Acadwrite

Introduction writing example: GNN transfer learning for molecular property prediction

Run on 2026-10-26 · Acadwrite introduction writing · shown as generated

This is a full run of the second example on the introduction writing page. The input was a topic and a short description of findings. The AI searched and read the literature, planned the argument, and saved a cited introduction as a note in the workspace. Below are the input, the process, the argument, the full text and the references.

Note: The input is the built-in example on the page. The introduction and references are the raw workflow output with no manual edits (output language: Chinese). Every reference was checked: DOIs resolve and titles match (Gilmer et al. 2017 is an ICML paper without a DOI and links to Semantic Scholar).

What the author entered

Topic

图神经网络迁移学习用于分子性质预测 (GNN transfer learning for molecular property prediction)

Findings and conclusions

To address the scarcity of high-quality molecular data, we designed a GNN transfer learning framework pre-trained on large low-fidelity data. On 28 million protein–ligand interactions it improved sparse-task performance 8-fold while needing an order of magnitude less high-fidelity training data.

Search and output settings
Years: anySources: English journal papersOutput: Chinese

What the workflow did

The whole run happened in the workspace. Besides academic search, the AI used web search to check specific figures such as the size of PDBbind; all search records stay in the workspace.

16
search rounds
14
papers read in full
12
references
6
paragraphs
  1. Read the task brief: topic, findings and output language
  2. Ran 16 searches on GNN molecular representations, self-supervised pre-training, few-shot learning and negative transfer, multi-fidelity learning and binding affinity
  3. Checked specific figures with web search and read 14 papers in full
  4. Wrote the introduction as bottleneck → method basis → pre-training progress → two gaps → this study, with numbered references
  5. Re-read the draft to check numbers against citations

Argument plan

1Bottleneck

High-fidelity labels are expensive and sparse; small data is intrinsic to drug discovery; PDBbind has fewer than 20k complexes with measured affinity.

2Method basis

Molecules are graphs; message passing networks unified property prediction and GNNs became a core tool.

3Pre-training progress

MolFormer, MolE and MG-BERT show that large-scale pre-training improves transferability.

4Gap 1: unstable gains

Ablations show self-supervised graph pre-training often gives no significant gain; heterogeneous tasks invite negative transfer.

5Gap 2: multi-fidelity cascades

Real pipelines produce data of different scale and fidelity that standard transfer methods cannot use as a whole; existing fixes target single endpoints.

6This study

Decouple fidelity from scale: pre-train on abundant low-fidelity data and calibrate with scarce high-fidelity data.

The generated introduction (as is)

Underlined text marks inline citations that link to the source.

分子性质预测是药物发现与材料设计链条上的关键环节——无论是先导化合物的早期筛选、ADMET 性质的成药性评估，还是虚拟筛选中的打分排序，都依赖于对分子结构与性质之间定量关系的可靠刻画。然而，可靠的性质标签通常来自湿实验测定或高精度量子化学计算，成本高、周期长，并受数据保密性的限制。由此形成的"小数据"并非偶发异常，而是药物研发流程的内在特征（Pallikkavaliyaveetil et al., 2025）；在真实场景中，分子往往只带有极少量人工标注，标签稀缺已成为制约监督式人工智能模型效能的普遍问题（Wang et al., 2025）。以蛋白–配体结合亲和力为例，被广泛使用的 PDBbind 数据集仅收录不足两万个带有实验亲和力标注的复合物结构，而大量已有的亲和力测量因缺少配套结构信息而无法被直接利用（Wei et al., 2025）。如何在有限的高精度数据条件下仍得到可用的性质预测模型，因而成为分子机器学习长期未解的难题。

图神经网络（GNN）为这一任务提供了自然的建模范式。分子可视为以原子为节点、化学键为边的图，GNN 通过消息传递机制逐层聚合局部化学环境，直接从分子图上学习与任务相关的表示。Gilmer 等人将此类模型统一为消息传递神经网络（MPNN）框架，并在分子性质预测基准上取得了当时最优的结果，同时指出后续工作应转向规模更大或标签更精确的数据（Gilmer et al., 2017）。此后，GNN 迅速成长为人工智能辅助药物发现的基础工具，被广泛用于分子性质预测、虚拟筛选、分子生成与合成路线设计等任务，并衍生出几何等变建模、可解释性、不确定性量化与规模化架构等方向（Zhang et al., 2025）。这类模型无需人工设计描述符即可端到端地学习结构–性质关系，恰好契合了性质预测对结构信息的需求。

为缓解标注稀缺，"大规模预训练 + 下游微调"逐渐成为主流路线。一方面，自监督预训练在巨量无标注分子上学习通用表示：MolFormer 在约 11 亿条 SMILES 序列上预训练，其学到的表示在多个下游基准上优于监督式和自监督的图神经网络与语言模型（Ross et al., 2021）；MolE 在约 8.42 亿个分子图上进行两阶段预训练，微调后在 22 项 ADMET 任务中的 10 项超过了公开最优结果（Méndez-Lucio et al., 2024）。另一方面，预训练也被直接嫁接到图级模型上处理具体端点：MG-BERT 将局部消息传递纳入 BERT 式结构，以掩码原子预测在大规模无标注数据上预训练，仅需少量微调便在 11 个 ADMET 数据集上稳定优于当时的最优方法（Zhang et al., 2021）。这些结果共同表明，只要预训练数据足够丰富，分子表示的可迁移性就能显著提升。

不过，预训练带来的收益并不稳固。系统的消融研究显示，在小分子数据上，自监督图预训练在不少设置下的增益并不具有统计显著性，有时甚至不如不做预训练；而当下游任务规模较小时，超参数的影响甚至超过预训练任务本身的选择（Sun et al., 2022）。任务之间的异质性同样会侵蚀迁移效果：当不同性质的标签服从不同分布、甚至生化关联本就微弱时，模型必须跨异构任务传递知识，容易受到负迁移的干扰（Wang et al., 2025；Eraqi et al., 2025）。这提示我们，"预训练数据规模大"本身并不足以保证迁移有效，预训练目标与下游任务之间的匹配程度才是决定性因素。

更根本的限制来自数据的存在形式。真实发现流程往往是由筛选漏斗串联的多阶段过程，各阶段产生规模与保真度各异的数据：早期环节可以低成本获得海量但精度有限的测量，后期环节的实验数据精确却稀少。已有的迁移方法大多隐含假定预训练与微调数据同分布，难以整体利用这种多保真度级联；针对图神经网络的实证分析也表明，常规迁移技术总体上无法有效汲取这类级联中的信息（Buterez et al., 2024）。研究者已提出若干局部修补方案：例如在 171 万条计算得到的低精度 log D 数据上预训练、再在约 1.9 万条实验数据上微调，可显著提升 GNN 对 log D7.4 的预测精度（Duan et al., 2023）；利用 45 万余个由结构预测模型生成的合成蛋白–配体复合物预训练打分函数，再以 PDBbind 的实验结构微调，同样优于现有亲和力预测方法（Wei et al., 2025）。然而，这些尝试多针对单一端点或单一一对保真度，尚缺少能够同时处理大规模、跨保真度与跨任务的通用迁移框架。

针对上述问题，本文设计了一种基于图神经网络的迁移学习框架，其基本思路是把"精度"与"规模"解耦：先以海量低精度数据学习可迁移的结构与相互作用模式，再用少量高精度数据完成面向目标性质的适配。我们在涵盖 2800 万条蛋白–配体相互作用数据的多保真度场景中对该框架进行了系统考察。结果显示，它能够把低精度数据中的信息有效注入稀疏任务，使稀疏任务的预测性能提升 8 倍，同时所需的高精度训练数据减少一个数量级。这表明，将大规模低成本数据视为"预训练资源"、将稀缺的高精度实验数据视为"校准资源"，是缓解分子性质预测数据瓶颈的一条可行路径，也为多保真度分子数据的系统性利用提供了可复用的框架。

References

  1. Nazreen Pallikkavaliyaveetil, Sriram Chandrasekaran. Small Data, Big Challenges: Machine- and Deep-Learning Strategies for Data-limited Drug Discovery. Advanced Drug Delivery Reviews, 2025. https://doi.org/10.1016/j.addr.2025.115762
  2. Ze-Yu Wang, Tian-Yi Jiang, Huan Ma, et al. Few-shot Molecular Property Prediction: A Survey. arXiv.org, 2025. https://doi.org/10.48550/arXiv.2510.08900
  3. Jin-Hang Wei, Yu-Pu Zhang, Peter A. Ramdhan, et al. GatorAffinity: Boosting Protein-Ligand Binding Affinity Prediction with Large-Scale Synthetic Structural Data. bioRxiv, 2025. https://doi.org/10.1101/2025.09.29.679384
  4. J. Gilmer, S. Schoenholz, Patrick F. Riley, et al. Neural Message Passing for Quantum Chemistry. International Conference on Machine Learning, 2017. https://www.semanticscholar.org/paper/e24cdf73b3e7e590c2fe5ecac9ae8aa983801367
  5. Odin Zhang, Hai-Tao Lin, Xujun Zhang, et al. Graph Neural Networks in Modern AI-aided Drug Discovery. Chemical Reviews, 2025. https://doi.org/10.48550/arXiv.2506.06915
  6. Jerret Ross, Brian M. Belgodere, Vijil Chenthamarakshan, et al. Large-scale chemical language representations capture molecular structure and properties. Nature Machine Intelligence, 2021. https://doi.org/10.1038/s42256-022-00580-7
  7. Oscar Méndez-Lucio, Christos A. Nicolaou, Berton A. Earnshaw. MolE: a foundation model for molecular graphs using disentangled attention. Nature Communications, 2024. https://doi.org/10.1038/s41467-024-53751-y
  8. Xiao-Chen Zhang, Chengkun Wu, Zhi-Jiang Yang, et al. MG-BERT: leveraging unsupervised atomic representation learning for molecular property prediction. Briefings in bioinformatics, 2021. https://doi.org/10.1093/bib/bbab152
  9. Ruoxi Sun, Hanjun Dai, Adams Wei Yu. Does GNN Pretraining Help Molecular Representation?. Neural Information Processing Systems, 2022. https://doi.org/10.48550/arXiv.2207.06010
  10. Basem A. Eraqi, D. Khizbullin, S. Nagaraja, et al. Molecular property prediction in the ultra‐low data regime. Communications Chemistry, 2025. https://doi.org/10.1038/s42004-025-01592-1
  11. David Buterez, J. Janet, Steven J Kiddle, et al. Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting. Nature Communications, 2024. https://doi.org/10.1038/s41467-024-45566-8
  12. Yan-Qing Duan, Li Fu, Xiao-Chen Zhang, et al. Improved GNNs for Log D7.4 Prediction by Transferring Knowledge from Low-Fidelity Data. Journal of Chemical Information and Modeling, 2023. https://doi.org/10.1021/acs.jcim.2c01564

Keep reading

Another example: thyroid cancer single-cell atlas

10 searches, 4 papers read, an introduction with 17 references.

About introduction writing

How the argument is planned, papers are found and citations are checked.

Try it with your own research

Enter your topic and findings. Acadwrite searches real papers, plans the argument and writes a cited introduction into a workspace note.

Write an introduction