Boom5426 commited on
Commit
5f69f4b
·
verified ·
1 Parent(s): c5e7ca7

Delete code/VCDesign_9page_final_revision_spec.md

Browse files
code/VCDesign_9page_final_revision_spec.md DELETED
@@ -1,162 +0,0 @@
1
- # VCDesign:9 页收口的最终修改规格
2
-
3
- **依据:** `main (13).pdf` 与当前 §5.2 间距截图。本文是待执行的修改规格;未修改 LaTeX 源码,未重编译,未重算实验。PDF 文本抽取词数仅用于估计编辑量,不是排版行数或页数保证。
4
-
5
- ## 0. 本轮目标与锁定项
6
-
7
- 目标是在官方模板下,让完整两段 Discussion and conclusion 结束于第 9 页以内。摘要、题目、两段结论文字、三项贡献的划分、Figure 1–4 的尺寸和内容、Table 1 的数据、全部主结果与数据来源保持不变。公式、指标、数据划分、模型权重、超参数选择过程及统计含义不得更改。
8
-
9
- 不以缩小字号、页边距、图宽、图中文字或全局行距解决超页;不修改官方 `.sty`;不使用负 `\vspace`、`\enlargethispage` 或强制缩放掩盖问题。图表尺寸和图内内容需要在编译前后逐一核对。
10
-
11
- ## 1. P0:先定位 §5.2 标题后异常空白
12
-
13
- 当前 PDF 第 7 页:标题文字下缘约为 471 pt,正文首行上缘约为 535 pt,中间约 64 pt。第 10 页还有约 35 行结论内容。空白的高度不能直接当作可回收篇幅。
14
-
15
- 检查 §5.2 标题、Figure 2/3 与下一节附近的手工 `\vspace` / `\vfill`、强制分页、过强的浮动体限制、`\FloatBarrier` / `\Needspace`、不可拆分环境以及页面竖直拉伸。只有源码与日志才能确认具体原因。
16
-
17
- 先移除被确认无必要的局部约束,使标题和正文恢复正常模板间距,图仍与对应讨论保持接近。若原因是页底对齐造成的伸缩空白,不把 `\raggedbottom` 当作默认投稿修复:它可能只是把空白移到页底,并不减少页数。先解决导致页面内容无法继续流动的约束;不要新增全局格式覆盖。
18
-
19
- 在正式文字压缩前编译一次,记录实际正文末页、结论起止页与 §5.2 间距。后续每轮使用这一编译结果,而不是按“预计节省行数”累加。
20
-
21
- ## 2. 撤回上一轮不准确的压缩写法
22
-
23
- - **§5.3:** `+0.0002` 对应 full atlas 与 size-controlled 的比较;`0.0078` 对应 size-controlled 与 own-response-masked 的比较。不能把两者都写成“at matched atlas size”。本轮保留 §5.3 原段落与所有关键数值。
24
- - **§5.4:** 不能把 row-matched 对照写成分离了 measurement volume 与 identity coverage。也不能把 same-identity 的 3M vs M 对照写成相同测量总量下重复测量胜过跨情境分配。保留“固定身份、raw targets、未 masking、每身份三次测量 vs 一次”的真实设计。本轮不重写 §5.4。
25
- - **§5.5:** 四项反向变化包括特征移除,不能说 atlas 和 candidate features 均固定。改为固定 atlas 后调整 basis rank 或 candidate features;保留 effect-only 范围。
26
- - **§5.1:** 保留 unit-direction targets 在同一批 held-context pools 上选择的事实;“外部基线运行前固定”不等于完全独立的评估集选型。
27
- - **Table 1:** masking 只针对 effect branch。保留其信息范围说明及 §3.3 的 68% 暴露说明,不能把 masked 候选统一写成整个模型从未见过。
28
- - **§5.2:** 不删除小预算下对 text ranker 未分离这一条件。主结果与边界应同时保留,不能只留 B=50 的正结果。
29
-
30
- ## 3. A 轮:优先压重复定义和导读,结论与主结果不动
31
-
32
- 下面给出精确替换范围与建议文字。作者年份和 Section/Figure 引用应恢复为项目已有的 citation/reference keys,不得猜测新键。数学符号沿用当前源文件已有宏(尤其预算平均 BU),不要另建不一致的指标命名。
33
-
34
- ### A1. §3 总起句
35
-
36
- 删除重复总起句;保留节标题及 Figure 1 在引言中的引用。
37
-
38
- 定位起句:`VCDesign separates the experimental decision, the scoring mechanism, and the measured utility`。
39
-
40
- 执行:删除以上指定的重复句,不删除后续定义、公式或特征说明。
41
-
42
- ### A2. §3.1 全节文字
43
-
44
- 保留输入、候选集、预算、共享评分器与知识输入;只移除已经在引言解释过的求解路线清单。保留 Mehrjou et al. 的现有 citation key。
45
-
46
- 定位起句:`A query q = (xs, xt) specifies source and target states`。
47
-
48
- A query \(q=(x_s,x_t)\) specifies source and target states, with transition \(d_q=x_t-x_s\). Given a feasible candidate set \(C_q\) and budget \(B\), a solver ranks candidates and selects the first \(B\) for testing (Mehrjou et al., 2022). Our reference scorer \(s(q,c)=f(x_s,x_t,c)\) uses candidate knowledge \(k_c\) available before profiling, allowing the library to vary without a fixed output vocabulary.
49
-
50
- ### A3. §3.2 第一段
51
-
52
- 保留评价作为独立贡献:预算内前缀、越靠前的有用候选受到更高奖励、独立实测效用、多解与求解器无关。以下 CRISPRi 实例、公式和统计量全部不动。
53
-
54
- 定位起句:`Evaluation should reflect the experimental choice`。
55
-
56
- We evaluate the first \(B\) candidates that the budget allows researchers to test, rewarding useful interventions earlier in this prefix. Candidate utility is defined by independently measured held-out outcomes rather than query-identity recovery, so multiple interventions can be useful for one target. This decision-level criterion applies across solvers, while the definition of a useful outcome remains application-specific.
57
-
58
- ### A4. §4 开头
59
-
60
- response-unseen 必须指自身响应不可用于 effect learning,不得改成整个系统从未见过该候选。
61
-
62
- 定位起句:`VCDesign-CED instantiates the framework for the response-unseen setting`。
63
-
64
- VCDesign-CED instantiates the framework where a candidate’s own response is unavailable for effect learning. It combines direct candidate scoring with Candidate Effect Distillation (CED), which learns how deployment-available biological knowledge maps to perturbational effects and aligns the inferred effects with the requested transition (Figure 1b,c).
65
-
66
- ### A5. §4.1 第一段首句
67
-
68
- 仅删除重复定义;保留紧接着的 STRING / MAP-KG 特征描述、引用以及之后所有公式和 raw/unit 配置说明。历史响应监督在后文的训练响应与 ridge 方程中仍完整给出。
69
-
70
- 定位起句:`Candidate Effect Distillation (CED) uses measured responses as supervision for a mapping`。
71
-
72
- 执行:删除以上指定的重复句,不删除后续定义、公式或特征说明。
73
-
74
- ### A6. §5.1 第一段
75
-
76
- 保留屏幕来源、三个真正留出情境、候选规模、跨全部训练情境遮蔽、评估情境不参与训练、检索定义、外部方法边界。附录 K562 控制仍有入口,不重复其次要数字。
77
-
78
- 定位起句:`We use CRISPRi Perturb-seq screens from K562 and RPE1`。
79
-
80
- We use CRISPRi Perturb-seq screens from K562 and RPE1 (Replogle et al., 2022) and HepG2 and Jurkat (Nadig et al., 2025). K562 provides disjoint training, development and locked evaluation identities; RPE1, HepG2 and Jurkat are held out in turn, with candidate pools of 1,273, 764 and 1,450 (Section A). Five-fold cross-fitting removes each evaluated candidate’s responses from all training contexts; the measured setting retains them. The evaluation context never supplies training responses (Section D.10). Profile retrieval matches historical responses to the requested transition by cosine; CellNavi and GEARS adaptations are detailed in Section D.6. No optimization solver is instantiated. K562 context-transfer controls are reported in Section D.13.
81
-
82
- ### A7. §5.1 第二段
83
-
84
- 不是为了省字删除选型来源。必须保留同一 held-context pools 上选择 unit targets,以及外部基线前固定。恢复简明统计读法,避免读者误把差值中位数当作中位数之差。
85
-
86
- 定位起句:`The locked K562 comparison uses prespecified raw-target CED`。
87
-
88
- The locked K562 comparison uses prespecified raw-target CED and \(\overline{\mathrm{BU}}\), fixed before scoring. For held contexts, unit-direction targets were selected on the same pools through a preregistered comparison; this choice and BU@20 were fixed before external-baseline runs (Section E.1). Levels and paired gains are medians, except that hit-count gains are means; 95% cluster-bootstrap intervals keep trained models fixed (Section C).
89
-
90
- ### A8. Table 1 导读段
91
-
92
- 保留 Table 1 作为主结果地图,不再逐字解释 upper/lower block;caption 与表本体保留,尤其 68% 暴露范围和 GEARS/CellNavi 脚注不删。
93
-
94
- 定位起句:`Table 1 reveals a first organizing principle`。
95
-
96
- Table 1 compares the evaluated computational routes under a common decision objective while distinguishing their candidate-response requirements. VCDesign-CED leads the evaluated response-optional solvers on held-out identities and in masked held-out contexts, whereas profile retrieval performs best when candidate responses are measured elsewhere. We next examine the source of response-unseen gains and the use of measured responses in mixed libraries.
97
-
98
- ### A9. §5.5 全节文字
99
-
100
- 保留三个相关系数、分层置换检验、4/5 反向变化、每个 held context 都成立,以及 effect-only 范围。不能写 candidate features fixed,因为变化包含 drop MAP-KG / STRING。
101
-
102
- 定位起句:`Prediction accuracy is informative for design-model selection only insofar`。
103
-
104
- Across masked atlas and input configurations, candidate-specific prediction accuracy and effect-only design utility are positively rank-correlated within each held context (Spearman \(\rho=0.31,0.69,0.71\); stratified permutation \(p<10^{-4}\); Figure 4c). Within a fixed atlas, however, four of five changes to basis rank or candidate features move prediction and design in opposite directions in every held context. Thus, prediction accuracy does not reliably select the effect predictor that yields the better finite-budget ranking.
105
-
106
-
107
- ## 4. B 轮:A 轮重编译仍超页时,等义压缩引言重复叙述
108
-
109
- 只替换引言第三、第四段,保留四段结构、首两段、三项贡献和全部 Results 数据。这里改变表达长度,不重新定义已经确���的任务与贡献。不动摘要和结论。
110
-
111
- ### B1. 引言第三段:仅在 A 轮编译后仍超页时使用
112
-
113
- 保留首次表述的原始范围、输入输出、独立评价贡献、context/response 区别与 CED。不重新讨论选题,不增加任何首创主张。
114
-
115
- 定位起句:`Here we introduce VCDesign (Figure 1)`。
116
-
117
- Here we introduce VCDesign (Figure 1), which ranks a feasible intervention library given source and target states and budget \(B\), prioritizing the \(B\) candidates to test. To our knowledge, this is the first explicit formulation of target-conditioned cellular intervention design as finite-budget ranking over a variable candidate set. Its decision-aligned evaluation scores this executable prefix using independently measured held-out outcomes, rewarding multiple useful interventions and comparing different computational routes under one downstream objective. VCDesign also distinguishes transfer to a new cellular context from ranking candidates whose own responses are unavailable for effect learning. For the latter, VCDesign-CED uses historical measurements to learn a mapping from deployment-available biological knowledge to perturbational effects, then combines their alignment with the requested transition with direct candidate scoring.
118
-
119
- ### B2. 引言第四段:与 B1 同轮执行
120
-
121
- 主要发现和 0.049、1.11 均保留;明确这些汇总数字来自三个 held-out contexts,而 K562 identity 实验只共享增益方向。删掉跨句重复推论,不删 Results 中任何主结果。
122
-
123
- 定位起句:`We evaluate this framework across four CRISPRi Perturb-seq contexts`。
124
-
125
- Experiments across four CRISPRi Perturb-seq contexts show that historical perturbation evidence can improve finite-budget selection across candidate identities and cellular contexts. In three held-out contexts, masking candidates’ own responses from effect learning still yields a pooled BU@20 gain of 0.049 over the same base ranker and 1.11 additional top-5% candidates among 20 selections on average; held-out K562 identities also show gains. When historical candidate responses are available, profile retrieval performs better, and availability-based routing improves mixed-library selection. Controlled comparisons show gains from additional perturbation measurements and losses after removing candidate knowledge, whereas additional contexts at matched measurement volume and the tested richer effect maps offer no comparable benefit. Prediction accuracy is associated with design utility across atlas and input changes, but does not reliably order nearby configurations. Together, these results establish a task and evaluation foundation for using virtual-cell models to prioritize experiments under finite budgets.
126
-
127
- ## 5. 本轮明确不动的内容
128
-
129
- - §3.3 关于 effect-branch masking 与固定 base scorer 的区别。
130
- - §5.2 的 held-context 与 held-identity 主结果、配对增益及置信区间、B=50 的 hit-count 增益与小预算条件、realized-outcome 的未通过预注册中位数标准。
131
- - §5.3 的三个不同 masking 对照、混合库 routing 的主增益,以及全测量情况下 fusion 仍劣于 retrieval 的结果。
132
- - §5.4 的 atlas 与身份覆盖混杂说明、same-identity 对照的真实条件、非线性模型的负结果和知识移除结果。
133
- - Figure 1–4 / Table 1 本体、全部公式及两段结论。Table 1 caption 与脚注本轮也保留,避免为了少量行数削弱信息范围说明。
134
-
135
- ## 6. 编译与验收
136
-
137
- 1. 分轮提交:先版式修复,再 A 轮文字替换;仅在仍超页时执行 B 轮。每轮都完成正常构建流程,确保交叉引用收敛。
138
- 2. 比较修改前后:题目与摘要不变、结论逐字不变、图的显示尺寸与图内内容不变、主表数字不变、所有主要效果量及置信区间不变。
139
- 3. 保留中位数、配对中位数、hit-count 配对均值的区别;不对任何 median levels 做减法来替代 paired median gains。
140
- 4. 看渲染而不只看日志:§5.2 标题与正文恢复正常距离;无漂浮标题、图文重叠、裁切、图注脱离图、超出页边界文字或转移到其他段落的大空白。
141
- 5. 报告完整科学正文的实际末页,不能用删除/移动必要科学内容到声明或附录的方式获得“9页”。声明的摆放和计页以当年官方模板与指南为准。
142
- 6. 页数达标即停止。若 B 轮后仍超页,报告实际剩余内容高度及浮动体阻塞位置,不得自行继续删主结果、改结论或缩图。
143
-
144
- ## 7. 词数变化(仅供规划)
145
-
146
- 按 PDF 抽取的文字与建议段落,以空白分隔计词;数学和引用的计数与 LaTeX 排版不同,不构成节省行数承诺。
147
-
148
- | 改动 | 原段词数 | 新段词数 | 净减少 |
149
- |---|---:|---:|---:|
150
- | A1 | 20 | 0 | 20 |
151
- | A2 | 128 | 56 | 72 |
152
- | A3 | 111 | 57 | 54 |
153
- | A4 | 79 | 45 | 34 |
154
- | A5 | 18 | 0 | 18 |
155
- | A6 | 161 | 107 | 54 |
156
- | A7 | 63 | 60 | 3 |
157
- | A8 | 107 | 59 | 48 |
158
- | A9 | 134 | 69 | 65 |
159
- | B1 | 210 | 123 | 87 |
160
- | B2 | 226 | 146 | 80 |
161
-
162
- A 轮约��少 368 个词;B 轮再减少 167 个词;合计约 535 个词。异常空白是否能转化为页数收益,只能以源码修复后重编译结果判断。