Venus-TIGER

Protein expression is the gate every engineered protein must pass through: a designed enzyme or therapeutic antibody that folds beautifully in silico is worthless if it cannot be produced in soluble, functional form. Yet expression has attracted far less computational attention than stability or binding affinity, and the mutations governing it are hard to reason about — distributed across the sequence, weakly correlated with structure, and rarely conserved in ways evolutionary models capture directly.

Venus-TIGER — "Tool for Identification of Governing Expression Residues" — predicts which amino acid substitutions raise or lower expression, and which positions matter most. It was developed by Fan Jiang, Mingchen Li, and colleagues in Liang Hong's group at Shanghai Jiao Tong University, with collaborators at East China University of Science and Technology and the Shanghai Artificial Intelligence Laboratory, and posted to bioRxiv in January 2025.

Its central design choice is to treat the entire mutational landscape as the unit of learning. Rather than scoring one variant at a time, Venus-TIGER assembles a two-dimensional matrix — sequence positions along one axis, the twenty possible substitutions along the other — from a protein language model's own token predictions, and trains that pseudo-landscape to match the corresponding matrix of measured expression values. Loss is computed only where measurements exist, turning the sparsity of deep mutational scanning data from an obstacle into a masking pattern. Generalization is inherited from the underlying language model, so a checkpoint trained on one set of proteins can be applied to a target it has never seen.

#Key Features

• Landscape-level training: A single training example is a whole protein's position-by-substitution matrix, so the model learns the shape of an expression landscape rather than isolated variant scores.
• Masked landscape loss: A binary cross-entropy objective is evaluated only at measured entries, extracting signal from sparse deep mutational scanning data without imputing the unmeasured majority.
• Frozen backbone, no inference overhead: Feature extraction layers stay frozen and only the language head is trained, adding no parameters or compute at inference over the base language model.
• Zero-shot and few-shot transfer: The trained checkpoint predicts expression effects for proteins absent from training, and small amounts of target-specific data improve accuracy further when a full mutational scan is cost-prohibitive.

#Technical Details

The architecture is a frozen ESM-2 650M feature extractor paired with the trainable language modeling head inherited from ESM-2. The head's per-position logits over amino acid tokens are read directly as the predicted expression landscape and compared against the binarized ground-truth matrix. Training data comes from ProteinGym: 17 expression-related deep mutational scanning datasets spanning 71,125 variants across proteins expressed in bacterial, fungal, and mammalian hosts, including CP2C9, ERBB2, PTEN, RASK, S22A1, and VKOR1, binarized against the wild-type expression score. Evaluation uses a strict leave-two-out protocol — one dataset held out for validation and another for test, the remaining 15 for training — averaged across 272 validation-test combinations, with Adam at learning rate 1e-4 and early stopping.

Venus-TIGER reached an AUC of 0.759 on the largest dataset (S22A1, n = 9,803) and 0.749 on the smallest (OPSD, n = 165), while one-hot CNN and LSTM baselines fell near chance (AUC 0.487 and 0.504 on S22A1; 0.407 for the LSTM on OPSD) and MLP regressors on ESM-1v, ESM-2, and ProtT5 embeddings hovered around 0.5. Baseline precision, recall, and F1 approached zero on these imbalanced datasets, where beneficial mutations are rare. Two out-of-sample cases served as independent tests: T7 RNA polymerase (883 residues) and a 142-residue VHH single-domain antibody, with wet-lab measurements produced by the authors. Adding 20% to 80% of target-specific T7 RNAP data improved performance monotonically over the zero-shot baseline.

#Applications

The model is aimed at protein engineers choosing which variants to build. Expression yield is the first filter in nearly every biologics, enzyme, and biocatalysis campaign, and ranking single-site mutants computationally shifts screening burden away from the bench. The two validation targets map onto the main use cases: T7 RNA polymerase represents workhorse enzymes whose titer determines reagent cost, the VHH nanobody biologics whose manufacturability decides whether a candidate advances. The residue-level output also flags positions where loop placement, surface exposure, or amino acid identity systematically shift expression.

#Impact

Venus-TIGER extends the pseudo-likelihood machinery that made protein language models effective zero-shot variant-effect predictors to a property those models were not built to capture, by supervising the landscape matrix rather than individual scores. Its accompanying analysis of the sequence and structural correlates of expression, and of the relationship between stability, activity, and expression, addresses a property poorly characterized relative to folding and binding. The work is a preprint awaiting peer review, and its wet-lab validation rests on two proteins. No code repository or model weights have been released; supporting data accompanies the manuscript as supplementary information, under a CC BY-NC-ND license.