted Recurrent Selection on Chalkiness

The C0S0 individuals derived from three rounds of random pollination using 16 parents were used as the testing population for recurrent selection to develop low-chalkiness germplasm. The average PGWC of 200 selected plants was 38.08% (Table 3). The PGWC of 30% individuals with the lowest PGWC ranged from 14.0% to 29.0%, 23.08% on average, which was significantly lower than that of the 30% individuals selected based on the predication model, ranging from 15.0% to 41.0%, 26.58% on average. The PGWC of the lowest 30% of selected plants based on C1S0 PS and MARS was 14.97% and 17.85%, respectively. In C2S0, the PGWC of MARS was 14.97% on average, which was still significantly higher than that of PS. Interestingly, there was no significant difference observed in C3S0; the PGWC of PS and MARS ranged from 6.0% to 17.0% and from 7.0% to 20.0%, with averages of 11.88% and 12.97%, respectively. A similar trend was observed for the PGWC of advanced lines. In the first two cycles, the PGWC of PS was significantly lower than that of MARS; however, the PGWC of PS and MASR was quite similar in C3, with averages of 12.64% and 12.84%, respectively.

Table 3. Genetic gain in PGWC using phenotypic and marker-assisted recurrent selection.

Cycle	Methods	PSP	PPR		PAL		DTM		LAC (%)	IAC (%)	HAC (%)
			Range (%)	Average (%)	Range (%)	Average (%)	Range (d)	Average (d)
C0	PS	38.09 ± 29.74	14.0–29.0	23.08 ± 4.24	12.67–35.33	21.02 ± 4.72	123–145	133.65 ± 6.32	48.33	25.0	26.67
	MARS	38.09 ± 29.74	15.0–41.0	26.58 ± 6.41 **	11.67–42.67	23.04 ± 8.06 *	123–143	131.72 ± 4.86	95.0	5.0	0
C1	PS	24.63 ± 16.86	9.0–20.0	14.97 ± 2.75	9.33–27.33	17.24 ± 4.73	124–147	134.02 ± 6.40 *	53.33	23.33	23.33
	MARS	29.76 ± 18.54	6.0–35.0	17.85 ± 5.62 **	9.67–38.67	21.82 ± 8.54 **	123–144	131.75 ± 4.94	96.67	3.67	0
C2	PS	19.73 ± 14.17	6.0–20.0	12.90 ± 3.65	5.33–27.33	13.69 ± 5.25	123–147	134.32 ± 6.44 *	66.67	15.0	18.33
	MARS	22.81 ± 15.93	6.0–27.0	14.97 ± 5.62 *	5.67–26.67	15.95 ± 6.51 *	123–144	131.72 ± 5.00	98.33	1.67	0
C3	PS	14.07 ± 9.08	6.0–17.0	11.88 ± 3.14	5.33–23.00	12.64 ± 4.09	125–147	134.65 ± 6.41 **	71.67	11.67	16.67
	MARS	15.10 ± 8.65	7.0–20.0	12.97 ± 3.56	5.57–25.67	12.84 ± 4.72	123–144	131.73 ± 4.92	100	0	0

Note: PSP, PPR, PAL and DTM represent PGWC of the selected plants, the PGWC of the plants for recombination and the PGWC of advanced lines, respectively. LAC, IAC and HAC indicate low (<18%), intermedia (18–23%) and high (>23%) AC, respectively. PS and MARS represent phenotypic selection and marker-assisted recurrent selection, respectively. * and ** represent significant differences at p < 0.05 and p < 0.01 based on the t-test between PS and MARS.

4. Discussion

Consumers generally prefer rice with uniform shape and translucent endosperm, although preferences for eating quality characteristics vary across countries and regions [1,32]. Chalkiness refers to white, opaque areas, since starch granules in the chalky, opaque areas are loosely packed. Starch, including amylose and amylopectin, accounts for approximately 70–80% of dry matter in rice grains. Amylose is synthesized by AGPase (including four subunits, AGPL1, AGPS1, AGPS1 and AGPS2) and GBSSI while amylopectin is synthesized by the coordinated actions of AGPase, starch synthase, starch branching enzyme and starch debranching enzyme [33]. Therefore, it is not strange that some starch synthase genes are involved in chalkiness formation, since insufficient starch biosynthesis affects not only rice yield but grain quality. In the present study, 29 of 69 digenic interactions (epistasis) for DEC and PGWC were related to starch synthase genes such as GBSSI, GBSSII, AGPL1, AGPS2, BEII, SSI, SSIIIb and so on (Table 1). Among them, most were identified between starch synthase genes and markers linked to QTLs for chalkiness or grain size. GBSSII was not associated with chalkiness in this population [18]. It is interesting that GBSSII exhibited a high frequency of significant epistatic interactions with other markers, with nine and three interactions for DEC and PGWC, respectively. The interaction between GBSSI and GBSSII explained 5.91% and 2.92% of the variation in DEC and PGWC, respectively. In addition, markers for GS3 and GW5 were also found to be involved in epistasis with GBSSII in chalkiness formation. It is well known that GBSSII is responsible for the amylose biosynthesis in the chloroplast of leaves via fixing carbon by photosynthesis [34]. Foliar starch, known as transitory starch, affects not only plant growth but also grain quality, since it can be converted into sucrose and then transported into endosperm at night during the grain filling stage [35]. Overexpression of GBSSII, in the oslesv and osesv1 mutants, significantly increased the starch content in both leaves and endosperms and improved the grain weight and appearance qualities, although these phenotypic characteristics of the OE lines did not completely revert to wild-type levels [36]. In addition, the mutations of GBSSI created using CRISPR/Cas9 reduced, but did not abolish, GBSS activity in seeds due to partial compensation caused by the upregulation of GBSSII [37]. The above results suggested that GBSSII plays an important role in chalkiness by regulating the source–sink relationship. It is worth mentioning that pleiotropy of genes or markers showing epistasis in chalkiness formation was common. QTL mapping has revealed that epistasis plays an important role in the genetic control of chalkiness [15]. In the present study, of the major significant epistatic interactions for DEC and PGWC, 56 of 69 involved at least one marker related to a known QTL/gene for chalkiness. A total of 27 of the 69 significant interactions involved at least one marker related to grain shape or size. A total of 23 marker pairs were related to known QTLs of grain yield. These results implied that how to balance grain yield and appearance quality is a critical issue in modern rice breeding. Five significant interactions for DEC and PGWC involved at least one marker not in the regions with known QTLs, which indicated that epistasis between markers without main effects has long been seen as an important genetic basis for complex traits [15,26,27]. Although the interactions between genes with large main effects are more important in gene pyramiding, which is seen as one of the major MAS methods [38], the contributions of epistasis between the target genes and genes without main effects (background genes) should also not be neglected to achieve the expected results.

Recurrent selection has been wildly used in plant breeding for quantitative traits and has shown high efficiency [39,40,41]. Chalkiness is a complex trait not only affected by lots of QTLs or genes but also by environmental factors, cultivation and agronomic practices and physiological and metabolic processes during grain filling. The national standard for Grade 1 rice in China requires DEC to be less than 2% (GB/T 1354-2018 [42]); therefore, cultivars with Grade 1 rice are very difficult to develop, especially in hybrid rice, because of a narrowing genetic base. In the present study, each cycle has only two seasons, one for random pollination and another for single plant selection, which implies that genetic gain could be rapidly increased (Table 3). In the PS program, the PGWC of selected plants based on agronomic traits was decreased significantly, from 38.09% in C0S0 to 19.73% in C2S0; that is, PGWC was almost reduced by half in two years (Table 3). The improvement in PGWC might be attributed to accurate measurement and selection of this trait, a large scale of genetic recombination and the elimination of undesirable alleles for the next recurrent selection cycle [43,44,45]. Compared to PS, the genetic gain in PGWC of the MAS program was slower. The PGWC of MAS-selected plants was 29.76% in average, whereas that of PS was only 24.63%. However, the gap between PS and MAS was narrowed to 3.08% and 1.03% in C2S0 and C3S0, respectively. The difference in PGWC between the two methods was significant in C0S0, C1S0 and C2S0 as well. However, in C3S0, no significant difference was observed. In other words, phenotypic selection was more efficient at early stages of recurrent selection for low-chalkiness germplasm enhancement than MARS. After a few cycles, the genetic gain was similar. This may be mainly attributed to the higher selection accuracy of PS, since it was based directly on PGWC values. This value of each plant not only reflects the effects of the whole genome, but also includes the interaction between genes and the environment. However, there were only few markers considered at first; that is, the predicted value could not completely represent the actual value. In addition, chalkiness was highly susceptible to environmental factors, especially at the single-plant level [1,8,9,10].

This study found that both PS and MARS exhibited high efficiency in improving chalkiness. However, when other agronomic or quality traits of advanced lines were investigated, both methods showed limitations. For example, the days to maturity of advanced lines of C0S0 in PS were 133.65 d on average, which increased to 134.65 d in C3S0. However, that of MARS ranged from 131.72 d to 131.75 d; that is, the DTM under MAS was quite stable (Table 3). This indicates that there are more and more late-maturing plants selected cycle by cycle since delayed heading dates provide a greater chance to escape high-temperature stress during grain filling, thereby ultimately leading to lower chalkiness. The results above suggest that MAS has some advantages in low-chalkiness line development, since the influences of environmental conditions and cultural practices, especially during grain filling, could be overcome to a certain degree [8,9,10]. However, in the MAS program, using the prediction model, lines with intermediate AC and high AC disappeared after three cycles. It is not strange that GBSSI is involved in all molecular models for prediction (Table 2). The average AC of the G allele of the first intron/exon junction site of GBSSI was 24.32%, while that of the T allele was 16.75% [18]. On the other hand, the two alleles (C and A) of the sixth exon of GBSSI had a very similar AC (23.25 and 22.94%, respectively). The genotype G/C at these two loci had a higher AC than that of the G/A genotype. Therefore, the two loci of GBSSI were major contributors to AC, with the genotypes of G/A, G/C and T/A having, respectively, high, intermediate and low AC [46]. Some studies have reported that AC was associated with the formation of grain chalkiness [18,19,20,31,47]. We previously found that the low AC lines was lower in DEC than the lines with intermediate or high AC in the present population and a population of 239 indica varieties released in China [10,48]. Our findings in the present study provide a genetic explanation of the often-observed positive correlation between AC and chalkiness. AC, DEC and PGWC of the lines with the T allele of GBSSI1IF was significantly lower than those with the G allele. The above results suggest that GBSSI had pleiotropic effects (additive and epistasis) on AC and chalkiness, which caused low AC to exhibit the lowest chalkiness, while the genotype with intermediate AC had the highest chalkiness; therefore, almost no HAC or IAC plants remained after few cycles of MAS, since these two alleles were excluded during random recombination for the next cycle. Therefore, this might be the main reason why the frequency of the lines with a low AC increased from 48.33% in C0S0 to 71.67% in C3S0, while that of IAC and HAC decreased in the PS program (Table 3). Except for low chalkiness, preferences for eating quality characteristics vary across countries and regions. Grain shape, especially grain length and width significantly affected chalkiness [22,23,24]; therefore, MAS for these known genes, combined with visual selection for grain shape, could be a more efficient method for breeding low-chalkiness lines with acceptable eating quality (AC) and grain shape in rice.

5. Conclusions

This study explored the epistatic interactions between molecular markers associated with the percentage of grain white chalkiness (PGWC) and the degree of endosperm chalkiness (DEC) and evaluated the applicability of marker-assisted recurrent selection (MARS) for breeding rice with low chalkiness. GBSSI and SSIIa were consistently identified in optimal models, indicating their critical roles in regulating rice chalkiness. The incorporation of digenic epistatic effects significantly improved the adjusted coefficient of determination (R²) of the best predictive models across both experimental environments. Our results confirmed that MARS is an efficient strategy for reducing chalkiness and enhancing eating quality. Nevertheless, phenotypic selection may remain more effective for breeding rice varieties with intermediate or high amylose content.