怎样比较寻找遗传互作的策略？ How should strategies for finding genetic interactions be compared?
几种策略各自决定检验哪些 SNP 对或基因对。要判断哪种更好，需要固定哪些条件、比较哪些数字？这个网站用拟南芥的数据回答这个问题。开始面板练习 → Each strategy decides which SNP pairs or gene pairs to test. To judge which works better, what must be held fixed and which numbers should be compared? This site answers the question with Arabidopsis data. Start the panel exercise →
鼠标移到弧线上查看是哪一对，点击打开详情。Hover over an arc to see the pair; click to open it.
936 个品系的现蕾时间从 23.5 天到 139 天不等Time to visible buds ranges from 23.5 to 139 days across 936 lines
数据来自全球采集的 936 个拟南芥自然品系。性状 DTF1 是 16 °C 下从播种到莲座中央出现可见花蕾的天数（AraPheno 703），平均 64.5 天。下文简称"现蕾时间"。 The data cover 936 natural Arabidopsis lines collected worldwide. The trait, DTF1, is the number of days at 16 °C from sowing until visible flower buds appear in the centre of the rosette (AraPheno 703); the mean is 64.5 days. Below it is called "time to buds".
DNA 上某个位置在不同品系之间只差一个碱基，这样的位置叫 。拟南芥品系几乎完全纯合，每个 SNP 在一个品系里只有两种状态，记作 0 和 1。 A position where lines differ by a single base is a . Arabidopsis lines are almost fully inbred, so each line carries one of two states at each SNP, written 0 and 1.
常规的 一次检验一个 SNP。这篇论文研究两个 SNP 之间的：一个位点的效应随另一个位点的状态而改变。 A standard tests one SNP at a time. This thesis is about between two SNPs: the effect of one site changes with the state of the other.
两个 SNP 把品系分成 00、01、10、11 四组，本站称为。如果两个位点的效应只是相加，第二个位点在两种背景下引起的现蕾时间变化相同；有互作时，这两个变化不同。两者之差叫交互对比，检验互作就是检验它是否为 0。 Two SNPs split the lines into four groups, 00, 01, 10 and 11: the . If the two effects simply add, switching the second site changes time to buds by the same amount in both backgrounds; with an interaction the two changes differ. Their difference is the interaction contrast, and testing for interaction means testing whether it is zero.
已知的例子是 FRI 和 FLC：FRI 提高 FLC 的表达，FLC 推迟开花，两者都有功能时开花明显推迟。图中第三种情形按这个模式画，数字是示意值，不是测量值。 A known example is FRI and FLC: FRI raises FLC expression and FLC delays flowering, so flowering is clearly late only when both are functional. The third case in the figure follows this pattern with illustrative numbers, not measurements.
SNP 对的数目约为 SNP 数的平方的一半The number of SNP pairs is about half the square of the SNP count
任意两个 SNP 都可能互作。检验次数越多，单凭偶然得到很小 p 值的机会越大，所以判定门槛要随检验次数变严。最简单的做法是 ：用 0.05 除以检验次数。它有两个前提：零假设下的 p 值有效，检验族事先定好。所以分母是一个面板上全部合格的对，不是最后复检的 1,000 对；模型本身校准不足时（例如 Step 1 的 GLM），换门槛也补救不了。 Any two SNPs might interact. The more tests, the more likely a very small p-value appears by chance, so the bar has to tighten with the number of tests. The simplest rule is the : divide 0.05 by the number of tests. It rests on two conditions: p-values are valid under the null, and the family of tests is fixed in advance. So the denominator is every eligible pair on a panel, not the 1,000 pairs re-tested at the end, and a stricter bar cannot rescue a model that is badly calibrated to begin with (such as the GLM in Step 1).
是从 181 万个常见 SNP 中按连锁程度精简出的子集，Step 2 有说明。正式门槛只计四格人数合格的对：r² 0.2 面板为 0.05 ÷ 1,742,876,507 ≈ 2.87 × 10⁻¹¹。 A is a subset thinned from the 1.81 million common SNPs by linkage; Step 2 explains how. The formal bar counts only pairs with enough lines in every cell: for the r² 0.2 panel it is 0.05 ÷ 1,742,876,507 ≈ 2.87 × 10⁻¹¹.
各策略的区别在于先检验哪些对The strategies differ in which pairs they test
在同一个 SNP 面板上，A1、A2 和 B 用同一个检验模型和同一个门槛，区别只在送去检验的对。C1 检验的是基因对，用多自由度的联合检验，有自己的分母和门槛，不能和 SNP 对策略直接排名。On a given SNP panel, A1, A2 and B use the same test model and the same bar; they differ only in which pairs are tested. C1 tests gene pairs with a multi-degree-of-freedom joint test and has its own denominator and bar, so it cannot be ranked directly against the SNP-pair strategies.
边际筛选（GLM）Marginal screening (GLM)
先做普通 GWAS，挑出单独就有信号的 SNP，再两两配对检验。在本轮设置下计算量最小；单独没有信号的互作会漏掉。Run a standard GWAS, keep SNPs with a signal on their own, and test them in pairs. The cheapest in this round's settings; interactions with no single-SNP signal are missed.
边际筛选（混合模型）Marginal screening (mixed model)
与 A1 相同，但单 SNP 扫描加入了品系间的亲缘关系，挑出的 SNP 更少。As A1, but the single-SNP scan includes relatedness among lines, so fewer SNPs are kept.
全基因组 SNP 对扫描Genome-wide pair scan
不参考单 SNP 信号。先对面板内全部 SNP 对做快速初筛，再检查四格和秩，最后按预算用正式模型复检前 1,000 对。在本轮设置下计算量最大。Ignores single-SNP signals. All SNP pairs in the panel go through a fast screen, then the four-cell and rank checks, and the top 1,000 within the budget are re-tested with the full model. The most computing in this round's settings.
基因对检验Gene-pair test
把一个基因内的 SNP 压缩成几个主成分，再对两个基因的主成分乘积做联合检验，自由度按乘积项的秩确定。1a 只做了 35 个基因的小例子和已知真值的 oracle 检验（基因 ± 2 kb）；全基因组的规模只用于门槛分母和成本估算。2026-11 起暂定改为 294 个开花基因、基因 ± 500 bp；部分条件待商定，导师尚未确认，新流程还没有运行。Compress the SNPs in each gene into a few principal components, then jointly test the products of two genes' components, with degrees of freedom set by the rank of the product terms. In 1a only a 35-gene example and an oracle test with known answers were run (gene ± 2 kb); the genome-wide scale was used only for the bar's denominator and cost estimates. Since November 2026 the working plan is 294 flowering genes with ± 500 bp; some details are open, the supervisor has not confirmed it, and the new pipeline has not run.
只用编码区（已去掉）Coding regions only (dropped)
编码区内的常见 SNP 太少，每个基因中位数 7 个，FLC 只有 1 个。1a 试跑后不再保留，论文会说明原因。Coding regions hold too few common SNPs: a median of 7 per gene and only 1 in FLC. Dropped after the 1a trial; the thesis will give the reasons.
用已知答案的模拟数据评价各策略Strategies are scored on simulated data with known answers
真实数据里哪些互作是真的并不知道，所以要用模拟：保留 936 个品系的真实基因型，指定一对 SNP 产生互作，生成模拟的开花时间，再让各策略去找。答案已知，就可以逐项计分。 Which interactions in the real data are true is unknown, so we simulate: keep the real genotypes of the 936 lines, make one chosen SNP pair interact, generate simulated flowering times, and let each strategy search. With the answer known, each step can be scored.
真实的互作对是否在策略能检验的范围内。不在范围内的不可能被找到。Whether the true pair lies inside what the strategy can test. A pair outside that range can never be found.
在覆盖到的前提下，最后越过门槛的比例。Among covered pairs, the share that finally clears the bar.
覆盖率 × 条件检出率。只报告条件检出率，会高估覆盖率低的策略。Coverage × conditional detection. Reporting only conditional detection overstates strategies with low coverage.
每个模拟性状含 1 对互作 SNP，强、弱各 40 个。A1、A2、B 是完整流程，可以逐个性状配对比较。虚线下方的 C1 是旧试用规格的 （"已知答案再去验证"的对照）：已知答案后直接检验真实的基因对，没有经过搜索，检验和门槛也不同，所以不参与排名。 Each simulated trait contains one interacting SNP pair, 40 strong and 40 weak. A1, A2 and B are full pipelines and can be compared trait by trait. C1, below the dashed line, is the old trial spec's (a check made with the answer in hand): the true gene pair was tested directly once the answer was known, without a search and with a different test and bar, so it is not ranked.
无互作模拟：估计误报Null simulations: estimating false positives
另外生成 200 个不含互作的模拟性状（S0），检查 B 在两个面板上会不会报出假发现。一个性状只要整套搜索中有一对越过门槛，就算一次假发现；目标是这个比例不超过 5%，即。B 在两个面板上都是 8/200 = 4%，95% 1.7%–7.7%。这批 S0 没有检出超出 5% 的证据，但区间上界 7.7%，不能证明不超过 5%。S0 性状仍含加性背景效应，只是没有互作。 We also generate 200 simulated traits without interaction (S0) and check whether B reports false findings on the two panels. A trait counts once if any pair in the whole search clears the bar; the target is that this share stays at or below 5%, the . B scored 8/200 = 4% on both panels, 95% 1.7%–7.7%. This S0 batch shows no evidence of exceeding 5%, but with an upper bound of 7.7% it does not prove the rate is at most 5%. S0 traits still carry additive background effects; they lack only the interaction.
第一轮（1a）已推进到策略比较的试填Round 1 (1a) has reached a trial fill of the strategy comparison
试跑指流程能从头到尾运行，但参数没有冻结。计划在第 48 周（11 月 23 日前后）由用户和导师冻结规则，之后开始正式分析。A trial run means the pipeline runs end to end but the settings are not frozen. The plan is for the student and supervisor to freeze the rules in week 48 (around 23 November), with the formal analysis after that.
数据准备Data preparation
样本、SNP 面板、亲缘矩阵、模拟性状从哪里来Where the samples, SNP panels, kinship and simulated traits come from
研究：1a 完成Research: 1a done页面已上线Page live打开 →Open → 1单 SNP 扫描Single-SNP scan
GLM 和 MLM，为 A1、A2 提供候选GLM and MLM, supplying candidates to A1 and A2
研究：GLM 主成分个数待定Research: GLM PC count open页面已上线Page live打开 →Open → 2互作检测Interaction search
各策略的流程和误报检查How each strategy runs and how false positives are checked
研究：面板与 C 规格待定Research: panel and C spec open页面已上线Page live打开 →Open → 3互作的贡献What interactions add
3a：双核模型里 K∘K 的附加方差分量（全局估计，与策略无关）；3b：各策略选出的互作项能否提高折外预测3a: the extra variance component of K∘K in a two-kernel model (a global estimate, independent of strategy); 3b: whether each strategy's interaction terms improve out-of-fold prediction
研究：1a 完成Research: 1a done页面已上线Page live打开 →Open → 4生物学解释Biological interpretation
候选位点附近的基因和注释，以及注释能说明什么Genes and annotations near the candidates, and what they can show
研究：1a 示例完成Research: 1a examples done页面已上线Page live打开 →Open → 5策略比较Comparing strategies
用 80 个模拟性状完成一次面板比较One panel comparison on 80 simulated traits
研究：试填，正式比较未做Research: trial fill; formal comparison not done页面已上线Page live开始练习 →Start →六个步骤都有了页面。遗传学基础知识可以先看从番茄到碱基。All six steps now have pages. For genetics background, see From Tomato to Base Pair.
研究进度总览（2026-11-05 更新）Research status (updated 5 Nov 2026)
研究证据截至 2026-11-04：最近一次运行是 11 月 2 日的 r² 0.5 面板 S0 × 200 校准，最近一次整理是 11 月 4 日的 Step 5 比较表试填 v2。下面四栏分别是已经做完的、计划要做的、等谁决定的，以及因为还没决定而暂停的工作。Evidence is current to 4 Nov 2026: the latest run was the S0 × 200 calibration on the r² 0.5 panel on 2 Nov, and the latest write-up was version 2 of the Step 5 trial comparison table on 4 Nov. The four boxes list what is done, what is planned, who decides what, and what is on hold until those decisions are made.
- Step 0：数据、P1 和三个 P2 面板、两批模拟性状Step 0: data, P1 and the three P2 panels, two batches of simulated traits
- Step 1：GLM、MLM 扫描和 A1、A2 的候选Step 1: GLM and MLM scans, A1 and A2 candidates
- Step 2：A1、A2、B（r² 0.2）在 DTF1 和模拟上的完整流程；B（r² 0.5）在模拟上的完整流程和 S0 × 200 校准Step 2: full A1, A2 and B (r² 0.2) pipelines on DTF1 and on simulations; full B (r² 0.5) pipeline on simulations with the S0 × 200 calibration
- Step 3：M0、M1、M2 的折外预测和 3a 双核模型Step 3: out-of-fold prediction with M0, M1, M2 and the 3a two-kernel model
- Step 4：两个注释示例；Step 5：比较表试填Step 4: two annotation examples; Step 5: trial comparison table
"完成"指流程跑通并有记录，数字都是工作值，不是正式结果。“Done” means the pipeline ran and is recorded; the figures are working values, not formal results.
- 第 48 周（11 月 23 日前后）：冻结规则Week 48 (around 23 Nov): freeze the rules
- 第 49 周：A1、A2、B 的正式模拟；真实数据 DTF1、FT10 上的 A1、A2、BWeek 49: formal simulations for A1, A2 and B; A1, A2 and B on the real DTF1 and FT10 data
- 第 50 周：C 的规格定下后，在真实数据上运行 C，并做交叉验证Week 50: once the C specification is settled, run C on the real data and cross-validate
- 第 51 周：少量重复的完整 C 模拟；Step 2 收尾；方法部分草稿Week 51: a few full C simulation replicates; wrap up Step 2; draft the Methods
日程会随冻结结果调整。The schedule will move with the freeze decisions.
- 导师：SNP 对的合格规则（四格人数、是否加稳健性要求）、门槛的判定规则、GLM 的主成分个数Supervisor: eligibility rules for SNP pairs (cell size, any robustness requirement), the rule for crossing the bar, the number of GLM principal components
- 作者和导师：P2 面板选 r² 0.2 还是 0.5；模拟情景和重复次数，包括是否加一个因果位点对取自开花基因的情景Author and supervisor: P2 panel r² 0.2 or 0.5; simulation scenarios and replicates, including whether to add a scenario with causal pairs drawn from flowering genes
- 作者：B 的实现、各策略的候选数、预测协议、随机对照是否保持端点复用结构Author: B's implementation, candidate numbers per strategy, the prediction protocol, whether random controls keep the reuse structure of end points
- C 的检验规格：只用开花基因、基因 ± 500 bp 是暂定规格，导师未确认；其余条件仍在商定C's test specification: flowering genes only and gene ± 500 bp are a provisional spec not yet confirmed by the supervisor; the remaining conditions are still being discussed
- B（r² 0.5）在 DTF1 上的检验、预测和完整成本：等 P2 面板B (r² 0.5) on DTF1, its prediction and full cost: waiting for the P2 panel
- 新 C 的任何检验：等 C 的规格Any test of the new C: waiting for the C specification
- 正式模拟和正式的策略比较：等第 48 周冻结Formal simulations and the formal strategy comparison: waiting for the week-48 freeze
- 富集检验：1a 没有稳健的对，按工作流改为逐个写成案例Enrichment tests: round 1a has no robust pair, so the workflow switches to case write-ups
把各页写下的解释汇总起来带走Collect what you wrote on each page and take it with you
每个 Step 页面的末尾有一道写作题，Step 5 是第 7 题和它的自评。你写的内容只保存在这个浏览器里，换设备或清除浏览器数据后会丢失。导出的 Markdown 带网站版本号和研究证据日期，方便以后对照。Each Step page ends with a writing task; for Step 5 it is task 7 and its self-assessment. What you write is saved only in this browser and is lost on another device or after clearing browser data. The exported Markdown carries the site version and the evidence date so you can compare later.
| 页面Page | 状态Status | 已写字数Characters written |
|---|