FirstDay

Environment Setup andDepthBasic Knowledge Learning and Explanation

1.AIDDOverview: FromCADDToAIDD

2.Software Installation and Environment Setup

(1)anaconda

(2)vscode

(3)Configuration of Environmental Variables

(4)SwitchpipAndcondaMirror Source

(5)Creation of Virtual Environment

3.RDKITUse of the Toolkit

(1)Based onRDKitMolecular Reading and Writing

(2)Based onRDKitMolecular Drawing

(3)Based onRDKitMolecular Fingerprint and Molecular Descriptors

(4)Based onRDKitCompound Similarity and Substructures

4.Methods for Acquiring Comprehensive Drug Databases

(1)Based onrequestsBasic Crawling Operations

(2)Small Molecule DatabasePubChemData Acquisition（pubchempy / requests）

(3)Protein DatabasePDB、UniProtData Acquisition

5.Deep Learning-Assisted Drug Design

(1)Basic Concepts of Neural Networks andsklearnIntroduction to the Toolkit

(2)Basic Knowledge of Graph Neural Networks and Message Passing Mechanisms

(3)TransformerBasic Model Knowledge: Tokenization, Positional Encoding, Attention Mechanism, Encoder, Decoder, Pre-training-Fine-tuning Framework,huggingface Ecosystem Introduction

(4)Model Evaluation and Validation: Accuracy, Precision, Recall,F1Score,ROCCurve,AUCCalculation, Mean Absolute Error, Mean Squared Error,R2Scores, Explained Variance Scores, Cross-Validation, etc.

SecondDay

Top Journal Reproduction Series1——Representation Learning and Property Prediction of Molecules and Biochemical Reactions for Drug Discovery

Training Background:In artificial intelligence-assisted drug discovery (AIDD) in,Representation Learning and Property Prediction of Molecules and Biochemical ReactionsIs the cornerstone of the entire research process. The structure of a molecule determines its function, and how to effectively represent complex molecular structures and biochemical reaction processes in a form that computational models can understand is a precondition for achieving efficient prediction and optimization. By constructing reasonable molecular representations (such as graph neural networks,SMILESEncoding, fingerprints, etc.), we can letAIThe model captures key chemical features, which are then used to predict the physicochemical properties, biological activity, and toxicity of molecules, providing a reliable foundation for subsequent virtual screening, molecular generation, and reaction design. Therefore, this topic not only establishesAIDDThe core competency framework for modeling and predictive capabilities in China also lays a solid foundation for intelligent decision-making throughout the drug discovery process.

Training Content1:

Nature Machine Intelligence｜Application of Attention-Based Neural Networks in Chemical Reaction Space Mapping《Mapping the space of chemical reactions using attention-based neural networks》

1.Dataset

1.1.PistachioDataset: Includes260Tens of thousands of chemical reactions, from patent data, covering792A category of reactions. The data were deduplicated and filtered for validity (usingRDKit）。

1.2.USPTO 1k TPLDataset: Based onUSPTOPatent data, including44.5Ten Thousand Reactions, Generated through Atom Mapping and Template Extraction1,000A reaction template category.

1.3.Schneider 50kDataset: Public dataset, containing5Myriad Reactions,50A category used for comparison with traditional fingerprint methods.

2.Model.The study compared twoTransformerArchitecture:

2.1.BERTClassifier: An encoder-based model, after being pre-trained through masked language modeling, is fine-tuned on classification tasks, using[CLS]Tagged embeddings as reaction fingerprints (rxnfp）。

2.2.Seq2SeqModel: Encoder-Decoder structure that breaks down the classification task into hierarchical predictions of superclasses, categories, and specific reactions. Both adopt a simplified version.BERT(Hidden Layer256Dimension), Input is UnlabeledSMILESSequence, no reagents required-Reagent differentiation or atomic mapping.

3.Training. Model training is divided into two steps:

3.1.Pre-training:BERTThrough MaskSMILESToken prediction tasks are used for self-supervised learning to learn general representations.

3.2.Fine-tuning: Optimizing the model for classification tasks using cross-entropy loss, learning rate2×10⁻⁵, Sequence Length512. The evaluation adopts confusion entropy (CEN) and Matthews correlation coefficient (MCC) to address data imbalance.

Training Content2:

TOPJournal | Prediction of Biochemical Reaction Yield Based on Deep LearningPrediction of chemical reaction yields using deep learning》 

1.Data. The study used three types of data:  

1.1.Buchwald-Hartwig HTEDataset: Contains3955IndividualPdCatalysisC-NConjugation reaction, covering15Halide minerals,4Ligand,3Alkali and23A combination of additives, yield measured through standardized experiments, high data quality.

1.2.Suzuki-Miyaura HTEDataset: Contains5760A reaction, involving15Electrophilic/Nucleophile,12Ligand,8Alkaline and4A combination of solvents with uniformly distributed yields.

1.3.USPTOPatent Dataset: Extracted from publicly available patents, containing reaction yields of different scales (gram-scale and sub-gram-scale). The data is noisy and inconsistently distributed, requiring smoothing via neighboring reaction yields to improve model performance.

2.Model. The core model is based on pre-trainedrxnfp(Reaction Fingerprint)BERTArchitecture, with the addition of a regression layerYield-BERT. The input is a standardized reactionSMILES, capturing contextual information of the reaction center and key reagents through the self-attention mechanism. The model does not require handcrafted features (such asDFTCalculate descriptors), directly predict yield in an end-to-end manner. Experiments show that its performance is superior to traditional methods (such as random forests and molecular fingerprint concatenation), especially inHTEThe data is close to the prediction level of chemical descriptors, and the parameters are highly robust (little impact from hyperparameter tuning).

3.Training. Training is divided into two steps:  

3.1.Pre-training:BERTLearning through masked language tasksSMILESGeneral representation.

3.2.Fine-tuning: Adopt simpleTransformersLibrary andPyTorchFramework, toMSELoss Optimization Regression Layer, Learning Rate (2×10⁻⁵) anddropoutRate (0.1–0.8) as the main parameter tuning object.HTEData Adopted Randomly/Time Division Validation,USPTOData is smoothed by proximity reaction yield to mitigate noise impact. Small-sample experiments (5%Training data) shows that the model can quickly screen high-yield reactions and guide synthesis optimization.

Training Content3:

TOPJournal｜Based onT5ChemRepresentation Learning and Property Prediction of Biochemical Reactions in Models: 《Unified Deep Learning Model for Multitask Reaction Predictions with Explanation》

1.Data Sources and Processing.Through self-supervised pre-training andPubChemThe molecular dataset is used for training to achieve excellent performance in four different types of chemical reaction prediction tasks. The model handles reaction type classification, forward reaction prediction, single-step retrosynthesis, and reaction yield prediction.

2.Model Architecture and Principles。T5ChemThe model is based on natural language processing.“Text-to-Text Transfer Transformer”(T5)Unified deep learning model for framework development, adapted byT5Framework for handling multiple chemical reaction prediction tasks.T5ChemThe model includes an encoder.-Decoder structure, and introduces task-specific prompts and different output layers according to the task type, such as molecular generation head, classification head, and regression head, to handle sequence-to-sequence tasks, reaction type classification, and product yield prediction.

3.Training Process and Details.

3.1.T5ChemThe model was first introduced inPubChemThe97 millionPerform self-supervised pre-training on the molecule, usingBERTSimilar“masked language modeling”Target.

3.2.During the pre-training phase, in the source sequencetokensRandomly masked, the model's goal is to predict the correct masked elements.tokens。

3.3.After pre-training is completed, the model is fine-tuned in downstream supervised tasks using different task-specific prompts and output layers.

3.4.The model generates molecules during the testing phase.token by tokenin a predictive manner until generation“End of sentence marker”Or reach the maximum predicted length.

TheThreeDay

Top Journal Reproduction Series2——Representation Learning and Property Prediction of Proteins Facilitate Drug Discovery

Training Background：InAIDDIn China, proteins are the main targets of drug action, and the complexity of their structure and function determines the success or failure of drug design.Representation Learning and Property Prediction of ProteinsIsomorphic Labs-Target interactions and the discovery of candidate drugs are crucial steps. Proteins, especially enzymes, serve as the primary targets for drugs, and their functions, structures, and dynamic properties directly influence drug design and efficacy. This topic is explored through two cutting-edge research works:*《Enzyme function prediction using contrastive learning》Demonstrates how to use contrastive learning to extract high-quality functional representations from protein sequences, achieving precise prediction of enzyme functions;《CatPred》*An integrated deep learning framework was proposed for in vitro enzyme kinetics parameters, such asKm、kcatetc.) are crucial for establishing efficacy models and optimizing lead compounds. These methods have significantly improved the accuracy and generalization ability of protein modeling, providingAIProvides strong support for target discovery, mechanism understanding, and candidate drug screening.

Training Content1: 

Nature Communication｜A Comprehensive Framework for Deep Learning of Enzyme Kinetic Parameters In Vitro《CatPred: a comprehensive framework for deep learning in vitro enzyme kinetic parameters》

CatPred A comprehensive deep learning framework for predicting in vitro enzyme kinetic parameters (kcat、Km、Ki), to address the issues of high experimental measurement costs, sparse data, and poor generalization capabilities. This method not only provides accurate predictions but also introduces a quantification of prediction uncertainty, supporting out-of-distribution (out-of-distribution) Robust prediction of enzyme sequences. In addition, the authors constructed a new standardized dataset (CatPred-DB), and systematically compared various enzyme representation methods.

1.Data：CatPred The dataset used comes fromBRENDA AndSABIO-RK Database, the author constructedCatPred-DB, including:23197 Articlekcat，41174 ArticleKmAnd11929 ArticleKi Data, each record contains the amino acid sequence of the enzyme,AlphaFold OrESMFold Predicted structures, substratesSMILES Expression. The data was cleaned and standardized, with missing and duplicate values removed, and a logarithmic transformation applied to the parameters to conform to a normal distribution.

2.Model：CatPred Adopting a modular design, enzymes and substrates are represented through different neural network modules for representation learning, and probabilistic regression output (in the form of mean and variance of a Gaussian distribution) is used, allowing for uncertainty estimation.aleatoric + epistemic）。

3.Training

3.1.All models adopt the negative log-likelihood loss function (NLL) training, to simultaneously predict parameter mean and uncertainty.

3.2.Training Use-Verification-Test Trisection Method (80%-10%-10%), and establish“Out-of-training set”The test subset is used for generalization ability evaluation.

3.3.In order to assess uncertainty,CatPred Use10An ensemble of models trained with different initial parameters to quantifyepistemic uncertainty。

3.4.During model training, different similarities (sequencesidentity<99%、80%、60%、40%) The test set demonstrates its robustness.

Training Content2:

Science｜Based on Contrastive LearningProteinClassificationAttribute Prediction《Enzyme function prediction using contrastive learning》

1.Data Sources and Processing: CLEANThe model's training is based onUniProtHigh-quality data in the database, which contains approximately1.9Hundreds of millions of protein sequences.CLEANThe model takes amino acid sequences as input and outputs a list of enzyme functions sorted by probability (ECNumbering as an example). In order to verifyCLEANaccuracy and robustness, the authors conducted extensivein silicoExperiment, andCLEANApplied to an internally collected database of uncharacterized halogenases (total36Individual)ECNumbering annotations, followed by in vitro experimental validation through case studies.

2.Model Architecture and Principles: CLEANThe model adopts a contrastive learning framework, aiming to learn an embedding space for enzymes, where the Euclidean distance reflects functional similarity. Embedding refers to the numerical representation of protein sequences, which is machine-readable while preserving the important characteristics and information carried by the enzymes. InCLEANIn the task, with the sameECThe numbered amino acid sequences have smaller Euclidean distances, while those with differencesECThe sequence of numbers has a greater distance.

3.Training Process and Details:

3.1.During the training process,CLEANThe model is trained with a contrastive loss function, where positive samples are selected based on their proximity to the anchor (anchor) Embed negative sequences with small Euclidean distances to improve training efficiency.

3.2.Model uses language modelESM1bThe obtained protein representation serves as the input to a feedforward neural network, with the output layer generating refined, function-aware embeddings of the input proteins.

3.3.When predicting, by calculating the query sequence with allECThe pairwise distances between clustered centers are used to predict the input protein.ECNumbering.

3.4.CLEANTwo methods have also been developed to predict confidence from the output rankings.ECNumber: One is a greedy method, and the other is based onPThe method of value.

FourthDay

Top Journal Reproduction Series3——Deep Learning-Based Molecular Generation Aids Drug Discovery

Training Background：Molecular generation is a key technology in fields such as chemistry, biology, and materials science, and is of great significance for new drug development, new material design, and chemical reaction prediction. Traditional molecular generation methods rely on expert knowledge and trial-and-error experiments, which are time-consuming and costly. With the development of artificial intelligence technology, especially the application of natural language processing and diffusion models in molecular generation, we are now able to use computational models to accelerate this process. This course will introduce fromNLPThe design patterns of diffusion models, which are capable of understanding and generating molecular structures, thereby enhancing the efficiency and accuracy of molecular design. Through this course, participants will be able to master the latest technologies and methods in molecular generation, as well as how to apply these techniques to practical problems.

Training Content1：

Nature Communication｜Based onEnd-to-End Graph Generation Framework for Molecular Generation：《Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing》

1.Data Sources and Processing:Graph2EditsThe model used a publicly available benchmark dataset.USPTO-50k, including50016A reaction, these reactions are correctly atom-mapped and classified as10Different types of reactions. The dataset is divided into40k、5k、5kThe reactions were used for the training, validation, and test sets.

2.Model Architecture and Principles:Graph2EditsThe model is an end-to-end graph generation architecture based on graph neural networks (GNN) Predict the editing sequence of the product graph and generate intermediates and final reactants based on the predicted editing sequence order. This model merges the two-stage process of the semi-template method (identifying the reaction center and completing the synthon) into a one-pot learning process, enhancing its applicability in complex reactions and making the prediction results more interpretable. The core of the model consists of a graph encoder and an autoregressive model used to generate the editing sequence and apply these edits to infer intermediates and reactants.

3.Training Process and Details:

3.1.Graph2EditsThe model uses a directed message-passing neural network (D-MPNN) as the graph encoder to obtain atomic representations and global graph features, and predict atoms/Key editing and termination symbols.

3.2.Model Training UseTeacher Mandatory Policy, that is, using real editing sequences as model input. At each editing step, the model calculates the probabilities of all possible edits and selects the highest-scoring one.kAn editor applies these edits to the input graph to obtainkAn intermediate.

3.3.During the generation process, the generation branch will stop if the maximum number of steps is reached or the graph representation indicates termination.

3.4.Finally, according to the possibilities for the precedingkA sequence and graph are ranked, and collected as the final prediction result.

Training Content2

Nature Computational Science｜Molecular Generation Network Based on Equivariant Diffusion Model《Structure-based drug design with equivariant diffusion models》

1.Brief introduction. This paper proposes a structure-based drug design method (SBDD), utilizingSE(3)-Equivariant Diffusion Model (DiffSBDD) Generate novel small molecule ligands that match the binding site conditions of proteins. This method involvesSBDDThe problem is modeled as a 3D conditional generation task, capable of generating all atomic positions at once, overcoming the limitations of traditional autoregressive methods that lose global context due to sequential generation.DiffSBDDNot only supports de novo molecular design, but also enables property optimization, negative design, and molecular fragment modification (inpainting) and other tasks are applied flexibly.

2.Data Summary. The study usedCrossDockedAndBinding MOADTwo datasets for training and evaluation.

2.1.CrossDockedThe dataset contains40,344Training Protein-Ligand Pair and130A test pair, validation set size is246Ensure that proteins from different sets come from different main enzyme classification categories to avoid overfitting.

2.2.Binding MOADThe dataset was filtered and used for testing, with analysis limited to samples that all methods could generate.78IndividualCrossDockedAnd119IndividualBinding MOADObjective. In addition, dataset processing involves removing corrupted entries and throughZenodoPublicly provide processed data and sampled molecules to ensure research reproducibility.

 3.Model Summary.DiffSBDDIsSE(3)-Equivariant diffusion models generate 3D molecular structures conditioned on protein binding sites, using3DThe graph representation (atomic coordinates and types) avoids the complex post-processing required to infer molecular structures from density maps in traditional methods. The model design respects rotational and translational symmetries in three-dimensional space.

FifthDay

Top Journal Reproduction Series4: Protein Combined with Molecular Dynamics-Dynamic Prediction of Ligand-Complex Interactions

Training Background:Protein-The prediction of ligand interactions is one of the core tasks in modern drug discovery and bioengineering, and its importance goes without saying. In the drug development process, accurately predicting the binding sites, three-dimensional structures, and affinities of proteins with small-molecule ligands not only helps to uncover the mechanisms of intermolecular interactions but also significantly accelerates the screening and optimization of candidate drugs, reducing research and development costs and time. Traditional experimental methods such asXAlthough X-ray crystallography and nuclear magnetic resonance are precise, they are time-consuming, costly, and struggle to meet the demands of large-scale screening. However, with the rapid development of deep learning and artificial intelligence technologies, computational methods in protein-Show great potential in ligand prediction.

Research Content1: 

Nature Communication｜Interaction-Aware Proteins-Ligand Docking and Affinity Prediction Model《Interformer: an interaction-aware model for protein-ligand docking and affinity prediction》

1.Brief Introduction: This study proposes a method namedInterformerBased onGraph-TransformerUnified Model Architecture for Protein-Ligand docking and affinity prediction. Addressing the limitation of existing deep learning models that neglect modeling non-covalent interactions between protein and ligand atoms,InterformerIntroduced the Interaction-Aware Mixture Density Network (MDN) to explicitly capture hydrogen bonds and hydrophobic interactions, combined with a negative sampling strategy and pseudoHuberLoss Function, Optimizing Interaction Distribution through Contrastive Learning to Enhance the Accuracy of Docking Poses and the Robustness of Affinity Prediction.

2.Dataset: The study usedPDBBindTime-split Test Set (333samples) to evaluate docking accuracy,PosebustersBenchmarking validates physical reasonableness, and internal real-world datasets test generalization ability. Training data is sourced fromPDBBindCrystal Structure Database。

3.Model：InterformerBased onGraph-TransformerArchitecture, including:(1) The graph represents a module, with atoms as nodes and adjacency relationships as edges;(2) Masked Self-Attention (MSA) mechanism, throughIntra-BlocksAndInter-BlocksCapture ligands separately/Interactions within and between proteins;(3) Interactive PerceptionMDN, integrating four Gaussian distributions to simulate van der Waals forces, hydrophobic interactions, and hydrogen bonds;(4) The edge output layer integrates node and edge features to predict energy;(5) The Pose Scoring and Affinity modules predict the correct pose and experimental affinity values based on virtual nodes.

4.Training Details: The training is divided into two stages: first, the energy model is trained based on crystal structures to generate negative samples, and then the pose scoring and affinity models are trained jointly with both positive and negative samples. Optimization is performed using negative log-likelihood loss.MDNBinary cross-entropy loss optimization posture scoring, pseudoHuberLoss (σ=4) Optimize Affinity Prediction (UnitIC50、Kd、KI, after negative log normalization). Monte Carlo sampling generates candidate poses,

Research Content2:

Nature Communication｜Molecular Dynamics-Driven Protein-Dynamic Prediction of Ligand Complex Structures《DynamicBind: predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model》

1.Brief Introduction: This study proposes a method namedDynamicBindA deep learning method for predicting ligand-specific proteins-Ligand complex structure. Traditional molecular docking methods often treat proteins as rigid or only partially flexible, making it difficult to handle large-scale conformational changes in proteins. While molecular dynamics simulations can capture dynamic conformations, the computational cost is high.DynamicBindBy constructing smooth energy landscapes through equivariant geometric diffusion networks, protein simulations without ligands are efficiently modeled (apo) State to Ligand Binding (holo) Conformational transition of the state, without relying onholoStructure or extensive sampling.

2.Dataset: Research based onPDBbind2020Database (19,443A protein-Ligand complex crystal structure), divided by time:2019Data from previous years were used for training and validation,2019The data from the year was used for testing. Additionally,Major Drug Targets (MDT)Test set (599Yes), focusing on kinase,GPCRSuch as the main drug targets, requirementsAlphaFoldPredicted structure and crystal structurepocket RMSD>2Å, ensure the test difficulty. Achieve this during training.AlphaFoldSample of protein sections generated by interpolating predicted structures with crystal structures.

3.Model：DynamicBindIt is an equivariant generative model based on graph neural networks, using coarse-grained representations (proteins asCαNodes and side chain dihedral angles representation, ligands represented by heavy atom nodes), outputs include translation, rotation, torsion angle updates of proteins and ligands, as well as binding affinity andcLDDTConfidence Score. The model learns fromapoToholoThe“morph-like”Transform, optimize energy landscapes, including63.67Million parameters.

4.Training Details: Training in8BlockNvidia A100 80GB GPUCarry out on5Day, input is additionmorphTransformed ProteindecoyConformations and ligand conformations with added Gaussian noise, the goal is denoising operation. The loss function includes eight terms (ligand and protein translation, rotation, torsion, etc.), throughKabschAlgorithm AlignmentapoAndholoStructure, combined with diffusion noise to adjust conformational transitions. Iterative reasoning.20Update the initial structure.