Rewards

Reward models for search-based protein design.

RewardModel combines the outputs of explicitly selected metrics into a scalar reward for one search checkpoint.

For specialized metrics (RF3, AF2, TMOL), see rfproteina.metrics.

class rfproteina.rewards.RewardModel(rewards: Mapping[str, Any])

Bases: object

Aggregate metric outputs for one search checkpoint.

Parameters:

rewards – Map metric.output keys to weight, clip, and normalize_by settings. Values are clipped, divided by normalize_by, and multiplied by weight. Defaults are weight=1 and normalize_by=1, with no clipping.

The search strategy owns a separate instance for each checkpoint and runs the metrics explicitly listed there before aggregating their outputs.

aggregate_metrics_output(metrics_output: dict[str, Any]) → float

Aggregate a metrics output dict into a scalar reward using configured weights.

Values are clipped per clip_values before weighting. metrics_output is never mutated, so downstream reporting still sees the raw metric values.

Parameters:

metrics_output – Dictionary of metric results keyed by "metric_name.output_key" (e.g. "rf3.plddt").

Returns:

Weighted scalar reward, or -inf if a weighted key is missing.