LoFlexMDM-Medium-Shared-BracketSAFE

Variable-length discrete diffusion model for bracket SAFE molecule generation.

Paper: https://arxiv.org/pdf/2602.18695

Training

Hyperparameter	Value
Learning rate	0.0001
Global batch size	2048
Block size	256
Training steps	50000
Weight decay	0.01
Dataset	Bracket SAFE (datamol-io/safe-gpt, ~1.17B molecules)
Checkpoint	EMA weights at step 50000

W&B run: learned-noise-icml/safe_lflexmdm_no_lora_fix_b_um_a1

Unconditional generation (de novo)

1024 sampling steps, 1000 molecules per run, mean ± std over 5 seeds (from paper Table 1 / appendix).

conf.	p	Validity (%)	Diversity	Uniqueness (%)	Quality (%)
yes	1.0	98.300 ± 0.100	0.900 ± 0.000	99.400 ± 0.200	51.500 ± 0.600
yes	0.9	99.200 ± 0.100	0.890 ± 0.000	99.600 ± 0.100	69.200 ± 1.100
yes	0.5	99.800 ± 0.000	0.850 ± 0.000	96.900 ± 0.200	70.800 ± 1.000
no	1.0	98.400 ± 0.100	0.900 ± 0.000	99.200 ± 0.100	52.800 ± 1.100
no	0.9	99.200 ± 0.100	0.890 ± 0.000	99.300 ± 0.100	66.800 ± 0.900
no	0.5	99.500 ± 0.000	0.860 ± 0.000	99.500 ± 0.100	71.200 ± 0.600

Conditional generation (fragment-constrained)

Means over 5 runs (from paper Table 2). Tasks: LD (linker design), ME (motif extension), SD (scaffold decoration), SG (superstructure generation).

Task	Validity (%)	Diversity	Uniqueness (%)	Quality (%)
Linker design	99.6	0.576	64.4	51.7
Motif extension	99.9	0.608	79.2	53.6
Scaffold decoration	99.8	0.601	82.6	40.5
Superstructure generation	100.0	0.593	72.6	37.0

Usage

See the https://github.com/dhruvdcoder/LoFlexMDM release repository for training and evaluation instructions.