Chatterjee Lab's pCoMole Shrinks Proteins Without Losing What Makes Them Work
A new discrete flow matching framework shrinks proteins like GFP and Cas9 while juggling multiple drug-like properties under hard biochemical constraints.
- pCoMole guides pre-trained Edit Flows to shrink biomolecules while optimizing multiple properties under hard constraints (paper).
- Uses an augmented Tchebycheff utility and Doob-h transform, approximated with short Monte Carlo rollouts over candidate edits.
- Wet-lab validated: two 229-residue eGFP variants with 10 deletions retained green fluorescence in BL21 cells.
- Also demonstrated on Cas9 orthologs (SpCas9, St1Cas9, St3Cas9, GeoCas9) preserving PAM specificity after shrinking.
- Compresses peptide binders into peptidomimetics optimizing seven ADMET and binding properties simultaneously.
- Full code, checkpoints, and configs released on Hugging Face, accepted at NeurIPS 2026.
pCoMole shrinks biomolecules under hard design constraints
Protein and peptide optimization often begins with a working molecule that needs to become smaller without losing the properties that made it useful. A NeurIPS 2026 paper from the Chatterjee Lab introduces pCoMole, a guidance method that edits existing biological sequences while balancing multiple objectives and excluding designs that violate specified constraints. The authors evaluated it computationally on GFP, Cas9 orthologs, and peptidomimetics, then confirmed fluorescence for two edited GFP variants in cells.
| Component | What pCoMole provides |
|---|---|
| Starting point | An existing sequence and a pretrained Edit Flow model |
| Allowed edits | Insertions, deletions, and substitutions |
| Objectives | Multiple property scores combined according to user-specified preferences |
| Constraints | Zero target weight for terminal sequences that fail feasibility rules |
| Guidance | Short Monte Carlo rollouts that estimate the value of candidate edits |
| Experimental evidence | Two shortened eGFP variants retained fluorescence in BL21 cells |
A constrained editor for known leads
Protein and molecular generators commonly sample new sequences from scratch. Lead-optimization programs have a narrower task: preserve a known molecule’s useful behavior while changing its size, stability, permeability, toxicity, binding, or manufacturability. Coordinating those goals usually requires a generator, several property predictors, constraint filters, and a separate candidate-ranking stage.
pCoMole incorporates preferences and feasibility rules into the editing process. A developer supplies an initial sequence, objective functions, trade-off weights, and hard conditions such as a maximum length or minimum property threshold. The method then guides local edits toward terminal sequences that score well and remain within the feasible set.
From preferences to edit probabilities
Edit Flows model sequence changes as a stochastic path through variable-length discrete states. Each transition can insert, remove, or replace a token, allowing the model to revise part of a protein or molecular string without reconstructing the entire sequence.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
