Research map / Probabilistic and scientific modeling / Molecules and proteins

pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Steers a pre-trained Edit Flow with a Doob-h transform over an augmented Tchebycheff, feasibility-gated preference, approximated via short Monte Carlo rollouts, for multi-objective constrained protein/peptide sequence editing.

Problem

Editing existing biomolecular sequences (GFP, Cas9, peptide binders) to improve multiple objectives under hard feasibility constraints in discrete, variable-length spaces, which existing de novo or shrinkage-only generators do not jointly support.

Approach

Defines a feasibility-gated terminal preference from an augmented Tchebycheff utility, realizes the tilt via a Doob-h transform of a pre-trained Edit Flow CTMC, and approximates the harmonic function h_t with short Monte Carlo rollouts over candidate edits.

Key results

• On GFP shrinkage to 213 residues, pCoMole (UniRef_S Edit Flow) reaches excitation error 5.82 nm vs 22.36 (RayGun) and 20.13 (SCISOR), with brightness 43.5 vs 9.4 and 19.0.
• Two 229-residue pCoMole eGFP variants (10 deletions, 1-2 substitutions) retained clear green fluorescence in BL21 cells, while a 228-residue negative control did not.
• On Cas9 PAM-preserving shrinkage, PAM exact match under CICERO rose from 4% to 88% for St3Cas9 and from 20% to 72% for GeoCas9 with PAM guidance.

Surprising findings

• pCoMole with a UniRef-pretrained Edit Flow (no GFP fine-tuning) outperformed the GFP-fine-tuned version on excitation error (5.82 vs 8.44 nm), suggesting gains come from the sampling scheme rather than the specialized prior.
• SCISOR reached the low GFP quality tier faster than pCoMole (14 s vs 266 s), showing pCoMole is not faster at easy thresholds.

Limitations

• Higher computational cost from rollout-based guidance than unguided samplers, notably at easy quality tiers.
• Property predictors (FPredX, PeptiVerse, Admetica, Cas9 classifier, Protein2PAM) are computational oracles; wet-lab validation covers only eGFP.
• Approximation guarantees rely on rollout accuracy assumptions; the method may return the input sequence when no feasible terminal is found.

Experimental setup

Datasets: UniRef GFP sequences (UniProt P42212 start) Cas9-family sequences peptidomimetic SELFIES (28k) PDB peptide binder targets

Benchmarks: GFP shrinkage (FPredX excitation/brightness) Cas9 shrinkage (PAM match, Cas9-likelihood) peptidomimetic binder compression (seven drug-related properties) wet-lab eGFP fluorescence in BL21

Metrics: excitation error from 488 nm brightness PAM match rate Cas9 likelihood permeability half-life affinity motif score validity rate time-to-usable

Baselines: RayGun SCISOR CMOMO MolDQN GCPN

Builds on: Edit Flows discrete flow matching Doob-h transform SCISOR RayGun

Closest papers by meaning

• ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.86)
• A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.86)
• Variable-Length Generative Protein Design via Generalized Poisson Flow (NeurIPS 2026 · similarity 0.86)
• PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.86)
• Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.85)
• On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization (NeurIPS 2026 · similarity 0.85)