4 - Generative modeling

Question 1

What is QSAR modeling?

Answer 1

Quantitative Structure-Activity Relationship modeling

Question 2

What is the purpose of virtual screening in drug discovery?

Answer 2

To identify potential drug candidates from a large library of compounds

Question 3

What does GNN stand for?

Answer 3

Graph Neural Networks

Question 4

What is zero/few-shot learning?

Answer 4

A machine learning approach that requires very few examples to learn

Question 5

Define proteochemometrics.

Answer 5

A field that combines computational chemistry and bioinformatics to analyze biological data

Question 6

What is contrastive learning?

Answer 6

A technique that learns to differentiate between similar and dissimilar data points

Question 7

What is the estimated number of synthesized molecules in the chemical space of drug-like compounds?

Answer 7

Approximately 10^8 synthesized molecules

Question 8

What is the goal of distribution learning in molecule generation?

What model types are used, and how is the process evaluated?

Answer 8

Goals:
- To mimic the property distribution of known molecules
- learn the syntax and produce new, realistic molecules

Models: VAE (Variational Autoencoders), GANs (Generative Adversarial Networks), Diffusion models

Evaluation:
- validity, uniqueness
- similarity to reference molecules
- FCD (Frechet ChemNet Distance)

Question 9

What does goal-directed learning optimize for?
What are typical use cases?

Answer 9

A specific objective, such as bioactivity

• drug candidate optimization
• Hit-to-lead or lead-to-candidate transitions
• Finding high-potency molecules with low toxicity

Question 10

What is conditional generation in the context of molecular design?

Answer 10

Generating molecules based on explicit input conditions

Question 11

What is the significance of SMILES in molecular representation?

Answer 11

A notation that encodes molecular structures in a textual format

Question 12

What does LSTM stand for?

Answer 12

Long Short-Term Memory, a type of recurrent neural network that’s designed to handle sequential data.
They can work better with long-term dependencies (don’t forget important things over time) than regular RNNs. (Using gates)

Question 13

True or False: The LSTM can achieve up to 98% correct compounds in SMILES generation.

Answer 13

True

Question 14

What is a key challenge in generating molecules using machine learning?

Answer 14

Unclear synthesis methods for the generated molecules

Question 15

What is mode collapse in goal-directed generative models?

Answer 15

The tendency to repeatedly generate the same or very similar molecules

Question 16

What is Pareto ranking in multi-objective optimization?

Answer 16

Selecting non-dominated solutions based on multiple conflicting objectives

Question 17

Fill in the blank: The scoring function in continuous optimization can be described as _______.

Answer 17

Hill-climbing

Question 18

What are some strategies to promote diversity in generated molecules?

Answer 18

• Select diverse high-scoring molecules
• Mix exploitation with randomized exploration
• Penalize redundant solutions
• Expand around key substructures
• Sample proportionally to reward
• Reuse diverse SMILES variants during training

Question 19

What does the acronym VAE stand for?

Answer 19

Variational Autoencoder

Question 20

What is the role of scoring functions in drug design?

Answer 20

To evaluate and rank generated molecules based on desired properties

Question 21

What is the estimated number of all drug-like molecules?

Answer 21

Approximately 10^60

Question 22

What is the significance of using a memory buffer in molecule generation?

Answer 22

To penalize rewards for similar molecules and promote diversity

Question 23

What does the term ‘chemical space’ refer to?

Answer 23

The vast array of possible molecular structures and properties

Question 24

What is the purpose of hybrid models in molecule generation?

Answer 24

To combine different modeling approaches for improved generation

Question 25

What is the primary benefit of using diffusion models in molecular generation?

Answer 25

To generate diverse, high-quality molecular structures

Question 26

What is the importance of scoring functions in multi-objective optimization?

Answer 26

They help balance trade-offs between conflicting objectives in drug design

Question 27

What is the primary goal of sampling candidates in generative models?

Answer 27

to generate diverse and valid molecules that explore the learned chemical space, enabling discovery of novel structures with desired properties

Question 28

What is conditional generation in the context of molecular design?

Answer 28

Generate molecules that explicitly satisfy user-defined conditions, such as property values, shapes, or target sequences ## Footnote It involves learning a shared latent space between molecules and conditions.

Question 29

Which model types are commonly used in conditional generation?

Answer 29

* Conditional VAEs * Conditional RNNs and Transformers * Conditional Diffusion Models

Question 30

What is an advantage of conditional generation?

Answer 30

Direct control over properties or constraints via conditioning input ## Footnote This eliminates the need for a scoring function.

Question 31

What is a disadvantage of conditional generation?

Answer 31

Requires paired data (molecules with known properties or targets) ## Footnote This limits generalization to unseen or rare conditions.

Question 32

What is one of the metrics for evaluating generative models?

Answer 32

Chemical/Structural Validity and Non-Redundancy ## Footnote This includes metrics like the percentage of syntactically correct molecules.

Question 33

What does the term 'diversity of generated molecules' refer to?

Answer 33

Historically: internal diversity; fraction of unique compounds/scaffolds New: #Circles (clustering)

Question 34

What does property-based similarity involve?

Answer 34

Compare distributions of molecular descriptors, e.g., logP, MW, TPSA ## Footnote Metrics include Kullback-Leibler (KL) divergence and Kolmogorov-Smirnov (KS) distance.

Question 35

What is Fréchet ChemNet Distance used for?

Answer 35

To compare how molecules are perceived by a pretrained neural network ## Footnote It maps molecules to Gaussian distributions.

Question 36

What is the importance of desirability in goal-directed learning?

Answer 36

The objective is to produce molecules that satisfy all specified goals, not just valid molecules ## Footnote Desirable molecules must also be suitable for real drug discovery.

Question 37

Fill in the blank: Generative deep learning methods can readily be used to generate new _______.

Answer 37

[molecules]

Question 38

What are the three strategies of training generative models?

Answer 38

* Distribution learning * Goal-directed learning * Conditional generation

Question 39

What is a key aspect of evaluating the quality of generated molecules?

Answer 39

Desirability, synthesizability, QED, etc. ## Footnote This evaluation is not straightforward.

Question 40

What different approaches exist for molecule generation?

Answer 40

1. Autoregressive models Sequence-based (e.g., SMILES generation) Graph-based (generate atoms and bonds step by step) 2. Variational Autoencoders (VAEs) Encode molecule to a latent space, sample, decode 3. Generative Adversarial Networks (GANs) Compete a generator vs. discriminator to create realistic molecules 4. Diffusion models Gradually transform noise into a molecule (high diversity and quality) 5. Rule-based / Template-based Apply known reaction templates or synthesis rules 6. Matrix-based Generate adjacency matrices and atom types (less common)

Question 41

Model types for distribution learning

Answer 41

- autoregressive models - generative adversarial network - variational autoencoder - flow-based model - diffusion models - hybrid models

Question 42

Validity, novelty, uniqueness definition (in molecule generation)

Answer 42

Validity: % of syntactically correct molecules Novelty: % of molecules that were not in the training set Uniqueness: % of unique molecules

Question 43

What models are used in goal-directed learning (molecule generation)?

Answer 43

1. Reinforcement learning 2. Genetic algorithms 3. Reward-guided transformers (continuous optimization) 4. Iterative distributed learning

Question 44

What is the Fréchet ChemNet Distance?

Answer 44

A metric used to evaluate how similar two sets of molecules are — typically: Generated molecules vs. Training (reference) molecules. Method: compare real molecules and generated molecules by checking: Do their feature Gaussian distributions (in a neural network) look similar? A pretrained NN (ChemNet) is used