Protein Function Prediction

Predict 500 anonymous Gene Ontology (GO) function labels from protein sequences using homology retrieval, protein language models, annotation-pretrained models, and a multilayer ensemble.

Final release: V5.0.0 · Highest reported leaderboard score: 64.3818

The final submission is pairgogpt_v2_allkbag2.csv, covering 28,450 proteins and 500 binary labels. The score uses the report's Macro-F1 × 100 convention, equivalent to 0.643818 on a 0–1 scale. The project owner confirmed this file and its score for this release on 2026-10-17.

Download V5.0.0 · Analysis report · Source archive

Report and repository guide

Explore the original report directly on GitHub: chapter guide · all 10 figures · all 9 tables · downloadable table data.

README.md                    Project overview and current V5 result
docs/
  README.md                  Report navigation and version context
  report/                    Abstract, chapters 1–8, references, appendices
  figures.md                 Complete illustrated gallery
  tables.md                  All report tables rendered for GitHub
assets/figures/report/        10 original PNG figures and provenance manifest
data/report-tables/          9 UTF-8 CSV tables and provenance manifest

The chapter pages retain the report's Chinese text, preceded by short English summaries. Original figures and tables document the historical allbag2 analysis; the current V5 artifact is the separately supplied allkbag2 CSV described above.

Task and data

Item	Description
Training set	113,796 labeled proteins
Test set	28,450 proteins
Targets	500 anonymous GO-related binary labels
Input	Protein sequences using the competition's nonstandard letter encoding
Metric	Macro-F1: the unweighted mean of F1 across the 500 labels
Submission	protein_id, followed by label_0 through label_499; predictions are 0 or 1

The report describes a sparse, imbalanced task: the training label matrix has about 5.08% positive entries and an average of 25.42 labels per protein. Sequence length, homology coverage, label correlations, and annotation-depth variation guide the feature design.

Method

1. Align sequence encoding. Infer the letter-to-amino-acid mapping from within-training-set alignment statistics and BLOSUM62. Feed the mapped sequences to pretrained models.
2. Build complementary predictors. Combine MMseqs2 homology transfer within the competition training set, k-mer ridge regression, ESM2-650M embeddings with kNN and neural heads, and annotation features from GO-GPT v1 and four ProteInfer checkpoints.
3. Stack the base models. Combine 11 base predictors through a LightGBM protein–label pair model and an MLP stacker, then blend their predictions at 2:1.
4. Fuse annotation signals per label. Logistic regression combines the base logit, selected GO-GPT and ProteInfer terms, homology interactions, and generated-term counts. Term selection uses correlation within the training folds.
5. Learn cross-label corrections. A LightGBM pair stacker uses the fusion logit as its initial score and learns corrections from model logits, direct annotation features, protein-level prediction shape, related labels, inferred label hierarchy, and homology statistics.
6. Average seeds and apply thresholds. The final allkbag2 variant averages two seeds and adds GO-GPT score-space kNN to the 15 predictor inputs used by the preceding allbag2 variant. Each label has its own decision threshold.

flowchart TD
    A[Competition sequences] --> B[Encoding alignment]
    B --> C[MMseqs2 homology]
    B --> D[ESM2 representations]
    B --> E[GO-GPT v1 and ProteInfer]
    C --> F[Base predictors and stacking]
    D --> F
    F --> G[Per-label logistic fusion]
    E --> G
    G --> H[LightGBM protein-label pair stacker]
    E --> H
    F --> H
    H --> I[Two-seed average and label thresholds]
    I --> J[Final binary prediction CSV]

Loading

The v2 in pairgogpt_v2_* is a local pipeline iteration name. The archived workflow uses the pretrained GO-GPT v1 model.

Results and development history

The local results below come from the analysis report and the source archive's reports/cv_log.csv. Values are Macro-F1 × 100; dev, clean, and leaderboard evaluations are different measurements.

Development stage	Dev	Clean
Base ensemble blend_sm	47.69	47.59
Add GO-GPT per-label fusion	53.93	53.83
Add ProteInfer per-label features	56.02	55.94
Pair stacker pairgogpt_v2_pi	60.39	60.32
Direct ProteInfer features, four checkpoints, ridge and shape features pid4rs	62.29	62.24
Preceding two-seed ensemble allbag2	62.724	62.679
Final supplied two-seed ensemble allkbag2	62.761	62.720

The previous PairGO submission recorded 62.1406 on the leaderboard. The final supplied CSV is recorded at 64.3818, an increase of 2.2412 points. Exact leaderboard scores are supplied by the project owner; this publication did not perform a new leaderboard query.

Artifact history: the original report and source archive describe allbag2 as their final submission and associate it with the earlier recorded 64.38/64.3818 result. Their bundled CSV differs from allkbag2. For V5.0.0, the owner explicitly selected the separately supplied allkbag2 file and confirmed 64.3818 as its score. Use the standalone V5 CSV and its checksum for the current final artifact; the report and source archive are preserved as historical materials.

Results in the original report

Figure 8. Local honest Macro-F1 across development stages. The original chart ends at allbag2 and retains its historical leaderboard annotations; it is not a new evaluation of the V5 allkbag2 file.

See the complete figure gallery for the workflow, sequence and label distributions, correlations, homology structure, and performance by subgroup.

Evaluation protocol

• Five folds keep identical sequences together. This groups exact sequences, rather than homology clusters.
• The report calls threshold evaluation “honest”: thresholds for a fold are tuned using the other folds.
• Dev contains 108,796 training proteins. Clean contains 100,504 after excluding 8,292 proteins used for candidate GO-term selection.
• A separate 5,000-protein confirmation set was used at two earlier model milestones. No confirmation-set score is reported for the final allkbag2 variant.
• The report identifies fully nested cross-fitting across ensemble layers as future work. Local selection results should be read with that limitation.

The report's findings include strong gains from annotation-pretrained models and cross-label context, alongside annotation differences among identical sequences. The reported training workflow uses competition data and downloaded pretrained weights; it does not provide an independent audit of those weights' original training-data overlap.

Files and reproduction

V5 asset	Contents
pairgogpt_v2_allkbag2.csv	Owner-confirmed final submission
protein_function_prediction_report.docx	Original Chinese analysis report, unchanged
protein_function_prediction_source.zip	Original release(1).zip, unchanged; source scripts, model-resource manifest, development logs, and historical allbag2 CSV
V5_manifest.json	File hashes, CSV checks, score provenance, and artifact-version notes
SHA256SUMS.txt	Download integrity checks

The source archive expands to release/. Start with its README.md, MODELS.md, and requirements.txt. The documented environment is Windows 11, Python 3.12, an RTX 4060 Laptop GPU with 8 GB VRAM, and 16 GB RAM. Competition train.csv, test.csv, and submit_template_v1.csv, pretrained weights, and large feature caches must be obtained or generated separately.

Stage	Main scripts in release/src/
Fold assignment and encoding map	02_make_folds.py, 13_encoding_map.py
Base predictors	03_baseline_kmer.py, 04_embed.py, 05_train_mlp.py, 06_knn.py, 07_mmseqs_knn.py, 17_train_mlp_cond.py
Base stacking	21_depth.py, 09_stack.py, 23_mlp_stack.py, 08_blend.py
GO-GPT features	20_gogpt.py, 26_camber_prep.py, 27_gogpt_merge.py
ProteInfer features	33_proteinfer.py, 34_proteinfer_terms.py, 40_pinf_ridge.py
Fusion, pair stacking, seed averaging, export	28_gogpt_fusion.py, 35_pair_gogpt.py, 41_avg.py, 38_make_sub.py

The included command sequence reproduces the earlier allbag2 workflow. Although its logs include allkbag2, the archive does not include the additional gogptq feature-generation step required for complete reproduction of that variant. The published final CSV is therefore the authoritative downloadable result; the source archive is a reference implementation with this documented gap. Training was not rerun during publication. Third-party model sources, revisions, checksums, and licenses are listed in MODELS.md and the bundled upstream license files.

Release history

Release numbers follow publication milestones and are separate from model-family names such as V25, V36, and PairGO.

Release	Artifact or experiment	Recorded result
V5.0.0	Final pairgogpt_v2_allkbag2.csv, report, and source archive	Leaderboard 64.3818
V4.0.0	PairGO pairgogpt_v2_pi.csv	Leaderboard 62.1406
V3.0.0	V36 controlled GO-feature pilot	Pilot CV 38.7896; promotion gate not passed
V2.0.0	V25 + GO + V30 Pro evaluation artifacts	Historical selection 48.6082; no test CSV
V1.0.0	V25 + GO prediction CSV	Historical selection 48.1454

This table displays scores on a 0–100 scale; earlier release titles may use the 0–1 scale. V1–V3 used different local protocols and should not be directly ranked against PairGO leaderboard results. PairGO's pi and allkbag2 files belong to the workflow described in the supplied report; they are distinct from the separate V41 fusion experiment.

---

蛋白质功能预测

本项目根据蛋白质序列预测 500 个匿名 GO 功能标签，融合训练集内部的同源检索、蛋白质语言模型、功能注释预训练模型与多层集成学习。

最终版本：V5.0.0 · 报告最高排行榜得分：64.3818

最终文件为 pairgogpt_v2_allkbag2.csv，包含 28,450 条蛋白质记录和 500 个二元标签。按报告采用的 Macro-F1 × 100 口径，64.3818 对应 0–1 标度的 0.643818。项目所有者于 2026-10-17 明确确认了本次发布的文件及其成绩。

下载 V5.0.0 · 数智分析报告 · 源码资料包

报告与仓库导航

直接在 GitHub 阅读：报告章节目录 · 全部 10 张图 · 全部 9 张表 · 表格 CSV 数据。

docs/report/ 按报告顺序整理摘要、第 1–8 章、参考文献及附录；每页先提供英文概述，再保留中文正文。assets/figures/report/ 存放提取的原始 PNG，data/report-tables/ 存放原始表格的 CSV，并附来源与校验记录。

原图表记录报告中的历史 allbag2 方案；当前 V5 最终结果仍以本次单独提供的 allkbag2 为准。

任务与数据

项目	说明
训练集	113,796 条带标签蛋白质记录
测试集	28,450 条蛋白质记录
预测目标	500 个匿名 GO 相关标签
输入	使用比赛非标准字母编码的蛋白质序列
评价指标	对 500 个标签分别计算 F1 后取算术平均，即 Macro-F1
提交格式	protein_id 加 label_0 至 label_499，标签值为 0 或 1

报告记录的标签矩阵阳性比例约为 5.08%，每条蛋白质平均有 25.42 个标签。标签稀疏且分布不均，因此方案同时关注序列表示、同源覆盖、标签关联与注释深度差异。

技术方案

1. 序列编码对齐。 根据训练集内部比对统计与 BLOSUM62 推断字母到标准氨基酸的映射，再输入预训练模型。
2. 构建多种预测来源。 使用训练集内部 MMseqs2 同源迁移、k-mer 岭回归、ESM2-650M 嵌入及其 kNN/神经网络分类头，并提取 GO-GPT v1 与四个 ProteInfer 模型的功能注释特征。
3. 形成基础集成层。 将 11 个基础预测器输入 LightGBM 配对堆叠与 MLP 堆叠，再按 2:1 融合。
4. 逐标签融合。 逻辑回归结合基础层 logit、折内筛选的 GO-GPT/ProteInfer 术语、同源交互和生成术语数量。
5. 学习跨标签修正。 LightGBM 以“蛋白质 × 标签”为样本，以融合 logit 为初始分数，利用模型输出、注释特征、预测形状、相关标签及标签层级等信息进行修正。
6. 双种子平均与逐标签决策。 allkbag2 在此前 allbag2 的 15 项预测输入上加入 GO-GPT 分数空间 kNN，平均两个种子的结果，再应用各标签的阈值。

文件名 pairgogpt_v2_* 中的 v2 表示本地流程迭代；归档方案实际使用的预训练模型为 GO-GPT v1。

报告中的总体技术路线

图 1 保留报告原始的 15 项模型输入与 62.72 本地成绩，对应历史 allbag2 流程；当前 allkbag2 增加了 GO-GPT 分数空间 kNN 输入。

创作历程与结果

以下本地结果来自分析报告及源码包 reports/cv_log.csv，统一采用 Macro-F1 × 100。dev、clean 和排行榜对应不同的评估数据。

迭代阶段	dev	clean
基础集成 blend_sm	47.69	47.59
加入 GO-GPT 逐标签融合	53.93	53.83
加入 ProteInfer 特征	56.02	55.94
配对堆叠 pairgogpt_v2_pi	60.39	60.32
四个 ProteInfer 模型、直接 logit、岭回归与预测形状 pid4rs	62.29	62.24
较早的双种子集成 allbag2	62.724	62.679
本次提供的最终双种子集成 allkbag2	62.761	62.720

此前 PairGO 版本的排行榜得分为 62.1406；本次最终文件记录为 64.3818，提高 2.2412 分。排行榜精确成绩由项目所有者提供，本次发布没有重新查询比赛平台。

文件沿革说明。 原报告及源码包曾将 allbag2 记作最终提交，并关联当时记录的 64.38/64.3818；包内 CSV 与本次 allkbag2 的预测内容不同。发布 V5.0.0 时，项目所有者明确选择单独提供的 allkbag2，并确认其成绩为 64.3818。当前最终结果以 V5 单独提供的 CSV 及校验值为准，原报告和源码包作为历史资料保留。

评估方式与结果边界

• 五折划分将完全相同的序列放在同一折；它不是按同源簇划分。
• 报告中的 honest 口径指每折阈值仅用其余折调优，再用于当前折计分。
• dev 包含 108,796 条训练记录；clean 再排除用于候选 GO 术语筛选的 8,292 条记录，共 100,504 条。
• 另有 5,000 条确认集记录，在此前两个关键阶段使用。最终 allkbag2 没有报告确认集成绩。
• 报告将各集成层之间的完整嵌套交叉拟合列为后续改进方向，应据此理解本地选模成绩。

报告发现，功能注释预训练模型与跨标签上下文带来了主要增益，而相同序列之间的标签差异体现了注释深度变化。报告描述的训练流程使用比赛数据和公开预训练权重；这些权重原始训练数据的重叠情况未经过独立审计。

下载内容与复现

V5 附件	内容
pairgogpt_v2_allkbag2.csv	项目所有者确认的最终预测文件
protein_function_prediction_report.docx	原始中文数智分析报告
protein_function_prediction_source.zip	原 release(1).zip，包含源码、模型资源清单、开发日志和历史 allbag2 CSV
V5_manifest.json	文件哈希、CSV 检查、成绩来源与版本对应说明
SHA256SUMS.txt	下载文件完整性校验值

源码包解压后目录为 release/。先阅读其中的 README.md、MODELS.md 与 requirements.txt。记录的开发环境为 Windows 11、Python 3.12、RTX 4060 Laptop 8 GB 显卡及 16 GB 内存。比赛原始数据、预训练权重和大型中间特征需要另行获取或生成。

主要流程依次包括折划分与编码映射、基础模型、GO-GPT/ProteInfer 特征、逐标签融合、配对堆叠、种子平均及 CSV 导出。对应脚本包括 02_make_folds.py、13_encoding_map.py、20_gogpt.py、33_proteinfer.py、28_gogpt_fusion.py、35_pair_gogpt.py、41_avg.py 和 38_make_sub.py，完整命令见包内说明。

包内命令针对较早的 allbag2 流程。虽然日志包含 allkbag2 的评估结果，但未收录其新增 gogptq 特征的生成步骤，因此该源码包不能单独保证完整复现本次最终 CSV。最终预测以发布附件为准；本次发布未重新训练模型。第三方资源的来源、版本、校验值和许可列于 MODELS.md 及上游许可文件。

发布版本

版本	内容	记录结果
V5.0.0	最终 allkbag2 CSV、报告及源码包	排行榜 64.3818
V4.0.0	PairGO pi CSV	排行榜 62.1406
V3.0.0	V36 受控 GO 特征试验	pilot CV 38.7896，未达到晋级门槛
V2.0.0	V25 + GO + V30 Pro 评估资料	历史 selection 48.6082，没有测试 CSV
V1.0.0	V25 + GO 预测 CSV	历史 selection 48.1454

表中统一使用 0–100 标度，早期 Release 标题可能保留 0–1 标度。V1–V3 的本地评估口径与 PairGO 排行榜不同，不应直接排名比较。Release 编号与 V25、V36 等模型名称独立；报告中的 PairGO pi、allkbag2 文件也与另外开展的 V41 融合实验区分。