pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows

The theoretical derivation is rigorous: Proposition 3.1 shows that exact Doob-$h$ rates yield the desired tilted terminal distribution $q_G(\cdot \mid x_0)$, the augmented Tchebycheff utility $U(x)$ is strictly monotone for Pareto dominance, and the log-domain rollout implementation (Appendix B) is numerically stable. Empirically, the underlying Edit Flow models exhibit low train-test gaps across domains (Table 1), and on fixed-length GFP shrinkage pCoMole dramatically surpasses RayGun and SCISOR in predicted brightness and excitation alignment, with even a UniRef-pre-trained flow remaining competitive. Cas9 editing is equally strong: in a held-out fixed-length benchmark, pCoMole achieves a 100% exact PAM match against 35% for SCISOR and 0% for RayGun.

The wet-lab validation is thin: only two shortened eGFP designs were tested via qualitative fluorescence imaging, with no quantitative fluorescence assays, replicates, or statistical analysis, and the sole negative control was manually selected rather than generated by a competing method. While the authors claim the rollout approximation preserves theoretical guarantees, Proposition C.3 only bounds one-step multiplicative kernel error; there is no finite-sample bound on how rollout variance accumulates over the full variable-length trajectory under pruning and incumbent tracking. Computationally, the main text never reports the key sampling budgets—candidate count $C$, rollout count $R$, or rollout horizon—making it impossible to assess whether the flat high-tier runtimes in Table 3 trade wall-clock efficiency for massive per-step compute. Finally, peptidomimetic results rest entirely on unvalidated in-silico predictors (PeptiVerse, Admetica), and modest drops in non-toxicity or solubility (Table 6) may reflect real predictor noise rather than true biological trade-offs.

In-silico evidence is robust: the objective-guidance ablation (Table 4) cleanly shows that removing brightness or excitation guidance degrades the corresponding metric, and the independent CICERO evaluation confirms that PAM preservation is driven by guidance rather than by chance. Comparisons to RayGun and SCISOR are fair within the shrinkage setting, but these baselines are inherently not multi-objective optimizers, so reporting superior brightness or Cas9-likeness partially reflects an asymmetric capability gap. The peptidomimetic benchmark benefits from post-hoc pocket retention checks with Boltz-2 and AutoDock VINA, though the affinity objective used during generation was the in-loop PeptiVerse predictor rather than the held-out docking score.

Code and model weights are publicly released on HuggingFace, which is a major strength. However, critical hyperparameters for the guided sampling algorithm—including the number of candidate edits $C$, the number of Monte Carlo rollouts $R$, the inverse-temperature $\beta$, and the rollout horizon—are absent from the main text and only discussed in Appendix D ablations. Edit Flow architecture details (layers, hidden dimensions, exact training compute) and the schedule $\kappa_t$ and weight $w_t$ in Equation 6 are likewise relegated to appendices. Without these values, an independent team cannot reproduce the computational budget or the exact guidance schedule, severely impeding reproducibility despite the open code release.