JanusDDG

JanusDDG predicts the change in folding free energy (ΔΔG) caused by amino acid substitutions, using only the protein sequence. It was developed by the Computational Biomedicine Unit at the University of Turin and released as a preprint in April 2025. The task matters for both protein engineering, where stabilizing substitutions are the goal, and clinical variant interpretation, where destabilization is a common mechanism by which missense variants cause disease.

The model's distinguishing feature is that it is built to obey the thermodynamics of the quantity it predicts. Because ΔΔG derives from Gibbs free energy, a state function, it must satisfy antisymmetry — the prediction for a mutation and its reverse must sum to zero — and transitivity across chains of mutations. Many published predictors violate both, inflating apparent accuracy by learning a bias toward destabilization. JanusDDG enforces antisymmetry architecturally rather than through data augmentation, so its direct and reverse predictions are exact negatives by construction, giving a direct-versus-reverse correlation of exactly -1.

The name refers to the two-faced Roman god: the model looks at the wild-type and mutant sequences simultaneously through what the authors call two-fronts multi-head attention. Queries and values are computed from the difference between wild-type and mutant embeddings while keys alternate between the two, letting the network read mutation-induced perturbation without discarding sequence context. It builds on DDGemb, an earlier sequence-based predictor that relies on the embedding difference alone.

#Key Features

• Architecturally enforced antisymmetry: The cross-interleaved attention design guarantees that the predicted ΔΔG of a mutation and of its reverse are exact negatives, rather than approximately so, eliminating a systematic bias present in most competing predictors.
• Single and multi-point mutations: The same network handles substitutions at one position or many, without a separate model or a per-mutation additivity assumption.
• Sequence-only input: Predictions require no experimental or predicted 3D structure, so the model applies to proteins whose structures are unknown while still matching structure-based methods.
• Fixed released checkpoints: The repository ships JanusDDG.pth and JanusDDG_fine_tuned.pth, applied to new proteins and mutations from a user-supplied CSV with no retraining step.

#Technical Details

Input representations come from ESM-2 at the 650M-parameter scale. Wild-type and mutant sequences are each embedded, their difference is computed as a third representation, and the wild-type and difference embeddings are passed through separate 1D convolutions for dimensionality reduction before entering the bidirectional cross-attention transformer. Global average and global max pooling reduce the attended representation, and a linear head emits the scalar ΔΔG. Training uses S2450, a curated 2,450-variant refinement of the S2648 single-mutation dataset; a fine-tuning stage extends the model to multi-point mutations, validated on the M28 set.

On the S669 blind test set of single-point variants, JanusDDG reaches a Pearson correlation of 0.69 over the combined direct and reverse sets and 0.55 on direct mutations alone, ahead of DDGemb (0.68 / 0.53), PROSTATA (0.65 / 0.49), and the structure-based ACDC-NN (0.61 / 0.46). On S461 it reports a Pearson correlation of 0.69 and Spearman 0.66, against 0.63 and 0.60 for the next-best structure-based methods; on S96 it reaches Pearson 0.52. For multi-point mutations, evaluated on PTmul-NR — a redundancy-filtered subset of 82 variants across 14 proteins built to remove overlap with the training data — the fine-tuned model achieves a Pearson correlation of 0.61.

#Applications

JanusDDG suits protein engineering campaigns where candidate stabilizing substitutions must be ranked before synthesis, particularly for enzymes and biologics whose structures have not been solved. Its multi-mutation capability fits directed evolution and combinatorial design workflows, where the effects of co-occurring substitutions rarely add linearly. In clinical genomics it supports variant interpretation by flagging missense changes likely to destabilize a protein fold, one of the recurring molecular mechanisms behind loss-of-function disease alleles. Usage is a single command over a CSV listing sequences and mutations in the format A30Y_C65G, which keeps the barrier low for wet-lab groups without deep learning infrastructure.

#Impact

JanusDDG sits within a growing line of work arguing that physical constraints belong inside the model rather than in a post-hoc correction, a position that matters here because antisymmetry violations are what make many published ΔΔG benchmarks difficult to compare. Its practical result — that a sequence-only predictor built on frozen protein language model embeddings can match structure-based tools on both single and multiple substitutions — extends stability prediction to the large fraction of proteins that lack experimental structures. It is distinct from the ΔΔG predictors already in the catalog, EBM-DDG and StaB-ddG, which both target binding free energy at protein-protein interfaces rather than folding stability. The code and both checkpoints are released under an MIT license with a Docker image and training data for reproduction, though the repository remains a research release with a small community footprint, and the work is a preprint whose benchmark results await peer review.