Skip to content

DETANGO

Summary

DETANGO resolves a fundamental limitation of standard protein language models (pLMs) such as ESM-1v: when a pLM flags a mutation as evolutionary implausible, it conflates structural destabilization with direct functional disruption.[1] DETANGO reprograms frozen pLM representations by factorizing total evolutionary plausibility into distinct stability-driven and function-driven components.[1] This exposes stable-but-inactive (SBI) variants, maps allosteric networks, and pinpoints functional residues without requiring supervised functional annotations.[1]

Mathematical Formulation and Architecture

1. Evolutionary Likelihood Deconvolution

The total mutation effect $e(\mathbf{x}^{\text{MT}})$ predicted by a pLM as the log pseudo-likelihood ratio between mutant sequence $\mathbf{x}^{\text{MT}}$ and wild-type sequence $\mathbf{x}^{\text{WT}}$ is factorized into additive structural ($s$) and functional ($f$) plausibility scores:[1] $$e(\mathbf{x}^{\text{MT}}) = \log \frac{p_e(\mathbf{x}^{\text{MT}})}{p_e(\mathbf{x}^{\text{WT}})} = s(\mathbf{x}^{\text{MT}}) + f(\mathbf{x}^{\text{MT}})$$ where: - $s(\mathbf{x}^{\text{MT}}) = \log \frac{p_s(\mathbf{x}^{\text{MT}})}{p_s(\mathbf{x}^{\text{WT}})}$ is the structural plausibility, capturing changes in folding free energy ($\Delta\Delta G$) or cellular abundance.[1] - $f(\mathbf{x}^{\text{MT}}) = \log \frac{p_f(\mathbf{x}^{\text{MT}})}{p_f(\mathbf{x}^{\text{WT}})}$ is the functional plausibility, reflecting residual evolutionary constraints specific to catalysis, binding, or allostery.[1]

2. Network Architecture & Loss Objective

DETANGO takes internal representations from frozen Transformer layers of ESM-1v and splits them into a stability projection subspace and a functional residual subspace:[1] 1. Stability Projector: Explicitly trained via mean squared error ($L_{\text{stab}}$) to predict structural stability measurements ($\Delta\Delta G$ from FoldX or cellular abundance from DMS data).[1] 2. Functional Residual Subspace: Calculated by subtracting the stability projection from the total evolutionary representation.[1] 3. Training Objective: Jointly minimizes reconstruction loss $L_{\text{recon}}$ and stability prediction loss $L_{\text{stab}}$:[1] $$\mathcal{L} = \mathcal{L}{\text{recon}}(s + f, e) + \lambda, \Delta\Delta G)$$ This enforces that structural plus functional plausibility faithfully reconstructs the pLM's original evolutionary log-likelihood.[1]}} \mathcal{L}_{\text{stab}}(\hat{\Delta\Delta G

3. Per-Residue Function Score

To map functional importance at the residue level, DETANGO calculates a per-residue Function Score $F(i)$ as the negative mean functional plausibility across all 19 non-wild-type amino acid substitutions at position $i$:[1] $$F(i) = -\frac{1}{19} \sum_{a \in \mathcal{A} \setminus {x_i^{\text{WT}}}} f(x_i \to a)$$ Residues with high $F(i)$ harbor disproportionate densities of SBI variants and are inferred to be critical functional sites.[1]

Benchmark Performance and Validation

1. Classification of Stable-but-Inactive (SBI) Variants

Evaluated on a benchmark of 11 proteins with multiplexed assays of variant effect (MAVEs) measuring both functional activity and cellular abundance (PTEN, TP53, NUDT15, etc.):[1] - NUDT15 Case Study: While ESM-1v labeled most mutations as deleterious and FoldX identified only destabilizing variants (e.g., V17–V19, C58–A59), DETANGO correctly disentangled non-destabilizing catalytic variants (E65, E66, E112) as true SBI variants.[1] - Benchmark Superiority: Across all 11 MAVE proteins, DETANGO consistently outperformed unsupervised baselines (ESM-1v, ESM-2, ESM-C, FoldX, ESM-IF1, Func-ESM) across AUROC, AUPRC, and nDCG metrics, matching or exceeding supervised classifiers without using functional label supervision during training.[1]

2. Residue-Level Functional Site Identification

Evaluated on Human Domainome 1 (408 full-length human proteins, 213,332 residues, 17,381 CDD-annotated functional residues):[1] - CDD Annotation Enrichment: DETANGO assigned significantly higher function scores to annotated functional sites compared to unannotated residues ($P < 0.001$).[1] - Zinc-Binding Precision: 100% of the 168 CDD-annotated zinc-binding sites across the benchmark received positive DETANGO function scores.[1] - Regression Baselines: Outperformed post-hoc linear, logistic, and MLP regression models combining ESM-1v and FoldX, proving the superiority of representation-level disentanglement over post-hoc curve fitting.[1]

3. Allosteric Landscape Mapping & Protein Engineering Insights

  • KRAS Allosteric Mapping: Evaluated on multidimensional DMS data for KRAS bound to six effectors (RAF1, PIK3CG, RALGDS, SOS1, DARPin K27, DARPin K55).[1] DETANGO achieved higher AUPRC than ESM-1v or FoldX $\Delta\Delta G$ at identifying experimentally defined allosteric sites (6–22 sites per partner) and prioritized residues forming the binding pocket of sotorasib, a clinically approved allosteric KRAS(G12C) inhibitor.[1]
  • 17-Protein Allosteric Benchmark: Across 17 proteins with 159 experimentally validated allosteric sites, DETANGO-FoldX achieved superior AUROC over FoldX and ESM-1v.[1]
  • Local Energetic Frustration & Engineering Targets: Residues in the top 10% of DETANGO function scores were enriched for stabilizing mutations (harboring $\ge 2$ stabilizing substitutions), reflecting local energetic frustration required for catalysis.[1] Conversely, solvent-exposed surface residues ($\text{rSASA} \ge 0.25$) in the bottom 5% of DETANGO scores represent candidate sites for thermostability engineering without compromising function.[1]

Limitations and Availability

Code and model weights were evaluated from preprint documentation; repository release and peer-reviewed replication are pending.[1]

See Also

Citations

[1] Ding, Z., Li, Y., Tu, Z., Luo, Y., Luo, H. (2026). DETANGO: a disentangled protein language model separating stability and function. bioRxiv preprint, doi:10.64898/2026.02.03.703560. Supports: mathematical formulation, network architecture, and all benchmark findings above. Location: Methods (Deconvolution & Architecture); Results (SBI classification, Human Domainome 1, KRAS allostery).