Dossier · Biomni Lab × JQ1/BRD4 BD1

Abstract · 摘要

JQ1 on BRD4,
redocked, benchmarked,
extended — and then
disputed by its own author.

Biomni Lab 在未获任何提示的情况下，围绕 BRD4 BD1 bromodomain (PDB 3MXF) 与配体 JQ1 完成了一轮完整的计算药物化学工作流：结构预处理、对接 pipeline 搭建、20 化合物 benchmark、LOO 交叉验证、五个新类似物设计，并在最后一个 cell 里主动起草了一份 confidence attribution 表，把自己关于 BD1/BD2 选择性的最强结论置信度评为 LOW。

这份档案完整保留其 41 个 cell 的流程、四次错误恢复与所有关键数值，以期回答一个问题：一个 autonomous agent 在计算药化这样强调"老手直觉"的领域里，能诚实到什么程度。

Headline Figures · 核心数据

1.218 Å

JQ1 Redock RMSD
(MCS, 30 heavy atoms)

0.603 r

Vina vs pKi (n=20)
within lit. 0.3–0.7 range

0.98 pKi

LOO-CV RMSE
linear regression

4 / 4

Runtime errors
self-recovered

20 / 20

ChEMBL compounds
successfully docked

Novel analogs proposed
with agent-rated confidence

01PART ONE

从 PDB 3MXF 出发

Cells 0 – 4 · BioPython · NeighborSearch · H-bond analysis

Cell 0

complete

t + 00:00

Download 3MXF, parse chains, enumerate HETATM residues and altlocs.

下载 3MXF.pdb，用 BioPython 解析链、HETATM 残基，以及关键残基的 altloc 备选构象。

# Chain A summary Chain A: 127 std residues · 208 waters · 214 HETATM # HETATM residues Chain A JQ1 ('H_JQ1', 1, ' ') ← target ligand Chain A IOD ('H_IOD', 169, ' ') Chain A EDO ('H_EDO', 170, ' ') × 3 (cryoprotectant) Chain A DMS ('H_DMS', 171, ' ') # Altloc conformers (residues with dual side-chain states) SER51·CB / OG altloc=A LYS72·CB→NZ (5 atoms) altloc=A THR103·CB / OG1 / CG2 altloc=A GLN127·CB→NE2 (5 atoms) altloc=A

Agent observation 无需提示，agent 自行记录了四个 altloc 残基位置，为后续构象选择（保留 altloc=A）埋下伏笔。

Cell 2

complete

t + 02:14

Verify sequence integrity, identify pocket residues, locate JQ1 centroid and nearby waters.

验证序列完整性、口袋残基识别、JQ1 质心与邻近水分子定位。

Residue range: 42 – 168 (gaps: none) JQ1 atoms: 31 JQ1 centroid: (28.75, 15.83, -2.34) Waters within 8 Å of JQ1 centroid: 11 HOH 211 dist = 5.40 Å HOH 223 dist = 5.77 Å ← later flagged as conserved HOH 337 dist = 6.53 Å HOH 174 dist = 6.92 Å ... Altloc residues' CA distance to JQ1 centroid: SER51 26.8 Å → far, not pocket-relevant LYS72 21.4 Å THR103 17.0 Å GLN127 21.5 Å → all excluded from pocket

Cell 4

complete

t + 03:02

Map water-mediated H-bond network; define the docking box; enumerate pocket-lining residues.

水分子介导的氢键网络测绘；对接盒几何定义；口袋衬里残基完整清单。

=== Pocket Water H-bond Analysis === Criterion: donor/acceptor distance ≤ 3.5 Å HOH 223 (5.77 Å from JQ1): → JQ1·OAG 3.21 Å ← direct to ligand → ASP 144·OD2 2.87 Å → ASN 140·OD1 2.82 Å → HOH 333·O 2.84 Å → HOH 264·O 3.41 Å === Docking Box === Center: (28.47, 15.04, -2.52) Size: 22 × 22 × 22 Å (standard for BRD4 BD1 shallow pocket) === Pocket-Lining Residues (≤ 5 Å of any JQ1 atom) === TRP81 PRO82 PHE83 GLN85 VAL87 LEU92 LEU94 TYR97 CYS136 TYR139 ASN140 ASP145 ILE146 MET149

Domain judgment · 领域判断 Agent 识别 HOH 223 是桥连 Asn140/Asp144/JQ1 的关键保守水，但没有把它塞进对接盒——计算药化圈内的"know but don't abuse"判断，后续受体制备中仅在 receptor_clean.pdb 里保留这一个水分子。

02PART TWO

从 22,285 到 20

Cells 6 – 8 · ChEMBL web client · pChEMBL aggregation

Cell 6

complete

t + 04:30

Query ChEMBL for BRD4 activity records with Ki/IC50 measurements.

从 ChEMBL 拉取 BRD4 的全部 Ki/IC50 活性数据。

Total raw records: 22,285 Assay-type breakdown: IC50 18,877 Ki 3,408

Cell 8

complete

t + 05:48

Filter by pChEMBL validity, aggregate per compound, stratify by potency, select a balanced 20-compound benchmark spanning ≥ 3 log units.

按 pChEMBL 有效性过滤，每化合物聚合，按活性分层，选出跨度 ≥ 3 个对数单位的 20 化合物基准集。

Records with valid pChEMBL: 14,872 Unique compounds: 8,837 pChEMBL range: 2.16 – 9.52 Stratification: weak (pKi 5 – 6) : 2,641 moderate (pKi 6 – 7) : 2,373 potent (pKi 7 – 8) : 2,316 very potent (pKi 8 +) : 743 Selected: 20 compounds pKi range: 5.26 – 8.91 (span: 3.65 log units)

ChEMBL ID	pKi (exp)	n	± σ	Bin
CHEMBL2393130	5.94	32	0.67	moderate
CHEMBL6027008	5.26	8	0.35	weak
CHEMBL5845072	5.26	8	0.46	weak
CHEMBL5772221	5.26	8	0.00	weak
CHEMBL4648362	5.93	8	1.53	moderate
CHEMBL2017291	6.94	41	0.64	potent
CHEMBL6025219	6.78	−	−	potent
CHEMBL2179387	6.66	−	−	potent
CHEMBL4846805	6.55	−	−	potent
CHEMBL1995703	6.89	−	−	potent
CHEMBL1957266 · JQ1 parent	7.21	−	−	potent
CHEMBL1232461	7.03	52	0.46	potent
CHEMBL3581647	7.79	36	0.43	potent
CHEMBL4078100	7.80	21	0.96	potent
CHEMBL4297454	7.56	21	1.35	potent
CHEMBL3987016	8.52	23	0.65	very potent
CHEMBL3895660	8.79	8	0.09	very potent
CHEMBL4171228	8.09	7	0.49	very potent
CHEMBL3900452	8.91	6	0.35	very potent
CHEMBL3893098	8.29	6	0.04	very potent

03PART THREE

受体制备，四次翻车，一次入场

Cells 10 – 22 · PDBFixer · OpenBabel · Vina 1.2.5 · MCS-RMSD

Cell 10

complete

t + 07:10

Receptor preparation with PDBFixer: missing atoms, hydrogens at pH 7.4, clean chain.

用 PDBFixer 处理受体：补全残缺原子、在 pH 7.4 加氢、仅保留 altloc A、保留 HOH 223。

PDBFixer missing residues: {} PDBFixer missing atoms: 1 residue affected PDBFixer hydrogens added at pH 7.4 Wrote 3MXF_fixed_full.pdb Wrote receptor_clean.pdb (protein + HOH 223, altloc A only) Wrote jq1_crystal.pdb Receptor atom count: ~1060

Cell 12

error → recovered

t + 09:34

First attempt: Meeko for receptor PDBQT conversion.

第一次尝试：用 Meeko 做 PDBQT 转换。

MGLTools prepare_receptor4.py: NOT FOUND ModuleNotFoundError from .chemtempgen import export_chem_templates_to_json ----> import gemmi No module named 'gemmi'

Self-recovery · Cell 14 Agent 没有尝试 pip install gemmi（会触发 compilation），而是改用 OpenBabel 作为备选路径。这是一个计算药化圈内常见的"知道 Meeko 不是唯一选择"的判断。

Cell 14

recovered

t + 10:58

Fallback path: OpenBabel for receptor and ligand PDBQT; RDKit ETKDGv3 for 3D embedding.

备选方案：OpenBabel 做 PDBQT 转换，RDKit ETKDGv3 做 3D 构象生成。

OpenBabel receptor conversion: 1 molecule converted receptor.pdbqt: 85,228 bytes ATOM/HETATM lines: 1,060 JQ1 mol: True, atoms: 31 ETKDGv3 embedding result: 0 (0 = success) Wrote jq1_3d.sdf JQ1 PDBQT: 1 molecule converted jq1.pdbqt: 3,460 bytes

Cell 16

error → recovered

t + 12:20

Run Vina redocking.

运行 Vina 再对接。

Command line parse error: unrecognised option '--log' # Vina v1.2.5 dropped the --log flag; agent must update syntax

Cell 18

recovered

t + 13:15

Vina 1.2.5 with corrected syntax, exhaustiveness = 16.

Vina 1.2.5 正确语法，exhaustiveness = 16，9 个构象。

mode |  affinity | rmsd l.b. | rmsd u.b.
-----+-----------+-----------+----------
   -8.189          0          0
   -7.362       1.056      1.662
   -6.143       3.237      5.584
   -6.091       3.356      5.837
   -5.984       3.977      7.075
   -5.909       2.942      6.116
   -5.893       4.128      7.260
   -5.839       4.280      7.397
   -5.663       3.497      5.970

Best Vina score (JQ1): −8.189 kcal/mol

Cell 20

artefact → recovered

t + 14:02

First RMSD attempt using atom-name matching between crystal and docked pose.

首次 RMSD 计算——按原子名匹配。

Crystal atom names: CAA CAB CAC CAD ... NBD CBE (descriptive) Docked atom names: C C S C C C C C N ... (generic from PDBQT→PDB) Common heavy atoms for RMSD: 0 Centroid distance (fallback): 0.131 Å Greedy-matched RMSD: 0.532 Å (likely underestimated)

Correct diagnosis Agent 识别出原子名不匹配是因为 PDBQT → PDB 转换不保留晶体结构的自定义原子名。它没有接受 0.131 Å 或 0.532 Å 这两个可能被"看起来很棒"的数字，而是启动了基于 RDKit MCS 的正确算法。

Cell 22

complete

t + 14:50

RMSD via RDKit maximum common substructure atom mapping.

通过 RDKit MCS 最大公共子结构做原子映射，正确计算 RMSD。

MCS: 30 atoms · 32 bonds

*** MCS-matched RMSD (30 heavy atoms): 1.218 Å ***

✓ RMSD < 2.0 Å — redocking PASSES validation threshold
  Interpretation: Vina correctly reproduces the JQ1 crystal binding mode.

入场券 · The gate-keeping check 计算药化 docking 任务的必经环节：若 self-docking 无法复现晶体构象，后续所有对接结论都不成立。Agent 在此之前完成这一步——未被提示。

04PART FOUR

20 次对接，一条 r = 0.603 的直线

Cells 24 – 34 · Parallel docking · Pearson · LOO-CV

Cell 24

complete · 20/20

t + 18:40

Dock all 20 benchmark compounds; record best Vina score for each.

对 20 个基准化合物依次对接，记录每个化合物的最佳 Vina 分数。

ChEMBL ID	pKi (exp)	Vina (kcal/mol)
CHEMBL3581647	7.79	−10.210
CHEMBL4297454	7.56	−9.987
CHEMBL3893098	8.29	−9.458
CHEMBL1232461	7.03	−9.397
CHEMBL3900452	8.91	−9.234
CHEMBL4171228	8.09	−9.166
CHEMBL3987016	8.52	−9.035
CHEMBL3895660	8.79	−8.913
CHEMBL1995703	6.89	−8.715
CHEMBL4648362	5.93	−8.587
CHEMBL6025219	6.78	−8.433
CHEMBL1957266 · JQ1	7.21	−8.124
CHEMBL5772221	5.26	−8.081
CHEMBL2017291	6.94	−8.092
CHEMBL5845072	5.26	−7.950
CHEMBL2179387	6.66	−7.847
CHEMBL4078100	7.80	−7.679
CHEMBL6027008	5.26	−7.697
CHEMBL2393130	5.94	−7.111
CHEMBL4846805	6.55	−7.064

=== Correlation ===
Pearson r ( |Vina| vs pKi_exp ):  0.603   (p = 0.0049, n = 20)
Literature benchmark for BRD4 / Vina: r ≈ 0.3 – 0.7   ✓ within range

=== LOO-CV ===
Leave-one-out RMSE:  0.98 pKi units
LOO Pearson r:       0.503
Linear model:  pKi = −0.787 × Vina_score + 0.355

Fig. 1

rendered

t + 22:00

Vina score (negated) vs experimental pKi.

散点图：|Vina| 分数对实验 pKi。

05PART FIVE

五个新分子，一把尺子

Cells 28 – 38 · Scaffold extension · SA score · ADMET

Cell 30

complete

t + 28:12

Design five JQ1 analogs (tBu-ester replacement), compute SA score, ADMET, Lipinski, PAINS, and dock each.

设计五个 JQ1 类似物（均为 tBu 酯的替换），计算合成可及性、ADMET 性质、Lipinski 和 PAINS 过滤，并对每个进行对接。

ID · 设计	MW	logP	TPSA	SA	Vina	Pred pKi	Lip.	PAINS
AN-01 · furan	440.9	5.06	82.5	3.48	−8.020	6.67	✗	—
AN-02 · pyrrole	439.9	4.80	85.2	3.54	−8.387	6.96	✓	—
AN-03 · ethanolamine	444.0	3.30	92.4	3.29	−8.883	7.35	✓	—
AN-04 · morpholine	470.0	4.05	72.6	3.34	−9.818	8.08	✓	—
AN-05 · CF3	490.6	5.90	69.4	3.43	−9.888	8.14	✗	—

=== Top-2 for synthesis (agent-ranked) === #1 AN-04 · JQ1-morpholine SMILES: Cc1sc2c(c1C)C(c1ccc(Cl)cc1)=N[C@@H](CC(=O)N1CCOCC1)c1nnc(C)n1-2 Pred pKi: 8.08 ± ~1.0 (LOO-RMSE) Rationale: tBu ester → morpholine amide; ZA channel H-bond + solubility #2 AN-03 · JQ1-ethanolamine SMILES: Cc1sc2c(c1C)C(c1ccc(Cl)cc1)=N[C@@H](CC(=O)NCCO)c1nnc(C)n1-2 Pred pKi: 7.35 ± ~1.0 (LOO-RMSE) Rationale: tBu ester → ethanolamine amide; ZA channel extension · removes labile ester

Caveat · 警告 LOO-RMSE 已达 0.98 pKi 单位，意味着 AN-04 与 AN-03 之间 0.73 pKi 的预测差距在噪声之内——agent 在下一幕主动指出了这一点。

06PART SIX

Agent 起草的 confidence attribution

Cells 36 & 40 · Unprompted self-critique · 置信度自评

Cell 36

audit

t + 35:10

Unprompted audit of completed steps: SMILES validity, scaffold diversity, conservation claims, methodological gaps.

未经提示的完整审计：类似物 SMILES 合法性、基准集骨架多样性、水分子保守性论据、方法论缺口。

[1] ANALOG SMILES VALIDITY CHECK OK AN-01 (JQ1-furan) atoms=31 chiral=[(16,'S')] OK AN-02 (JQ1-pyrrole) atoms=31 chiral=[(16,'S')] OK AN-03 (JQ1-ethanolamine) atoms=30 chiral=[(16,'S')] OK AN-04 (JQ1-morpholine) atoms=32 chiral=[(16,'S')] OK AN-05 (JQ1-CF3) atoms=34 chiral=[(19,'S')] [2] BENCHMARK SCAFFOLD DIVERSITY CHECK Mean pairwise Tanimoto (ECFP4): 0.166 (<0.3 = diverse) Max pairwise Tanimoto: 1.000 Compounds with Tanimoto ≥ 0.4 to JQ1: 4 / 20 → CONCERN: correlation may be inflated by scaffold bias. [3] HOH 223 CONSERVATION — WHAT WE ACTUALLY SHOWED We observed one conserved bridging water in this single crystal. We did NOT perform MD or cross-structure analysis to show it is conserved across BRD4 structures.

Cell 40

verdict

t + 37:58

Final confidence attribution: each major claim rated by the agent with its own reasoning.

最终置信度自评——agent 对每一个主要结论自评置信度与理由。

Medium

Retrospective SAR narrative (Part 3)

Based on Vina scores + known BRD4 pharmacophore. No MM-GBSA, no PLIP fingerprints computed.

Medium-High

SAR cliff explanation (CHEMBL1957266 vs CHEMBL3900452)

Tanimoto = 0.627 confirms structural similarity. ΔpKi = 1.70 is real (ChEMBL). ZA-channel piperazine H-bond driver is well supported by literature.

Low-Medium

Top-2 prospective ranking (AN-04 > AN-03)

LOO-RMSE = 0.98 pKi; the 0.73 pKi gap between AN-04 and AN-03 is within noise. Both are reasonable — ranking is not statistically distinguishable.

Low

BD1 / BD2 selectivity prediction (AN-05)

BD2 docking was NOT performed. Selectivity claim is based on structural reasoning only (BD1 Asp144 vs BD2 His437). No computational evidence generated.

编辑按语 · Editorial 这是这份档案真正的"不可替代之处"。一个 autonomous agent 在全流程结束后，主动起草了这张 confidence 表——并把自己最吸引人的结论（选择性预测）评为 LOW。这种"把未完成的工作标记为未完成"的能力，在当前大多数 AI agent 的输出里几乎见不到。

What the agent did well

入场券验证。主动做 JQ1 self-docking，得 RMSD = 1.218 Å，跨过了 < 2.0 Å 这道行业门槛，才允许后续结论成立。
四次错误自恢复。meeko/gemmi、Vina v1.2.5 CLI、PDB 原子名匹配、f-string 三元表达式——全部由下一个 cell 诊断并绕过。
HOH 223 的判断力。识别保守水但不把它塞进对接盒——计算药化的 "know but don't abuse" 式判断。
把数字放回 literature baseline。r = 0.603 与文献 BRD4/Vina 0.3–0.7 的直接对比，而非裸报数字。
对自己不利的数据说实话。AN-04 vs AN-03 的排名位于 LOO 噪声之内——agent 自己点出。

What it flagged as limits

Scaffold 偏向。基准集中 20 个化合物有 4 个 Tanimoto ≥ 0.4 相对 JQ1，相关性可能被骨架相似性抬高。
HOH 223 仅一晶体结构观察。没有 MD、没有跨结构分析来真正论证其"保守性"。
AN-05 LOO 误差 > 2 pKi。预测信号不可靠，不应作为强结论。
BD2 对接缺失。选择性预测仅凭结构推理，置信度 LOW，未生成计算证据。
无 MM-GBSA / PLIP。SAR 叙事仅基于 Vina 分数与已知药效团，置信度 MEDIUM。

THE Docking
DOSSIER

JQ1 on BRD4,
redocked, benchmarked,
extended — and then
disputed by its own author.

Contents

从 PDB 3MXF 出发

从 22,285 到 20

受体制备，四次翻车，一次入场

20 次对接，一条 r = 0.603 的直线

五个新分子，一把尺子

Agent 起草的 confidence attribution

What the agent did well

What it flagged as limits

JQ1 on BRD4, redocked, benchmarked, extended — and then disputed by its own author.

Contents

从 PDB 3MXF 出发

从 22,285 到 20

受体制备，四次翻车，一次入场

20 次对接，一条 r = 0.603 的直线

五个新分子，一把尺子

Agent 起草的 confidence attribution

What the agent did well

What it flagged as limits

JQ1 on BRD4,
redocked, benchmarked,
extended — and then
disputed by its own author.