LoMuS is introduced, a multi-representation-based deep learning model that predicts dataset-provided protein stability scores directly from the primary sequence that consistently gains across standard experimental stability benchmarks.
Abstract
Abstract Motivation Protein folding stability is a key determinant for understanding protein dynamics, including molecular function, pathogenicity, and protein engineering. Yet, accurate prediction of protein stability remains challenging due to high variability in available data, particularly when only sequence information is available and structural knowledge is limited or unavailable. In this work, we introduce LoMuS, a multi-representation-based deep learning model that predicts dataset-provided protein stability scores directly from the primary sequence. In the core of the model architecture, a fusion network integrates explicit physicochemical descriptors with low-rank adapted protein language model derived embeddings from the sequence that consistently gains across standard experimental stability benchmarks. Results We rigorously evaluate LoMuS across multiple settings, such as absolute folding stability scoring, mutation landscape stability scoring, held-out protein domains, out-of-distribution label regimes, and per-protein evaluation. LoMuS consistently outperforms sequence-only baselines, achieving an absolute performance gain of at least 10% in Spearman’s rank correlation across several benchmarks. Per-protein evaluations further demonstrate robust performance gains. Ablation analyses confirm that complementary signals from physicochemical descriptors and sequence embeddings are critical to the effectiveness of the proposed multi-representation approach. We believe LoMuS advances protein engineering research by improving the prediction and ranking of protein stability scores. Availability All codes including data preparation scripts, training and validation recipes, and experimental configurations for LoMuS are available at: https://github.com/kabir-ai2bio-lab/LoMuS.
Performance estimates in protein function prediction depend not only on model choice but also on upstream decisions that define the learning problem. Using antioxidant protein classification as a controlled case study, we evaluated how dataset harmonisation, protein representation, redundancy control, and partitioning strategy affect protein machine learning pipelines. We integrated 18,804 records from 12 publicly available dataset entries into a curated consensus dataset of 4,193 protein sequences. One-hot encoding and six pretrained protein language model representations were evaluated as model inputs and as similarity spaces for redundancy reduction and distance-aware splitting. Representation choice substantially altered dataset geometry, retained dataset size, class balance, and downstream evaluation. At representation-specific p90 thresholds, one-hot encoding retained the complete dataset, whereas pretrained embeddings retained between 5% and 25% of sequences. Distance-aware partitioning reduced apparent performance relative to random splitting by up to 0.15 MCC before redundancy control, while this difference narrowed after similarity filtering. Selected configurations nevertheless maintained high performance under stricter evaluation, reaching an MCC of 0.84. These findings show that performance estimates should be interpreted as outcomes of complete data-centric workflows rather than isolated properties of predictive models.
Three modeling frameworks are developed, including models based on handcrafted features, models using embedding representations extracted from ProteinMPNN, and ensemble models integrating a diverse set of state‐of‐the‐art predictors integrating a diverse set of state‐of‐the‐art predictors.
Yang Liu, Jian Zhang, Minghui Li· Protein Science· 0 citations
UniStab is introduced, an end-to-end framework for predicting stability changes across all mutation types by leveraging the implicit geometric reasoning of a pre-trained folding model and demonstrates state-of-the-art performance, particularly in the challenging scenarios of multi-point mutations and indels.
Hong Tan, Shenggeng Lin, Yi Xiong· Chemical Science· 0 citations
Predicting protein stability, like changes in melting temperature (ΔTm) caused by mutations, is a critical task in therapeutic protein engineering and drug discovery. This is reflected by a growing solution space, including both AI-based sequence and structure based methods. This paper demonstrates that accurate ΔTm prediction does not require structural input features, but can achieve state-of-the-art results with a careful training design for large sequence-based protein language models. We combine an autoresearch-inspired setup search with controlled ablation studies and show that a well-tuned sequence-only ESM2-650M model [6] outperforms structure-informed methods in our benchmark, achieving the lowest error (MAE/RMSE) and competitive Pearson correlation without pH or structural inputs. We further show that choices such as loss function, pooling strategy, auxiliary supervision, and finetuning regime materially affect performance.
Daniel Siegismund, Mario Wieser, E. Natali et al.· bioRxiv· 0 citations
Accurate protein function prediction (PFP) is essential for understanding biological systems. However, structure-based graph neural networks often rely on fixed-distance contact maps, which may inadequately capture continuous, multi-scale spatial topologies, while the long-tail distribution of Gene Ontology (GO) labels may bias prediction toward frequent functions. We propose DHST, a deep hybrid structure–topology framework that integrates sequence semantics from a pretrained protein language model with local structural information learned by a residual graph convolutional network. DHST further introduces site-specific persistent homology to encode multi-scale topological invariants and a topology-guided residue-wise gated fusion module to modulate structure–semantics representations using local topological embeddings. The fused residue features are aggregated through dual-path pooling, and a weighted binary cross-entropy loss is used to mitigate the adverse effects of label imbalance. On the PDB dataset, DHST achieved area under the precision–recall curve (AUPR) scores of 0.779, 0.481, and 0.557 for molecular function (MF), biological process (BP), and cellular component (CC), respectively; on the AF2 dataset, the corresponding scores were 0.729, 0.390, and 0.459. The model also demonstrated robust generalization to low-homology proteins and maintained strong predictive performance across GO terms with different levels of functional specificity. Ablation results supported the contributions of the main components.
Bin Lu, Fujun Xiang, Hai-Long Wang et al.· Applied Sciences· 0 citations
Sequence-based protein druggability classification can support early target triage when structural information is unavailable, uncertain, or inconsistently linked to druggability labels. We present DrugPLMFormer, a sequence-first retrospective screening framework that combines frozen protein language model embeddings with self-attentive BiLSTM encoding, Transformer-based long-range modeling, optional physicochemical feature fusion, and compute-budgeted BO–CTCM model selection. Hyperparameters were selected through multi-fidelity screening within an approximately 200-evaluation budget, using a validation objective that combined AUPRC and MCC to balance threshold-free discrimination with operating-point stability, rather than to imply unrestricted generalization. On ProTar-II, using a 50% sequence-identity homology-aware split, DrugPLMFormer achieved 95.98% accuracy, 96.01% F1-score, 96.42% sensitivity, 95.61% specificity, and 0.981 ROC-AUC. Without using external data for training, tuning, threshold selection, or early stopping, the selected model showed favorable held-out mean performance on ProTar-II-Ind (96.62% accuracy, 0.9688 ROC-AUC) and DPI_CDF (96.20% accuracy, 0.9696 ROC-AUC). Paired external analyses indicated that accuracy and F1-score differences were numerically favorable but not statistically significant, whereas the ROC-AUC improvement on DPI_CDF was statistically supported. Train-to-external homology analysis showed that most external proteins had less than 50% sequence identity to the training set, although residual dataset shift and label heterogeneity may still affect generalization. With cached PLM embeddings, downstream CPU inference required approximately 1.0–1.2 ms per sequence, excluding tokenization and ESM-2 embedding generation. Overall, DrugPLMFormer provides a reproducible, leakage-aware framework for retrospective sequence-based druggability screening and target prioritization, while prospective validation and experimental confirmation remain necessary before operational deployment.
Zahra Kafi, Khosro Rezaee, Hossein Eslami· Journal of King Saud Univers...· 0 citations