PROTO-NOOS

REINVENT4 samples batches of 128 molecules per step for 10 to 100 steps and stops early at a mean reward of 0.60. The declared reward is the geometric mean of QED and an AutoDock Vina docking term run through DockStream. In our campaign the docking term returned no data, so generation was effectively driven by QED alone. That is the first entry in our register of negative results.

A generic ChEMBL prior, a ScaffoldSimilarity diversity filter and a low stopping threshold limit distribution collapse. The retained pools hold 3,794,691 samples from 602 batches.

A deterministic layer applies Lipinski’s rule of five and an eNTRy-based entry heuristic. A learned accumulation model, trained on LC/MS measurements in a TolC-deficient E. coli strain with 37 descriptors and gradient boosting, reaches ROC AUC 0.845 on a random split and 0.834 on a scaffold split.

It annotates rather than gates. The predicted retention probability separates neither enzymatic activity (ROC AUC 0.44) nor whole-cell activity (0.59), so a hard threshold would only shrink the library without improving agreement with experiment.

A two-pool ordinary differential equation model tracks the free cytoplasmic pool and the target-bound pool through three processes: entry through the membrane, efflux, and association with and dissociation from the target (K_D = k_off / k_on).

The first pass uses a fixed kinetic prior and orders candidates for structure prediction. The second pass replaces the prior with the predicted K_D; only that pass gates, with a minimum engagement and a maximum K_D of 100 µM. Both passes share the same permeability and efflux mappings, so their agreement is not independent validation.

Ligands are validated and canonicalised in RDKit so that incompatible compounds drop out before GPU work. Boltz-2 co-folds protein and ligand and returns an affinity as log IC50, which we use as an operational affinity signal, together with ipTM and pLDDT confidence measures.

PoseBusters runs only on the benchmark path. On 848 BindingDB complexes, strict checking passes none, because the distance and overlap tests against the cofactor fail next to NADPH. With those checks disabled, 68 poses pass.

The receptor (PDB 6XG5) is stripped of alternative conformations, water and buffer and described with AMBER ff14SB and TIP3P water. The ligand is parameterised with GAFF2 and AM1-BCC through ACPYPE, starting from the Boltz-2 pose. Production dynamics runs 0.1 ns, one replica per compound, and yields a backbone deviation and a stable/unstable verdict.

Only 3.85% of runs come out stable and a single short trajectory carries no statistical power, so this stage does not gate. Energetics need longer replicas, free-energy methods or QM/MM; those belong to the biocomplex protocol.

Flux balance analysis on the iML1515 model of E. coli in aerobic M9 minimal medium covers two scenarios: full knockout of the target gene, and partial inhibition with enzyme occupancy taken from concentration over K_D. The outputs are a relative growth rate and a four-level phenotype label.

With the intracellular concentration fixed at 1 µM, growth is a deterministic function of affinity: a threshold rule on K_D (0.11 / 1.00 / 9.00 µM) reproduces 100% of the labels. The block therefore duplicates the affinity signal and is kept out of the objective.

Each quantity passes through a desirability function onto a common dimensionless scale, and the blocks are summed with explicit weights. Score v2 has two active blocks, 0.75 affinity and 0.25 QED; molecular dynamics acts as a gate with weight zero, and penetration stays an annotation.

The variant was chosen from 16 configurations on 1,003 records with measurements and checked under 200 weight perturbations of ±20%. That is a retrospective validation on the same cohort, not a result on untouched data.