什么是分子优化？（Molecule Optimization OA）

分子优化是药物发现的关键步骤，，可通过化学修饰分子从而改善候选药物的所需特性，目标是发现与已知起始分子相比具有改善的药物特性的分子（产生新的有效分子，使药物特性最大化，同时保持与输入分子的相似性）。例如，在先导优化中，可以改变先导分子的化学结构以提高它们的选择性和特异性。传统上，这种分子优化过程是根据药物化学家的知识和经验规划的，并通过基于片段的筛选或合成来进行。因此，它不可扩展或自动化。

药物中的小分子以及他的药理学是什么？

小分子（英语：small molecule）是一个有机化学概念。一般将分子量小于900道尔顿的有机化合物分子称为小分子。目前大部分药物都是小分子类药物，蛋白质、核酸等生物大分子的基本组成单位（如氨基酸、核糖核苷酸、脱氧核苷酸）也是小分子。小分子药物一般是抑制剂，通过干扰蛋白间相互作用起效。

药理学
目前大部分的药物都是小分子药物。小分子的分子量上限一般设在900道尔顿，这是因为900道尔顿以下的分子在人体内能较快速地扩散进入细胞，到达作用靶点。根据里宾斯基五规则，口服给药的小分子药物的分子量最好能低于500道尔顿，否则损耗率会出现明显上升。

1、A Deep Generative Model for Molecule Optimization via One Fragment Modification

通过一个片段修饰进行分子优化的深度生成模型

A Deep Generative Model for Molecule Optimization via One Fragment Modification | Papers With Code

这里面的数据集似乎不错？（有paired ...）

数据集的说明：

1. Use provided processed dataset

If you want to use our provided processed dataset, please check the directories below:

./data/logp06/ : the dataset of pairs of molecules with 0.6 similarity and different on penalized logp property.

./data/drd2_25 : the dataset of pairs of molecules with 0.6 similarity and different on DRD2 property. The property difference between each pair of molecules is greater than 0.25.

./data/qed_1 : the dataset of pairs of molecules with 0.6 similarity and different on QED property. The property difference between each pair of molecules is greater than 0.1. The QED properties of these molecules are greater than 0.7.

./data/drd2_25_qed6 : the dataset of pairs of molecules with 0.6 similarity and different on QED property and DRD2 property. The property differences on DRD2 between each pair of molecules are greater than 0.25 (i.e., $DRD2(Y)-DRD2(X)\geq 0.25$ ). The QED property of each pair of molecules should satisfy QED(X)<0.6≤QED(Y).

In each directory, you will see the following files:

1、multiple zipped tensors-*.pkl files. These binary files contain the processed data including pairs of molecules and their edit paths. The data in these *.pkl files should be used for model training. All the tensors-*.pkl files will be read into Modof as training data. If you are using your own training data rather than the provided one, you can generate such tensors-*.pkl using the data processing tools as will be described below.

Note: Due to the limit of file size, we only provide part of the processed file here. To use the whole training dataset, please use the provided data preprocessing script to preprocess the dataset. Please decompress the zipped file before using them to train the model.

2、train_pairs.txt file in logp06 dataset. This file contains all pairs of molecules used in Jin’s paper. This file is identical to train_pairs.txt file in (iclr19-graph2graph/data/logp06 at master · wengong-jin/iclr19-graph2graph · GitHub). Please note that the molecule pairs contained in tensors-*.pkl files are a subset of all the molecule pairs in train_pairs.txt.

File format: each line in train_pairs.txt has two SMILE strings, separated by an empty space. The first SMILE string represents the molecule with worse properties, and the second SMILE string represent the molecule with better properties.

3、one_ds_pairs.txt file. This file contains the pairs of molecules used in Modof.

File format: each line in one_ds_pairs.txt has two SMILE strings, separated by an empty space.

4、test.txt. This file contains the SMILE strings of single molecules that are used as the testing molecules in XXX’s paper. These molecules are also the testing molecules used in our Modof.

File format: each line in test.txt is a SMILE string of a testing molecule.

5、vocab.txt. This file contains all the substructures of training molecules in tensors-*.pkl files. These substructures are in SMILE strings.

File format: each line in vocab.txt is a SMILE string of a substructures. The i-th row represents the i-th substructure (i.e., ‘i’ here is the substructure ID).