Skip to content

Latest commit

 

History

History
103 lines (69 loc) · 3.94 KB

02-models.md

File metadata and controls

103 lines (69 loc) · 3.94 KB

Models and inference

Molecular property prediction estimates endpoints such as aqueous solubility from a structure. Use predictions to compare analogs against the same property criteria. Language models have a different role in HALO: they propose structures, interpret supplied evidence, and suggest subsequent investigations.

Types of results

Result Origin Example
Descriptor Calculation from molecular structure RDKit molecular weight in g/mol
Prediction Estimate from a molecular model ADMET-AI solubility in log₁₀ mol/L
Measurement Reported experimental result An assay value with its conditions and source

Predictions are comparable when they share an endpoint, unit, and method. A design run saves the predictions used for both its source and its candidates.

Predict molecular properties

1. List the supported properties

The catalog lists the ADMET properties accepted as design criteria, with definitions and value types.

from halo import chem

print(chem.admet_properties())

2. Load the predictor

ADMETAIPredictor combines ADMET-AI predictions with configured PyTDC scoring functions.

from halo.chem.predict.admet_ai import ADMETAIPredictor

predictor = ADMETAIPredictor.load()

3. Predict one molecule

predictions = predictor.predict("CCO")
print(predictions["solubility"])

Recorded example output (ADMET-AI 1.4.0).

1.263 log10 mol/L (admet_ai:Solubility_AqSolDB@1.4.0)

The printed result includes its value, unit, and method. Failed predictions include a reason; successful predictions for other properties remain available.

4. Predict a series

smiles = ["CCO", "CCN", "CCO"]
batch_predictions = predictor.predict_batch(smiles)

Results follow the input order and include repeats. Use predict for a single SMILES string and predict_batch for a sequence.

To compute descriptors and predictions together, pass the loaded predictor to chem.profile_molecule or chem.profile_molecules through predictors=(predictor,).

Prediction source Output
ADMET-AI classifiers Probabilities for endpoints such as hERG inhibition and human intestinal absorption.
ADMET-AI solubility model Aqueous logS in log₁₀ mol/L.
PyTDC scoring functions DRD2 activity score and penalized logP.

Property definitions · Predictor reference

Configure language models

Each agent uses instructions and an output schema for its task. HALO validates the returned structures, references, and selections before using them. Choose a model for each role according to the task it performs.

Configuration Default Purpose
agents.model gpt-6-luna Routing, synthesis, and strategy.
agents.max_tokens 1800 Output token limit for these roles.
agents.roles.optimizer.model gpt-6.1-terra Molecular proposals in the repository configuration.
agents.roles.optimizer.max_tokens 2048 Output token limit for molecular proposals in that configuration.

Set role-specific options

Add a role under agents.roles in the YAML configuration. Available roles are orchestrator, optimizer, researcher, analyst, and strategist.

agents:
  roles:
    researcher:
      model: gpt-6-luna
      max_tokens: 2400
      reasoning_effort: low

Save this configuration as settings.yaml and start the service:

uv run halo serve --config settings.yaml

Set reasoning_effort to none when specifying a temperature. Configure credentials through OPENAI_API_KEY; use OPENAI_BASE_URL for a server supporting the Responses API and structured output. Service setup

For provider interfaces, execution limits, usage, and prompt capture, see the inference reference.