FlexMDM-Medium-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_flexmdm_v2

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 (%)
no		98.900 ± 0.100	0.890 ± 0.000	99.600 ± 0.100	62.0 ± 0.7
yes		67.800 ± 0.300	0.940 ± 0.000	61.700 ± 0.700	5.500 ± 0.400

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	99.7	0.599	63.7	50.8
Motif extension	99.7	0.623	79.9	46.9
Scaffold decoration	99.6	0.615	84.3	39.0
Superstructure generation	99.7	0.616	74.5	35.8

Usage

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