Engineering

Data Profiling & Transformation

Our data foundation is a robust, large-scale dataset comprising 50 distinct protein datasets curated and processed from the ProteinGym benchmark—a comprehensive collection of deep mutational scanning (DMS) assays [3]. This diverse dataset encompasses thousands of variants across multiple protein families, functionally categorized according to experimental assays (e.g., thermal stability, binding affinity, growth). This structured approach ensures that models learn generalizable principles of protein fitness rather than overfitting to individual targets [4].

Through the curation of DMS assays and harmonization of labels by assay type (e.g., activity, thermal stability, binding), a multi-task training corpus suitable for both zero-shot and supervised settings was generated. Benchmarks such as ProteinGym provide a large and diverse ground truth—millions of mutants across hundreds of assays—enabling robust evaluation and supervised fine-tuning. The data were transformed through deduplication, mapping of assay readouts to normalized fitness scores, metadata harmonization (assay type, organism, experimental conditions), and sequence-level filtering (length, low-complexity masking).

Strategy Selection

Pre-trained LMs, such as the ESM series, provide transferable sequence embeddings and masked prediction capabilities that encode structural constraints. Fine-tuning on DMS data adapts these general-purpose representations into task-specific fitness predictors [5].

Our model selection focuses on ESM-2, which demonstrates robust zero-shot capabilities and supports masked token prediction for generation purposes [6]. We plan to employ supervised fine-tuning with contrastive learning on positive and negative samples. This approach not only enables the model to recognize patterns associated with high-fitness sequences but also actively avoids features linked to poor functionality, thereby significantly enhancing its discriminative power [7].

Point-by-point scanning mask prediction is selected for generating systematic variants: masked token prediction is efficient, interpretable (yielding per-site scores), and compatible with the training objectives of PLMs. Similar strategies have been successfully applied in protein design and variant prediction [8][9]. For a target sequence, PROTEUS will sequentially mask each residue position and query the fine-tuned PLM to propose alternative amino acids. By masking each residue, predicting substitutions, and mapping mutations, this process generates an interpretable library that outperforms random masking strategies.

Training the Scoring Function

First, we conducted benchmark tests and ablations on ProteinGym to ensure that the scoring function correlates with DMS readouts and improves upon zero-shot baselines.

The training of the scoring function proceeded through iterative DBTL cycles and involved two construction and optimization phases: (i) task-specific regression heads (e.g., XGBoost/LightGBM or simple MLPs) trained on ESM-2 embeddings for calibration per assay; and (ii) synthesis or integration of scorers to harmonize assay types into a unified, standardized cross-protein ranking score. This "computational referee" enables direct comparison of candidate designs across different functional objectives.

Model Fine-tuning

Model fine-tuning was iteratively refined through the DBTL cycle, progressing through four major construction and optimization phases:

Wet-lab Build

The gene sequences of the top-scoring variants and their wild-type counterparts were reverse-translated and codon-optimized (e.g., NovoPro). Primers were designed using SnapGene, and the selected vector backbone was prepared via PCR-based backbone amplification and ligation. For instance, in the construction of β-lactamase variants, these genes were cloned into the pHT01-ATAAAA plasmid, replacing the native ampicillin resistance gene. The constructed plasmids were verified by colony PCR and sequencing prior to transformation into competent E. coli cells (or appropriate B. subtilis) for expression, ensuring direct assessment of protein phenotypes and avoiding confounding background properties such as intrinsic drug resistance. For β-lactamase constructs, agarose gel electrophoresis confirmed successful PCR amplification of the gene fragment (~800 bp) and plasmid backbone (~7 kb), as well as successful ligation via colony PCR. Since the case study focused on β-lactamase—an antibiotic resistance determinant—risk minimization measures were implemented, including the use of non-clinical laboratory strains, physical containment, selection with lower-risk antibiotic concentrations, and prior approval from the Institutional Biosafety Committee.

Scoring Results & Structure Prediction

Validation Metrics: A rigorous three-tier benchmark (s3 [Experimental Score] > s2 [Control Score] > s1 [Baseline Score]) was adopted as the gold standard for "successful engineering." The merged scoring function was used to rank the candidate pool, and modifications were only considered successful when the condition s3 > s2 > s1 was strictly met, thereby robustly demonstrating the added value of our fine-tuning strategy. In our primary β-lactamase case study, tests conducted on 500 low-activity sequences from the relevant dataset A4GRB6_PSEAI_Chen_2020 revealed that our method enhanced the performance scores of 71.4% of these sequences. Across various datasets, the proposed approach yielded statistically significant gains compared to both random baselines and zero-shot controls, demonstrating that PROTEUS is highly successful.

Wet-lab Test

Initially, the wet lab team utilized structure prediction tools such as AlphaFold to assist in validation, aiming to detect overall structural disruptions (e.g., loss of domain core, disruption of active site geometry). Candidates with unsatisfactory structural rationality were deprioritized, a strategy that reduced the waste of wet lab resources.

Subsequently, tailored experimental plans were determined based on protein characteristics, selectively performing protein expression, purification, and functional validation. Functional readouts were selected according to specific properties: activity assays (enzyme kinetics), thermal stability (ΔTm, DSF), binding (ELISA or BLI), or growth-based selection (colony counting of β-lactamase-expressing strains on antibiotic plates).

As an example, experimental validation was conducted via an antibiotic resistance assay: Escherichia coli strains transformed with plasmids containing either the optimized β-lactamase gene, the native β-lactamase gene, or no plasmid (blank control) were cultured on ampicillin plates. Bacterial colony growth was observed and quantified, with mean values calculated. Results showed 233 colonies for the optimized strain, compared to 117 for the control and 10 for the wild type. The higher survival rate observed for the optimized β-lactamase variant demonstrates the success of the modeling approach in this case study.

Results Analysis

The test results provided clear validation for the PROTEUS pipeline. Summary statistics from the modeling group included macro-level tests showing improvement rates (e.g., a 71.4% success rate for the target set), which guided model weight updates and score recalibration. On ampicillin-containing medium, strains expressing optimized β-lactamase variants produced significantly more colonies, indicating enhanced antibiotic resistance conferred by PROTEUS-designed variants and confirming the model's predictive accuracy. All wet-lab measurements were integrated back into the scoring pipeline. For each tested variant, predicted scores were compared with measured outcomes, and correlations were calculated to classify improved versus non-improved variants, as well as to validate effects at each site. Experimental results (DMS-style readouts at tested positions) were added to the training corpus and used for retraining or further fine-tuning of the scoring function via transfer learning. ProteinGym-style cross-validation guided the detection of overfitting and the selection of retraining strategies. This iterative approach helped capture context-dependent effects and partially addressed the issue of epistasis identified through wet-lab data.

Model Optimization

The integrated scoring function was retrained using new wet-lab experimental labels. Combinatorial and iterative strategies were explored to capture epistatic effects (e.g., nested rounds of scanning performed on backgrounds already containing validated beneficial mutations). Additionally, scaling to larger foundation models or multimodal (sequence & structure) representations was implemented to better capture interaction effects. Plans were also documented for the integration of scoring functions, iterative model fine-tuning and RL hyperparameter optimization.

Future work in our project will focus on further improvements, including the exploration of more advanced base models such as ESM3 and the development of multi-objective optimization capabilities to simultaneously meet diverse design requirements such as activity, stability, and expressibility. We will also actively incorporate automated experimental platforms and high-throughput screening methods to significantly enhance the efficiency and data output of the DBTL cycle, thereby continuously improving the model's practical design performance.