Prompt Format

Format prompts with molecular properties and similarity constraints for conditional generation

Overview

ChemLactica models use a structured prompt format to condition molecule generation on desired properties and structural similarities. Prompts are constructed using special tags that the model learned during pretraining.

Example from the README: </s>[SAS]2.25[/SAS][SIMILAR]CC(=O)OC1=CC=CC=C1C(=O)O 0.62[/SIMILAR][START_SMILES] generates a molecule with SAS score ~2.25 and Tanimoto similarity ~0.62 to aspirin.

Prompt Structure

A complete prompt follows this pattern:

</s>[PROPERTY_TAG]value[/PROPERTY_TAG]...[SIMILAR]SMILES similarity[/SIMILAR]...[START_SMILES]

Components

1. EOS Token (</s>): Marks the start of a new prompt
2. Property Tags: Specify desired molecular properties
3. Similarity Tags: Reference molecules with similarity scores
4. Start Tag: [START_SMILES] signals the model to begin SMILES generation

Property Tags

Available Properties

ChemLactica models understand these molecular properties from pretraining:

SAS - Synthetic Accessibility Score

Estimates how difficult a molecule is to synthesize.

• Range: 1.0 (easy) to 10.0 (very difficult)
• Typical values: 2.0-4.0 for drug-like molecules

"</s>[SAS]2.25[/SAS][START_SMILES]"

QED - Quantitative Estimate of Drug-likeness

Measures drug-likeness based on molecular properties.

• Range: 0.0 to 1.0
• Typical values: >0.7 for drug-like molecules

"</s>[QED]0.9[/QED][START_SMILES]"

TPSA - Topological Polar Surface Area

Surface area of polar atoms, important for membrane permeability.

• Range: 0 to ~300 Ų
• Typical values: 20-140 Ų for oral bioavailability

"</s>[TPSA]80[/TPSA][START_SMILES]"

CLogP - Calculated LogP

Lipophilicity measure for membrane permeability.

• Range: Typically -2 to 6
• Typical values: 0-5 for drug-like molecules

"</s>[CLogP]2.5[/CLogP][START_SMILES]"

Molecular Weight

Total molecular weight in Daltons.

• Range: Variable
• Typical values: 150-500 Da for drug-like molecules

"</s>[MW]350[/MW][START_SMILES]"

Multiple Properties

Combine multiple property tags in a single prompt:

prompt = "</s>[SAS]2.5[/SAS][QED]0.85[/QED][TPSA]75[/TPSA][START_SMILES]"

Similarity Constraints

SIMILAR Tag Format

Reference molecules to guide structural similarity:

"[SIMILAR]{reference_smiles} {similarity_score}[/SIMILAR]"

• reference_smiles: SMILES string of the reference molecule
• similarity_score: Tanimoto similarity (0.0 to 1.0) using Morgan fingerprints (ECFC4)

Example: Aspirin Analogs

aspirin = "CC(=O)OC1=CC=CC=C1C(=O)O"
prompt = f"</s>[SIMILAR]{aspirin} 0.62[/SIMILAR][START_SMILES]"

This generates molecules with ~0.62 Tanimoto similarity to aspirin.

Multiple Similar Molecules

Include multiple reference molecules:

utils.py:175

def create_prompt_with_similars(mol_entry: MoleculeEntry, sim_range=None):
    """Create SIMILAR tags for reference molecules."""
    prompt = ""
    for sim_mol_entry in mol_entry.similar_mol_entries:
        if sim_range:
            # Use random similarity in range for generation
            prompt += f"[SIMILAR]{sim_mol_entry.smiles} {generate_random_number(sim_range[0], sim_range[1]):.2f}[/SIMILAR]"
        else:
            # Use actual Tanimoto similarity
            prompt += f"[SIMILAR]{sim_mol_entry.smiles} {tanimoto_dist_func(sim_mol_entry.fingerprint, mol_entry.fingerprint):.2f}[/SIMILAR]"
    return prompt

In optimization, similarity ranges (e.g., 0.4-0.9) are used to encourage exploration while maintaining structural relationships.

Prompt Construction in Optimization

The OptimEntry.to_prompt() method builds complete prompts for optimization:

utils.py:197

class OptimEntry:
    def to_prompt(
        self, 
        is_generation: bool,
        include_oracle_score: bool, 
        config,
        max_score
    ):
        """Construct prompt from molecule entries and configuration."""
        prompt = ""
        
        # Add historical molecules from optimization trajectory
        for mol_entry in self.mol_entries:
            prompt += config["eos_token"]  # "</s>"
            
            # Add similar molecules
            prompt += create_prompt_with_similars(mol_entry=mol_entry)
            
            # Add custom properties
            for prop_name, prop_spec in mol_entry.add_props.items():
                prompt += f"{prop_spec['start_tag']}{prop_spec['value']}{prop_spec['end_tag']}"
            
            # Add oracle score (for rejection sampling)
            if "rej-sample-v2" in config["strategy"]:
                if include_oracle_score:
                    prompt += f"[PROPERTY]oracle_score {mol_entry.score:.2f}[/PROPERTY]"
            
            prompt += f"[START_SMILES]{mol_entry.smiles}[END_SMILES]"
        
        # Add final generation prompt
        prompt += config["eos_token"]
        
        if is_generation:
            # Use similarity range for exploration
            prompt_with_similars = create_prompt_with_similars(
                self.last_entry, 
                sim_range=config["sim_range"]  # e.g., [0.4, 0.9]
            )
        else:
            # Use actual similarities for training
            prompt_with_similars = create_prompt_with_similars(self.last_entry)
        
        prompt += prompt_with_similars
        
        # Add properties for target molecule
        for prop_name, prop_spec in self.last_entry.add_props.items():
            prompt += prop_spec["start_tag"] + prop_spec["infer_value"](self.last_entry) + prop_spec["end_tag"]
        
        # Add desired oracle score for generation
        if "rej-sample-v2" in config["strategy"]:
            if is_generation:
                desired_oracle_score = generate_random_number(
                    max_score, 
                    config["max_possible_oracle_score"]
                )
                oracle_score = desired_oracle_score
            else:
                oracle_score = self.last_entry.score
            
            if include_oracle_score:
                prompt += f"[PROPERTY]oracle_score {oracle_score:.2f}[/PROPERTY]"
        
        if is_generation:
            prompt += "[START_SMILES]"
        else:
            prompt += f"[START_SMILES]{self.last_entry.smiles}[END_SMILES]"
            prompt += config["eos_token"]
        
        return prompt

Custom Properties

Define custom properties for specific optimization tasks:

example_run.py:98

# Define custom property specification
additional_properties = {
    "tpsa": {
        "start_tag": "[TPSA]",
        "end_tag": "[/TPSA]",
        "calculate_value": lambda mol_entry: f"{rdMolDescriptors.CalcTPSA(mol_entry.mol):.2f}",
        "infer_value": lambda mol_entry: f"{rdMolDescriptors.CalcTPSA(mol_entry.mol):.2f}",
        "value": None  # Calculated later
    }
}

# Pass to optimization
optimize(
    model, tokenizer,
    oracle, config,
    additional_properties=additional_properties
)

Oracle Score Property

When using rejection sampling (rej-sample-v2 strategy), oracle scores are included:

utils.py:214

if "rej-sample-v2" in config["strategy"]:
    if include_oracle_score:
        prompt += f"[PROPERTY]oracle_score {mol_entry.score:.2f}[/PROPERTY]"

This helps the model learn to generate molecules with higher oracle scores.

Example Prompts

Basic Property-Conditioned Generation

# Generate drug-like molecule
"</s>[QED]0.9[/QED][START_SMILES]"

# Generate easily synthesizable molecule
"</s>[SAS]2.0[/SAS][START_SMILES]"

# Generate with specific TPSA
"</s>[TPSA]80[/TPSA][START_SMILES]"

Similarity-Guided Generation

# Generate aspirin analog
aspirin = "CC(=O)OC1=CC=CC=C1C(=O)O"
f"</s>[SIMILAR]{aspirin} 0.7[/SIMILAR][START_SMILES]"

# Multiple similar molecules
ref1 = "CC(=O)OC1=CC=CC=C1C(=O)O"
ref2 = "CC1=CC=CC=C1"
f"</s>[SIMILAR]{ref1} 0.6[/SIMILAR][SIMILAR]{ref2} 0.5[/SIMILAR][START_SMILES]"

Combined Constraints

# Drug-like aspirin analog with good synthetic accessibility
aspirin = "CC(=O)OC1=CC=CC=C1C(=O)O"
f"</s>[QED]0.85[/QED][SAS]2.5[/SAS][SIMILAR]{aspirin} 0.62[/SIMILAR][START_SMILES]"

Optimization with Rejection Sampling

# Training prompt (is_generation=False, include_oracle_score=True)
training_prompt = (
    "</s>"
    "[SIMILAR]CC(=O)OC1=CC=CC=C1C(=O)O 0.65[/SIMILAR]"
    "[QED]0.88[/QED]"
    "[PROPERTY]oracle_score 145.50[/PROPERTY]"
    "[START_SMILES]CC(=O)OC1=CC=C(C=C1)C(=O)O[END_SMILES]"
    "</s>"
)

# Generation prompt (is_generation=True)
generation_prompt = (
    "</s>"
    "[SIMILAR]CC(=O)OC1=CC=C(C=C1)C(=O)O 0.75[/SIMILAR]"
    "[QED]0.90[/QED]"
    "[PROPERTY]oracle_score 150.00[/PROPERTY]"
    "[START_SMILES]"
)

Best Practices

Property Values

Use realistic property value ranges based on your target chemical space

Similarity Scores

Balance exploration (lower similarity) and exploitation (higher similarity)

Multiple References

Include 3-5 similar molecules to guide structural features

Prompt Length

Keep total prompt + generation under 2048 tokens

Always include the EOS token </s> at the start of prompts and the [START_SMILES] tag before generation.

Next Steps

Molecule Generation

Learn how to use prompts for generation

Sampling Strategies

Optimize generation with sampling parameters