REM3DI PCQM epoch 2 continued on PubChemQC COSMO
Initialization: Queimo/rem3di-pcqm500k-corrected at commit 2f5c3b40b36259a13adc8412689254054bb93288, runs/corrected_20261027/evaluation/paper_audit/author_pcqm500k_5ep/epoch_02.
Ten supervised epochs predict 153 NHB/OH/OT profile-bin areas plus cavity A and V. Head: Linear(640,256), SiLU, Linear(256,155). Encoder and learned preprocessors are trainable; MACE-POLAR is frozen. Train-only log transforms and standardization are in each checkpoint's label_transform.json. Loss equally weights mean bin MSE, A MSE and V MSE in standardized units. Invert standardization, then expm1 on bins and exp on A,V. Physical constraints are not enforced by the head.
Each completed epoch contains backbone/head weights, optimizer, scheduler, RNG state, metrics, label transform and SHA-256 manifest. hub_verified.json in the notebook workspace records the verified Hub commit. Epoch directories appear only after their training completes. See config.json and dataset_reference.json for the actual data revision. The source folder contains the complete workflow. The frozen MACE foundation weights and legacy REM3DI backend are inherited from the existing notebook; environment/source provenance is in provenance.json.
After all ten epochs, evaluation runs ridge, Kumo-Tabular large, and supervised REM3DI fine-tuning for each epoch on ESOL and FreeSolv, with three matched random and scaffold outer folds. Inner validation controls fine-tuning and ridge selection. Benchmark connectivity matches are excluded from COSMO training. Earlier PCQM/author pretraining exposure is not removed retroactively. Benchmark subsets match the existing notebook: ESOL 1108, FreeSolv 625.
Training and evaluation are complete: all ten COSMO epochs, all 360 requested cross-validation fits, and 36 matched PCQM epoch 2 control fits are saved.
Comparison report · Full epoch summary · Fold-level results · Training on W&B · Evaluation on W&B.
The control uses exactly the same cached benchmark geometries, saved folds, and fitting seeds. At COSMO epoch 10, mean RMSE improves over that control in all 12 method/dataset/split combinations. Three-fold variation is reported as sample SD; these comparisons do not establish statistical significance.
Expanded literature and classical baselines
The expanded comparison is complete: 36 published REM3DI ridge/Kumo/fine-tuning fits, 12 Monroe + Kumo fits, and 12 ECFP4 + RDKit 2D + Kumo fits, using the same retained ESOL/FreeSolv cohorts and saved random/scaffold three-fold splits. There are 456 fold results including all existing COSMO and PCQM evaluations.
Expanded report · All fold results · All epoch summaries · Expanded W&B comparison.
Monroe uses the released author encoder and native conformer preprocessing; Monroe + Kumo is a new downstream evaluation rather than the paper's TabPFN recipe. All Kumo baselines use four estimators and a 640-column cap. The report records preprocessing, native feature widths, and fold SDs.
OpenADMET / ExpansionRx
COSMO epoch 10 + Kumo Large and Monroe + Kumo Large are evaluated on all nine endpoints using the original 5,326 / 2,282 train/test split: 18 endpoint results. Both frozen encoders use the same cached conformers. Kumo uses four estimators, seed 42 and a 640-column cap. Metrics use the raw numeric CSV labels, matching the linked Monroe notebook's data convention; its downstream model is TabPFN.
OpenADMET report · Endpoint metrics · Updated combined report · W&B.
BigSolDB 2.0 / Solvaformer comparison
Two frozen-encoder concatenation baselines are complete: COSMO epoch 10 and Monroe, each using [solute embedding, solvent embedding, temperature] with Kumo Large and all 1,281 / 1,441 raw input columns.
The exact released 9,686-row test cohort was recovered from author prediction row IDs and checked against all labels, temperatures and molecular identities; the committed predictions reproduce the paper's MPNN+Charges metrics. This resolves a conflicting 9,250 test count in the paper prose. Kumo uses the 75,731-row non-test complement; the private neural-model validation subdivision is not reproduced. See the report for the actual Kumo context plan.
BigSolDB report and published-table comparison · Metrics · Updated combined report · W&B.
- Downloads last month
- 323