Cycling Optimizer

Iterative optimizer that cycles between a conditioning function and generator

This optimizer is open source. Any third-party models, product names, or trademarks referenced are the property of their respective owners, and Proto is not affiliated with them.

---

Source

proto-bio/proto-language/proto_language/optimizer/cycling_optimizer.py

View source

Cycling optimizer for iterative sequence refinement.

A generalized optimizer that cycles between a user-defined conditioning function and a generator:

1. Call conditioning function with current sequences (from result_sequences)
2. Pass conditioning output to generator’s sample() method (into proposal_sequences)
3. Accept passing proposals into result_sequences (failed stay unchanged)
4. Repeat for num_steps

This enables flexible optimization patterns such as:

• Protein Hunter: Structure prediction -> inverse folding cycles
• Evo2 with feedback: Constraint-guided prompt modification -> generation cycles

• Constraints are optional; if provided, must be filter constraints (have threshold set) - only passing proposals update result_sequences

How It Works

Cycling alternates a conditioning function with a generator, looping between two views of the design (for protein-hunter, sequence and structure) and keeping only proposals that pass every filter.

Cycling alternates a conditioning function with a generator. Each cycle conditions on the current sequences, generates proposals from that conditioning, and keeps only proposals that pass every filter:

for cycle in 1..num_steps:
    data      = conditioning_fn(result_sequences)
    proposals = generator.sample(data)
    result_i ← proposal_i   ⇔   proposal_i passes all filters   (else keep prior)

Cycling takes filter-only constraints (each needs a threshold; a scoring constraint raises an error). The built-in protein-hunter pipeline sets conditioning_fn to structure prediction (Boltz-2 / Chai-1 / AlphaFold3) feeding inverse folding: predict a structure, then redesign the sequence for it.

API Reference

ConfigCyclingOptimizerConfig Source

Configuration for CyclingOptimizer.

This optimizer cycles between a conditioning function and a generator. On each cycle, the conditioning function receives the current proposal sequences, produces conditioning data, which is then passed to the generator’s sample() method.

The conditioning function can be provided either:

1. Directly via the conditioning_fn parameter (programmatic use)
2. Via the pipeline field using a predefined pipeline (API/JSON use)

• Pipeline-specific constraints:
  • protein-hunter requires an inverse_folding generator
• Constraints are optional but if provided must be filter constraints (must have threshold set)

num_steps · integer · required

Number of conditioning-then-generation cycles to run.

num_results · integer

Candidate design trajectories for this optimizer. Overrides program-level count.

pipeline · string

Predefined conditioning pipeline. ‘protein-hunter’ uses structure prediction -> inverse folding.

protein_hunter · ProteinHunterPipelineConfig

Configuration for protein-hunter pipeline. Only used when pipeline=‘protein-hunter’.

seed · integer

Random seed for reproducible optimization, generator, and constraint tool streams.

tracking_interval · integer · default: "1"

Save history and log progress every N steps. Step 0 and final step always saved.

track_proposals · boolean · default: "False"

Save granular per-proposal results (accept/reject) in history snapshots.

verbose · boolean · default: "False"

Emit per-step debug information about proposals, scores, and acceptance through the logger.

Usage

python

>>> def my_conditioning_fn(sequences):
...     return [process(seq) for seq in sequences]
>>> optimizer = CyclingOptimizer(
...     target_segment=segment,
...     constructs=[construct],
...     generators=[generator],
...     constraints=[],
...     config=CyclingOptimizerConfig(num_steps=5, num_results=4),
...     conditioning_fn=my_conditioning_fn,
... )
>>> optimizer.run()

Metadata

Property	Value
Key	cycling
Class	CyclingOptimizer
Targets Single Segment	True
Uses GPU	False