植物育種，作為一門(mén)歷史悠久且關(guān)鍵的學(xué)科，長(zhǎng)期以來(lái)被視為“科學(xué)與藝術(shù)的結(jié)合”。然而，在短周期作物（如玉米、小麥、水稻）為代表的現(xiàn)代育種中，這一傳統(tǒng)的浪漫化描述正被徹底改寫(xiě)。現(xiàn)代數(shù)據(jù)驅(qū)動(dòng)的育種方式挑戰(zhàn)了“藝術(shù)”的地位，改變了以往依賴直覺(jué)和經(jīng)驗(yàn)的育種方法，將“藝術(shù)”還原為可量化、可預(yù)測(cè)的科學(xué)，并使育種過(guò)程變得更加系統(tǒng)化與工程化。
時(shí)代的變遷：從“少數(shù)地塊”到“數(shù)千個(gè)地塊”
傳統(tǒng)“育種家眼睛”的歷史背景
傳統(tǒng)上，育種家的視覺(jué)判斷在育種工作中至關(guān)重要，尤其是在早期階段的視覺(jué)篩選（1960年代）。在育種項(xiàng)目初期，種質(zhì)資源高度多樣化（如綠色革命時(shí)期的水稻），面對(duì)幾十到幾百個(gè)地塊，育種家依靠眼睛和經(jīng)驗(yàn)來(lái)挑選出表現(xiàn)優(yōu)異的植株。
現(xiàn)代育種的挑戰(zhàn)
然而，現(xiàn)代育種項(xiàng)目，如工業(yè)化生產(chǎn)純系（如DH或RGA），涉及數(shù)千個(gè)產(chǎn)量試驗(yàn)地塊。在這種規(guī)模下，主要性狀（如成熟期、株高）趨于一致，視覺(jué)評(píng)估變得幾乎不可能。為了在數(shù)千個(gè)地塊中選擇最佳品種，育種家需要更高精度的數(shù)據(jù)支持——例如，區(qū)分50克產(chǎn)量差異，純靠眼力已無(wú)法勝任。
現(xiàn)代育種：從“藝術(shù)直覺(jué)”到“數(shù)據(jù)洞察”
數(shù)據(jù)驅(qū)動(dòng)的育種方式
“數(shù)據(jù)至上”并不意味著育種家可以待在辦公室里或待在實(shí)驗(yàn)室只靠模擬育種，相反，它重新定義了育種家與田地的互動(dòng)方式?，F(xiàn)代育種家已經(jīng)從依賴經(jīng)驗(yàn)直覺(jué)的“藝術(shù)”轉(zhuǎn)向了通過(guò)數(shù)據(jù)和統(tǒng)計(jì)分析作出決策的“科學(xué)”。
過(guò)去所謂的“育種直覺(jué)”，實(shí)際上是育種家多年積累的經(jīng)驗(yàn)，能夠?qū)?shù)據(jù)進(jìn)行高級(jí)模式識(shí)別，在十年如一日，南繁北育，刻意練習(xí)的結(jié)果?，F(xiàn)代技術(shù)，如基因組學(xué)、數(shù)據(jù)分析和傳感器，可以將這些模式顯性化、可測(cè)試化和可規(guī)?；?，將“直覺(jué)”轉(zhuǎn)化為有數(shù)據(jù)支持的決策。
現(xiàn)代育種的田間觀察
現(xiàn)代育種家的田間觀察，不再是單純尋找“大穗子”或“大玉米棒”，而是帶著統(tǒng)計(jì)學(xué)和數(shù)量遺傳學(xué)的視角：
監(jiān)控非遺傳效應(yīng)：重點(diǎn)觀察對(duì)照組的長(zhǎng)勢(shì)，判斷是否存在空間趨勢(shì)，如不均勻的灌溉、施肥或病害。
排除干擾數(shù)據(jù)：識(shí)別并記錄因管理不善等導(dǎo)致的不適用地塊、行或區(qū)組，在數(shù)據(jù)分析中排除。
補(bǔ)充數(shù)據(jù)分析：確認(rèn)區(qū)組效應(yīng)，驗(yàn)證空間矯正模型的合理性，確保非遺傳效應(yīng)被最小化。
決策的算法化和可復(fù)現(xiàn)性
現(xiàn)代育種流程中的選擇是基于可量化、可預(yù)測(cè)的模型進(jìn)行的，而非個(gè)人“直覺(jué)”。如果兩名育種家擁有相同的環(huán)境、表型和基因型數(shù)據(jù)集及相同的模型，他們將做出相同的選擇。
現(xiàn)代育種的核心：系統(tǒng)工程與可預(yù)測(cè)的遺傳增益
現(xiàn)代精英育種的本質(zhì)是一套工程化、可量化管理的流程。育種的目標(biāo)是優(yōu)化作物性能，而不是單純的表現(xiàn)形式。通過(guò)使用BLUPs、GBLUP、機(jī)器學(xué)習(xí)~和環(huán)境協(xié)變量等工具，育種家能夠?qū)崿F(xiàn)系統(tǒng)化、可預(yù)測(cè)的遺傳增益。自動(dòng)化和標(biāo)準(zhǔn)化的流程正在逐步取代個(gè)體的“匠人精神”。
因此，現(xiàn)代植物育種更應(yīng)被視為一門(mén)嚴(yán)謹(jǐn)?shù)臄?shù)據(jù)科學(xué)與系統(tǒng)工程，旨在通過(guò)量化、可復(fù)現(xiàn)的方法，優(yōu)化和預(yù)測(cè)農(nóng)作物的改良。
智種評(píng)論
這篇文章為國(guó)內(nèi)育種界提供了非常直接、現(xiàn)實(shí)的提醒：現(xiàn)代育種已經(jīng)不再是“師傅帶徒弟”的經(jīng)驗(yàn)藝術(shù)，而是扎扎實(shí)實(shí)的定量科學(xué)與系統(tǒng)工程。
在今天的大規(guī)模、多環(huán)境和強(qiáng)競(jìng)爭(zhēng)的育種體系中，單純依賴“眼力”和“直覺(jué)”已經(jīng)無(wú)法支撐持續(xù)的遺傳進(jìn)步。真正能夠拉開(kāi)差距的是對(duì)數(shù)量遺傳學(xué)、統(tǒng)計(jì)建模、基因組預(yù)測(cè)、試驗(yàn)設(shè)計(jì)、數(shù)據(jù)清洗與空間校正等現(xiàn)代工具的掌握能力。
對(duì)育種從業(yè)者的幾點(diǎn)建議：
田里不能少走，但眼睛要變成“統(tǒng)計(jì)學(xué)家的眼睛”：現(xiàn)代育種家要具備識(shí)別非遺傳效應(yīng)、判斷試驗(yàn)質(zhì)量和理解空間趨勢(shì)的能力，而這些都需要通過(guò)數(shù)據(jù)分析來(lái)支持。
育種直覺(jué)不是“天賦”，而是長(zhǎng)期結(jié)構(gòu)化學(xué)習(xí)的產(chǎn)物。積累的經(jīng)驗(yàn)固然寶貴，但需要與可量化的工具融合，才能放大效能。
現(xiàn)代育種已是“系統(tǒng)工程”，這意味著流程設(shè)計(jì)、試驗(yàn)布局、數(shù)據(jù)質(zhì)量控制、預(yù)測(cè)模型選擇，比“挑穗子”更為重要。
多讀經(jīng)典，回到理論本源：現(xiàn)代育種者應(yīng)不斷深化對(duì)遺傳學(xué)、統(tǒng)計(jì)學(xué)和試驗(yàn)設(shè)計(jì)的理論理解，這些基礎(chǔ)理論是將“藝術(shù)”轉(zhuǎn)化為“科學(xué)”的關(guān)鍵。
推薦經(jīng)典書(shū)籍：
Falconer《Quantitative Genetics》
Lynch & Walsh《Genetics and Analysis of Quantitative Traits》
Bernardo《Breeding for Quantitative Traits in Plants》
Piepho《Experimental Design and Analysis for Plant Breeding》
這些書(shū)籍將幫助育種者更好地理解現(xiàn)代育種體系的底層邏輯，并在數(shù)據(jù)時(shí)代的競(jìng)爭(zhēng)中站穩(wěn)腳跟。這些書(shū)可能對(duì)大家讀起來(lái)有些吃力，建議去讀下《玉米育種學(xué)》第三版，再用AI工具去讀這些經(jīng)典理論，可能受益匪淺。
原文：
In a previous post, I mentioned that “Data is king; breeder’s eye is a myth.” I appreciate those who both agreed and disagreed with that statement. I feel I need to expound a bit on the context of that statement in an elite breeding program involving short-duration row crops.
- a breeding program that is just being established which might use a very diverse set of founders (e.g. 1960s green revolution rice breeding) may rely on visual phenotypic selection of breeders from F2 to field trials to select for obvious qualitative traits from a FEW DOZEN to FEW HUNDRED plots.
- an elite modern breeding program consisting of an industrialized inbred production (DH or RGA) and THOUSANDS of yield trial plots simply does not make sense for the traditional definition of a breeder’s eye. Elite crosses would mean the major traits would have been fixed, and traits such as maturity and height would be almost uniform in an elite population. Visual selection on a trait with a highly continuous variation such as yield will require a resolution of ~50 grams (good luck!). Visual evaluation of thousands of plots in several locations is virtually impossible.
The priority for early-stage trials is to make sure non-genetic effects are minimized. We can redefine the Breeder’s Eye as follows:
- Breeders should be familiar with the checks. evaluate the check plots scattered throughout the field. Vigor of check plots across large trials can give an idea of spatial trends in the field, which can be confirmed when data is analyzed.
- evaluate the range and rows (or columns and rows) and note the presence of range/row/block effects, typically cause by non-uniform irrigation, fertilizer application, disease incidence, and cultural management practices. These observations will complement data analysis.
- Take note to exclude rows, ranges, blocks or plots that are unusable in the analysis of data.
- If populations are blocked, take note of promising populations and pedigrees.
- As you go through the field prioritizing check plots, some entries may catch your attention. It is okay to spend time on those plots tp learn more about the traits, genes present, pedigree, etc. Breeders will do this plot-by-plot evaluation in late-stage trials when entries are lessMODERN, DATA-DRIVEN PLANT BREEDING IS NOT AN ART
I have posted several times on this topic, and each time there are misinterpretations. I am hoping this can be clarified here.
Plant breeding has often been described as also an “art.” Modern data-driven breeding challenges that idea not by removing something mystical, but by revealing that what we long called art was actually misunderstood science and experience.
Misconception 1: “Data-driven breeders don’t go to the field.” This is incorrect. Field observation remains central. Data-driven breeders integrate genomics, analytics, sensors, and digital workflows on top of strong field experience. However, modern breeders look at a field trial with the perspective of statistics and quantitative genetics, while traditional plant breeders look at a field trial searching for big panicles, large ears and other observable phenotypes.
Misconception 2: “Breeder intuition is art.” Not quite. What people once referred to as intuition is actually pattern recognition built from years of exposure to data, even before it was digitized. As Adam Grant explains in The Originals, intuition becomes reliable only when it emerges from repeated, structured learning. Modern tools now make those patterns visible, testable, and scalable, turning “intuition” into transparent, data-backed decision-making.
Misconception 3: “Creativity disappears in data-driven breeding.” Creativity isn’t gone, it has moved. What used to be viewed as the “art” of picking plants is now the engineering creativity of designing superior pipelines: smarter testing schemes, simulations to shape crossing blocks, optimized selection strategies, integrated phenotyping and genomic prediction. This creativity is systematic and measurable, not mystical.
The truth: Modern breeding is not an art. It is a discipline rooted in data, experimentation, and intentional design. What used to be labeled “art” was simply scientific reasoning and pattern recognition before we had the tools to quantify them.
And by the way, different crops and programs would have different stages of transitioning “art” out of breeding activities, and that is okay.WHY MODERN PLANT BREEDING IS NOT AN ART
Plant breeding in elite, data-driven pipelines is not an art because decisions are governed by quantitative genetics, reproducible analytics, and structured workflows aimed at delivering predictable genetic gain, not personal intuition or aesthetic preference.
1. Modern breeding is governed by quantifiable, predictive models, not intuition. Elite pipelines operate on statistical genetics, quantitative trait models, genomic prediction, and algorithmic selection. Decisions are made using heritabilities, GxE models, haplotype effects, and prediction accuracies, not on personal “feel” or aesthetic judgment.
2. Selection is algorithmic and reproducible. In advanced programs, if two breeders have the same datasets (phenotypes, genotypes, environmental metadata) and models, they will make the same selection decisions. Reproducibility is a hallmark of science and engineering, not art.
3. Pipelines follow structured, stage-gated workflows. Elite breeding programs run using defined breeding schemes, molecular QC and selection, stage-gate advancement criteria, multi-environment testing, decision-support tools. These are standardized engineering workflows aimed at throughput, quality control, and risk reduction, unlike the free-form nature of artistic creation.
4. Big data and analytics drive advancement decisions and may rely on BLUPs and GBLUPs, machine learning, high-throughput phenotyping, environmental covariates, and many more. None of this is interpretable as “art”; it is data science, statistics, and bioinformatics.
5. The objective is performance, not expression. Art aspires to expression, originality, or aesthetic outcomes. Breeding pipelines aim for measurable genetic gain, product delivery, predictable trait introgression, and improvement in farmer ROI.
6. Outcomes are benchmarked against objective KPIs. Breeders are held accountable for rate of genetic gain, deployment timelines, market share of products, trait delivery metrics and pipeline efficiency. These are engineering KPIs, not artistic judgments.
7. High-performing pipelines reduce individual “craftsmanship”. With automation, analytics, and standardized workflows, crossing decisions use algorithms such as optimal contribution selection and advancement relies on genomic predictions or multi-environment statistical models in later stages. The personal “touch” disappears, the system produces the gain; the individual does not “craft” varieties.
8. Complex traits require systems engineering, not artistic intuition. Yield, adaptation and trait indices involve polygenic architectures, nonlinear GxE interactions, and physiological trade-offs. These require integrated systems biology + data engineering.
9. Success is validated through external empirical testing. New varieties must outperform checks across multi-location trials, show stability, and pass regulatory benchmarks. Art is not validated against empirical performance curves; breeding is.
來(lái)源：Mark Nas等
