Skip to content

Molecular reaction-bank measurement v1

  • Stability: development
  • Revision: v1-development
  • Scope: fixed replay banks constructed from a frozen RxnFlow starting policy

This convention describes one target-specific molecular experience bank before post-training. The native study fixes the RxnFlow policy, synthesis environment, MolGenBench target, docking and eligibility contract, bank size, and two-view confirmation rule. A profile exposes construction-view summaries for one bank. The independent view qualifies the stability of the six-bank ordering and does not contribute members or get averaged into a coordinate.

All coordinates use the 512 unique molecules in the frozen construction-view bank. The direction of reward and coverage coordinates is contextual because their relationship to downstream policy utility has not yet been evaluated. Lower generated-action and oracle costs retain their direct operational interpretation.

mean_target_reward

Arithmetic mean of the frozen Q07820 terminal reward. The reward is max(-vina_affinity, 0) * QED after the study's molecular eligibility and failure contract. Unit: docking_qed_reward. A molecule-specific eligibility or conformer failure is preserved as an explicit scientific zero in the native probe and remains outside the eligible bank pool. A ligand atom type rejected by the pinned Vina parser is also explicit scientific ineligibility. Every other parser, receptor, toolchain, docking-process, and timeout failure remains an infrastructure failure and cannot enter a bank or this mean as zero. Role: target_affinity. Sample unit: eligible bank molecule. Direction: contextual until downstream utility is scored.

mean_generated_action_count

Arithmetic mean number of molecular construction actions in a sampled RxnFlow trajectory. Unit: generated_molecular_actions_per_trajectory. Role: cost. Sample unit: bank trajectory. Direction: lower.

mean_oracle_seconds

Arithmetic mean selected-target oracle wall time among bank members. Unit: seconds_per_molecule. This is realized wall time under the pinned oracle runtime rather than a hardware-independent compute measure. Role: cost. Sample unit: eligible bank molecule. Direction: lower.

unique_scaffold_fraction

Number of distinct achiral Bemis--Murcko scaffolds divided by bank size. Acyclic molecules use the study's explicit sentinel. Unit: fraction. Role: coverage. Sample unit: bank. Direction: contextual.

unique_ring_systems_per_record

Number of distinct connected ring-system signatures divided by bank size. Rings belong to one system when they share at least one atom. Unit: ring_systems_per_record. Role: coverage. Sample unit: bank. Direction: contextual.

unique_route_fraction

Number of distinct ordered reaction-name sequences divided by bank size. Unit: fraction. Role: coverage. Sample unit: bank. Direction: contextual.

unique_reaction_families

Count of distinct reaction families represented by the bank. Unit: reaction_families. Role: coverage. Sample unit: bank. Direction: contextual.

embedding_mean_nearest_neighbor_distance

Arithmetic mean nearest-neighbor Euclidean distance within the bank after the frozen RxnFlow GINE state embeddings are centered on the eligible construction pool, projected by deterministic full SVD onto at most 32 nonzero components, scaled by the retained singular values, and L2-normalized. Unit: normalized_embedding_distance. Role: coverage. Sample unit: eligible bank molecule. Direction: contextual.

Readiness and edge cases

A bank is ready for profile comparison only when source, target, starting-policy, measurement-reliability, and bank-confirmation checks pass under matching contracts and seeds. The adapter rejects missing memberships, nonfinite values, digest mismatches, and any record that opened target-training outcomes. No uncertainty interval is exported because the native study estimates ordering stability across views rather than coordinate-level sampling intervals. Profile conformance does not establish that any coordinate predicts post-training utility.