AI4S for Programmers
August 13, 2026 · View on GitHub
两天，零生物背景，学会读懂一份抗体／蛋白数据集，并在五分钟内判断它可不可信。 AI for Science 里最难的部分不是科学，是数据工程——而那是程序员的主场。
🔗 ai4s.runixcloud.io · 中文授课 · English overview
关键词：AI for Science 入门 · AI4S 教程 · 生物信息学 入门 · 抗体数据 · 蛋白数据清洗 · 生物数据工程 · 程序员转 AI 制药 · IMGT Kabat Chothia 区别 · 表位 互补位 · UniProt 归一化 · 交叉源验证 · AlphaFold 局限
⚠️ 范围声明：这门课只覆盖 AI for Science 的生物数据方向（抗体与蛋白质的数据采集、 清洗、实体归一与质量验证）。材料、气候、天文等其它 AI4S 方向不在范围内，模型训练也不讲。 一门讲「怎么识破夸大数据」的课，没有资格夸大自己的覆盖面。
两天学什么
| 阶段 | 内容 | 时长 |
|---|---|---|
| 1 · 建立词汇表 | 蛋白质是字符串、抗体是 4 条链的对象、UniProt 登录号是主键 | 150 min |
| 2 · 看懂真实数据 | FASTA / PDB / 三套编号体系；打开一条真实记录逐字段对照 | 150 min |
| 3 · 建立判断力 ★ | 亲手把四个真实的 bug 跑出来：词边界假阳性、别名劫持、实算表位、断言式质检 | 210 min |
| 4 · 拿去用 | 三维验收、十个必问问题、五分钟抽查法、红旗清单 | 60 min |
学完做结业考试，答对 16/20 可下载结业证明 （浏览器本地生成，不是第三方认证）。
The premise
Biology is not hard because it is complicated. It is hard because every introduction assumes you already have the vocabulary. So this course never explains a biological concept on its own terms first:
| Biology | What you already know |
|---|---|
| Protein sequence | string over a 20-letter alphabet |
| Antibody | an object holding 4 strings |
| Variable / constant region | instance fields / shared class members |
| CDR | the few substrings that decide behaviour |
| KD (affinity) | a float on a log scale — read it like p99 latency |
| UniProt accession | primary key / UUID |
| Protein aliases | display names, and they collide |
| IMGT / Kabat / Chothia | competing schema versions |
| Cross-source validation | you cannot verify a source with itself |
| QC flags | a known-issues list |
Every data problem in this field is one you have already met in software engineering. What is missing is the vocabulary, not the ability.
What is in this repo
content/ course text, one HTML fragment per page + pages.py (metadata & SEO)
assets/ stylesheet, client JS, the 35-question bank, OG image template
labs/ six runnable Python labs + the public sample dataset
build.py stdlib-only static site generator
dist/ the built site (committed; CI fails if it is stale)
The six labs
Every lab is a real script producing real output — no pseudocode, no toy data.
| Lab | What it does | Network |
|---|---|---|
lab1_read_one_record.py | Print one antibody record field by field | offline |
lab2_data_checkup.py | COUNT(*) / GROUP BY health check over the dataset | offline |
lab3_substring_bug.py | Reproduce three classes of name-matching false positive | offline |
lab4_cross_source.py | Ask HGNC and UniProt the same question, watch them disagree | required |
lab5_compute_epitope.py | Compute an epitope from 3D coordinates via neighbour search | offline |
lab6_assert_qc.py | Data QC written as assertions, with a deliberate trap | required |
git clone https://github.com/Runix-lab/ai4s-for-programmers
cd ai4s-for-programmers
python3 -m venv .venv && source .venv/bin/activate
pip install -r labs/requirements.txt
python labs/lab1_read_one_record.py
About the sample dataset
labs/data/sample_antibodies.csv holds 29 antibody-antigen complexes and is built
entirely from public RCSB PDB data by labs/data/build_sample.py — structures and
sequences from RCSB, CDRs numbered locally with ANARCI/abnumber, epitopes computed with
gemmi. The build script is in the repo, so the dataset is reproducible rather than
asserted.
It deliberately keeps its defects: 3 single-domain antibodies with no light chain,
2 rows missing a UniProt cross-reference, 1 with an ambiguity code — all flagged in
QC_flags. A teaching dataset scrubbed clean installs intuitions that do not survive
contact with real data.
The sample also independently confirms two claims the course makes:
29/29 rows have identical Kabat and Chothia CDR-H3, and 16/29 rows have an antigen
chain that is not A. Run lab2 and count for yourself.
Accuracy
Every checkable claim was re-verified against primary sources before publication — database REST APIs, the original numbering-scheme papers, and the structures themselves. That pass found and corrected fourteen errors in the first draft, including one in the table this site uses as its flagship example. Two worth naming, because they get copied around unexamined:
- Kabat and Chothia define CDR-H3 identically. The widely repeated claim that all three schemes carve different H3 boundaries is wrong; the differences are in H1 and L1.
- VHH nanobodies come from camelids only. Shark single-domain antibodies are VNAR, from IgNAR — a separate evolutionary origin, not a kind of VHH.
A course about verifying other people's data has no business hiding its own corrections. Found an error? Open an issue — it gets fixed in public.
Building the site
No npm, no framework, no build dependencies beyond the Python standard library.
python3 build.py # build into dist/
python3 build.py --check # what CI runs: fails if dist/ is stale or lint trips
build.py also lints for the SEO mistakes that are invisible in a browser: over-long
titles (measured in display width, since a CJK glyph is twice as wide), duplicate or
badly sized meta descriptions, and dead internal links. OG images are rendered to PNG
by headless Chrome at build time; the step is skipped where Chrome is absent.
dist/ is committed so the site deploys with no build step. --check in CI is what
keeps that honest — edit a source file without rebuilding and the build fails.
Licence
- Content (course text, glossary, quiz questions): CC BY-SA 4.0
- Code (
build.py,labs/,assets/*.js,assets/*.css): MIT
The sample dataset is derived from RCSB PDB, which places its data in the public domain (CC0). Check the terms of any other database before reusing its data — that is question 10 of the course, and it applies to us too.
社交预览图
docs/social-preview.png（1280×640）可在 GitHub 仓库
Settings → General → Social preview → Upload an image 上传，
决定这个仓库被分享到任何地方时的卡片长什么样。
Built by Runix.