STAR-VAE: Latent Variable Transformers for Scalable and...
𝚐𝔪𝟾𝚡𝚡𝟾@gm8xx8
38 views
Nov 05, 2025
~1 min read
Advertisement
1
STAR-VAE: Latent Variable Transformers for Scalable and Controllable Molecular Generation
STAR-VAE (IBM) pairs a Transformer encoder, an autoregressive decoder, and a latent Gaussian bottleneck trained on 79M molecules encoded with SELFIES (a string format that guarantees 100 % chemically valid molecules). LoRA conditioning and classifier guidance make it property-aware without full retraining.
- Achieves perfect scores for validity, uniqueness, and novelty (1.000 each on MOSES)
- GuacaMol KL ≈ 0.916 (best among VAEs)
- Improves docking performance: mean binding scores change from −5.0 to −6.3 on 1SYH and from −5.2 to −5.4 on 6Y2F, indicating stronger predicted target binding
STAR-VAE (IBM) pairs a Transformer encoder, an autoregressive decoder, and a latent Gaussian bottleneck trained on 79M molecules encoded with SELFIES (a string format that guarantees 100 % chemically valid molecules). LoRA conditioning and classifier guidance make it property-aware without full retraining.
- Achieves perfect scores for validity, uniqueness, and novelty (1.000 each on MOSES)
- GuacaMol KL ≈ 0.916 (best among VAEs)
- Improves docking performance: mean binding scores change from −5.0 to −6.3 on 1SYH and from −5.2 to −5.4 on 6Y2F, indicating stronger predicted target binding
2