Research map / Graph learning / Graph self-supervised learning
SALSA: Semantically-Aware Latent Space Autoencoder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SALSA, a SMILES transformer autoencoder with a supervised contrastive loss over graph-edit (1-GED) mutant sets, yielding a latent space that better tracks molecular structural similarity, interpolation quality, and physicochemical properties.
Problem
SMILES-based autoencoders map structurally similar molecules to distant latent codes, producing disorganized latent spaces that hurt downstream drug discovery tasks.
Approach
Combine a transformer autoencoder's reconstruction loss with a SupCon contrastive loss over sets of 1-graph-edit mutants generated by add/replace/remove operators, so structurally similar molecules cluster in latent space while reconstruction prevents degenerate collapse.
Key results
- SALSA-32 reaches Spearman rho 0.878 for GED-EuD correlation versus 0.514 for the Naive autoencoder (Contra-32 0.888).
- SALSA produces interpolants with lower endpoint Tanimoto distance than the Naive decoder across five compound classes.
- SALSA achieves the highest property-distance correlation among models for nine of ten physicochemical properties.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the contrastive objective to SMILES reconstruction (ablation) | SALSA-32 versus Naive-32 · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.514 ± 0.53 -> 0.878 ± 0.20 | Helped |
| Add the contrastive objective to SMILES reconstruction (ablation) | SALSA-32 versus Naive-32 · ChEMBL23 anchor-supermutant evaluation | Kendall's τ: 0.467 ± 0.48 -> 0.824 ± 0.23 | Helped |
| Add reconstruction loss to the contrastive-only encoder (ablation) | SALSA-32 · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.888 ± 0.20 -> 0.878 ± 0.20 | Hurt |
| Reduce latent dimension from 32 to 16 (ablation) | SALSA · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.878 ± 0.20 -> 0.849 ± 0.24 | No clear effect |
| Reduce latent dimension from 32 to 8 (ablation) | SALSA · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.878 ± 0.20 -> 0.807 ± 0.28 | No clear effect |
| Reduce latent dimension from 32 to 4 (ablation) | SALSA · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.878 ± 0.20 -> 0.587 ± 0.46 | Hurt |
| Reduce latent dimension from 32 to 2 (ablation) | SALSA · ChEMBL23 anchor-supermutant evaluation | Spearman's ρ: 0.878 ± 0.20 -> 0.351 ± 0.57 | Hurt |
| Use SALSA rather than the Naive decoder for spherical interpolation | Molecular interpolation · Five compound classes | Endpoint-interpolant Tanimoto distance | Helped |
| Use SALSA rather than Naive or Contra representations | Physicochemical property representations · 10 physicochemical properties | Spearman's ρ between PropΔ and EuD | Mixed |
Surprising findings
- The contrastive-only encoder (Contra) performs slightly better than SALSA on GED-EuD correlation, and the paper attributes this to the regularizing effect of reconstruction.
- The paper notes it is not obvious why SALSA beats both Naive and Contra on property awareness.
Limitations
- Semantics are defined only as 1-GED structural similarity via a restricted set of node-level mutations (no ring breaking or disconnection).
- Naive autoencoder shows bimodal GED-EuD distributions and SALSA correlation flattens for larger n-GED.
- Contrastive-only training can collapse similar molecules, so the method relies on the reconstruction term.
Experimental setup
Datasets: ChEMBL21 ChEMBL23
Benchmarks: GED-EuD correlation on ChEMBL23 supermutants interpolation Tanimoto distance on five compound classes property-distance vs EuD correlation over 10 RDKit properties
Metrics: Spearman's rho Kendall's tau Tanimoto distance Spearman's rho for Prop-Delta vs EuD
Baselines: Naive autoencoder Contra encoder
Builds on: transformer supervised contrastive loss denoising adversarial autoencoder ChemVAE
Closest papers by meaning
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.87)
- SMI-Editor: Edit-based SMILES Language Model with Fragment-level Supervision (ICLR 2025 · similarity 0.86)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.86)
- MOL-Mamba: Enhancing Molecular Representation with Structural & Electronic Insights (AAAI 2025 · similarity 0.86)
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.86)
- Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.