find evaluation reports for AI used in biological studies

Evaluation reports for artificial intelligence (AI) and machine learning (ML) in biological studies encompass a broad range of applications, from protein structure prediction and genomic regulatory element identification to clinical diagnostics and risk modeling. These reports utilize standardized benchmarking datasets, quality assessment tools like PROBAST and QUADAS-AI, and reporting frameworks such as DOME and REFORMS to evaluate accuracy, generalizability, and reproducibility.

Structural Biology and Protein Prediction

Recent evaluations of AI in structural biology focus on the accuracy of 3D-structure prediction for proteins and biomolecular complexes.

• AlphaFold2 (AF2) Performance: A community assessment found that AF2 adds approximately 25% novel residue coverage to model proteomes with high confidence (pLDDT > 70) compared to previous homology models (Direct, High; PMID: 36344848). While highly accurate for stable domain geometry, AF2 shows performance declines for intrinsically disordered regions (IDRs), ligand-bound states, and large multi-component assemblies (Direct, High; PMID: 41326937, PMID: 36344848).
• Boltz-1 Benchmarking: The open-source model Boltz-1 was benchmarked against commercial tools like Glide and traditional tools like AutoDock Vina (Direct, High; DOI: 10.26565/2220-637x-2025-44-06). Boltz-1 achieved protein root mean square deviation (RMSD) values of less than 1 Å and outperformed AutoDock Vina in ligand-binding rankings across diverse datasets (Direct, High; DOI: 10.26565/2220-637x-2025-44-06).
• RNA Structure Prediction: Benchmarking deep learning for RNA structure remains challenging due to limited data in the Protein Data Bank (PDB) compared to proteins (Direct, High; PMID: 38552946). The RNA3DB dataset was developed to address structural dissimilarity and prevent data leakage during training (Direct, High; PMID: 38552946).

Genomics, Transcriptomics, and Regulatory Elements

AI evaluation in genomics focuses on identifying transcription factor binding sites (TFBS), single-cell analysis, and spatial data recovery.

• TFBS Prediction: A comparative analysis demonstrated that Support Vector Machine (SVM)-based models generally outperform Position Weight Matrices (PWMs) in most scenarios, particularly when evaluating imbalanced datasets using the Area Under the Precision-Recall Curve (AUPRC) (Direct, High; PMID: 40702706).
• Spatial Transcriptomics Deconvolution: The ST-deconv approach utilizes self-encoding and contrastive learning to improve the resolution of spatial transcriptomic data (Direct, High; PMID: 40896262). Benchmarking against methods like cell2location and SPOTlight showed ST-deconv consistently exhibited lower error rates (RMSE) and higher clustering purity in mouse olfactory bulb and cancer datasets (Direct, High; PMID: 40896262).
• Bacterial Variant Calling: Deep learning variant callers like Clair3 and DeepVariant achieved median F1 scores of 99.99% for SNPs in bacterial nanopore sequencing data, significantly outperforming traditional Illumina-based methods in repetitive or variant-dense regions (Direct, High; PMID: 39388235).

Clinical Diagnostics and Medical Imaging

Systematic reviews and meta-analyses have evaluated the diagnostic accuracy of AI in specialized clinical contexts.

• Bone Metastases: A systematic review of 16 studies found that AI models achieved a pooled sensitivity of 0.88 and an AUC of 0.95 for detecting tumor bone metastases on CT and MRI, performing comparably or superiorly to radiologists (Direct, High; PMID: 39966724).
• Risk Prediction Models: Evaluation of 29 models for Ovarian Hyperstimulation Syndrome (OHSS) identified high risks of bias in most studies due to retrospective designs and lack of external validation (Direct, High; PMID: 40826343). Similarly, reporting quality for AI in pediatric diabetes (PMID: 38241075) and patient-reported outcomes in oncology (PMID: 39499409) remains generally low, often failing to address data transformations or model robustness (Direct, High; PMID: 38241075, PMID: 39499409).

Methodological Frameworks and Reporting Standards

Standardized reporting is essential for addressing the "black box" nature of AI and identifying "shortcut learning," where models rely on unintended dataset artifacts (Direct, High; DOI: 10.1038/s42256-020-00257-z).

• DOME Recommendations: These community-driven standards cover Data, Optimization, Model, and Evaluation to enhance reproducibility in biological ML (Direct, High; DOI: 10.1038/s42256-020-00257-z).
• REFORMS Checklist: This 32-item framework was established to prevent common errors in ML-based science, focusing on study goals, computational reproducibility, and data leakage (Direct, High; PMID: 38691601).
• Interpretability: Studies utilize "post-hoc" methods (e.g., feature importance scores) and "model-based" methods (e.g., modular designs) to extract biological meaning from trained models (Direct, High; PMID: 36387282). For example, the CeLEry model provides uncertainty estimates for recovered cell locations in single-cell RNA-seq (Direct, High; PMID: 37422469).

How do the DOME recommendations differ between their general application and their specific interpretation for proteomics and metabolomics?

What are the primary sources of data leakage identified in the REFORMS checklist and how can they be mitigated?

Which quantitative performance metrics are recommended for evaluating AI models used in imbalanced biological datasets?

Unverified Citations

The following sources failed to support their assigned claims after 3 verification rounds designed to ensure only high-confidence, relevant references are retained:

• PMID:34316068 — Standardized reporting is essential for addressing the "black box" nature of AI and identifying "shortcut learning," whe...
  Failed: conclusion — The cited paper provides DOME recommendations for reporting standards but does not mention the term 'shortcut learning' or the identification of dataset artifacts as a primary goal.

Additional evaluation reports for AI in biological studies highlight the impact of specialized architectures like transfer learning and graph neural networks, the critical role of data preprocessing in metabolomics, and the performance of specific algorithms in drug repurposing and surgical modeling.

Specialized Biological Model Evaluations

Benchmarking has revealed that deep learning (DL) models frequently outperform traditional methods when handling high-dimensional or small-scale biological datasets.

• Drug Repurposing (DR): A comprehensive evaluation of ten methods across eight datasets identified Overlap Matrix Completion (OMC) and Bound Nuclear Norm Regularization (BNNR) as top-performing approaches, generally exceeding the accuracy of Matrix Factorization and standard deep learning models like DRDM (Direct, High; PMID: 41366216).
• RNA-Binding Protein (RBP) Prediction: The RBP-TSTL framework, which utilizes a two-stage transfer learning strategy, was benchmarked against nine existing predictors across four species. It demonstrated superior performance in terms of the Area Under the Precision-Recall curve (AUPRC), especially in cases where annotated data for specific species like E. coli or Salmonella were scarce (Direct, High; PMID: 35649392).
• Biological Sequence Classification: The BBATProt framework was evaluated across five independent datasets, including carboxylesterases and antimicrobial peptides. It integrated CNN, Bi-LSTM, and Temporal Convolutional Network (TCN) layers, showing that hierarchical feature extraction improved robustness across diverse functional prediction tasks (Direct, High; PMID: 41212592).

Data Preprocessing and Architectural Insights

Evaluations show that the choice of data preparation and model architecture is as critical as the algorithm itself for accuracy and interpretability.

• Metabolomics Preprocessing: A comprehensive evaluation of 5,000 models revealed that sampling-based methods for missing value imputation are superior to traditional probabilistic models or filling with zeros (Direct, High; PMID: 35323644). Furthermore, log fold-change transformations provided the most consistent performance for classification tasks, while biomass normalization had only a subtle influence on overall accuracy (Direct, High; PMID: 35323644).
• Surgical Modeling: Benchmarking of machine learning models for predicting surgical case duration across 14 studies indicated that tree-based algorithms (e.g., XGBoost, CatBoost) were generally more accurate than standard historical averaging or deep learning multilayer perceptrons, especially when training datasets contained fewer than 1,000 records (Direct, High; PMID: 37931236).
• Scaling Factors: In immunological modeling of TLR-4 signaling, feature scaling was found to be essential for the accuracy of neural networks (NN) but did not impact the performance of Support Vector Machines (SVM) or Naive Bayes (NB) (Direct, High; PMID: 40360553).

Clinical Utility and Reporting Challenges

Clinical evaluation reports emphasize the gap between high statistical accuracy and practical implementation.

• Medical Imaging Performance: In ultrasound elastography for breast tumor classification, deep learning models achieved a pooled sensitivity of 0.94, outperforming traditional machine learning (0.87). However, the report noted that internal validation often yields overoptimistic results compared to external validation (Direct, High; PMID: 41184781).
• Reporting Quality Discrepancies: In pediatric diabetes research, medical papers tended to have higher reporting quality than engineering journals, particularly in describing model examination techniques (Direct, High; PMID: 38241075).

What specific metrics are recommended by the DOME guidelines for evaluating regression models in biological studies?

How does the use of external validation datasets affect the sensitivity and specificity of AI-based diagnostic models in the provided evidence?

What are the comparative advantages of using graph neural networks versus convolutional neural networks in spatial transcriptomics deconvolution?

Unverified Citations

The following sources failed to support their assigned claims after 3 verification rounds designed to ensure only high-confidence, relevant references are retained:

• PMID:39705058 — 41% accuracy, significantly mitigating the overfitting typically seen in small, function-specific datasets
  Failed: conclusion — The claim of '41% accuracy' is not supported; the paper actually reports accuracies ranging from 62.62% to 83.41% depending on the model and cutoff.
  Possible alternatives (unverified): PMID:40826343 (36% topic match)
• PMID:39499409 — 8% of oncological machine learning studies provided open-source code or data
  Failed: conclusion — The paper reports that 5.8% (not 8%) of the studies provided open-source code or data.
  Possible alternatives (unverified): PMID:34316068 (35% topic match); PMID:35649392 (35% topic match)

Further evaluation reports on AI in biological studies extend into bioprocess optimization, the generation of synthetic molecular data, and specialized diagnostic frameworks for medical "black-box" models. These reports emphasize the integration of mechanistic systems biology with data-driven machine learning to enhance model reliability.

Specialized Domain Frameworks and Registries

The transition from general recommendations to field-specific implementation is facilitated by centralized registries and nuanced interpretations of reporting standards.

• The DOME Registry: This structured database implements community recommendations by providing unique identifiers and a "DOME score" for supervised ML publications in biology (Direct, High; PMID: 39661723) «✓ PMID:39661723». The DOME score is calculated as the number of valid answers to the recommendations divided by the total number of questions, offering a preliminary measure of transparency and quality (Direct, High; PMID: 39661723) «✓ PMID:39661723».
• Proteomics and Metabolomics BOLOs: Specific interpretations of the DOME recommendations for mass spectrometry-based fields emphasize "Be on the lookout for" (BOLO) items (Direct, High; PMID: 35119864) «✓ PMID:35119864». These include ensuring training and test data are disjoint at the molecular structure level (e.g., preventing stereoisomers from biasing statistics) and accounting for instrument performance events over time, such as maintenance or calibration shifts (Direct, High; PMID: 35119864) «✓ PMID:35119864».

Bioprocess and Enzyme Design Evaluations

Benchmarking in biocatalysis and manufacturing focuses on the synergy between physics-based simulations and deep learning.

• CHO Cell Medium Optimization: Evaluation of culture medium design for monoclonal antibody (mAb) production in Chinese hamster ovary (CHO) cells highlights the use of Constraint-Based Modeling (CBM) and Kinetic Modeling (KM) (Direct, High; PMID: 39571767) «✓ PMID:39571767». While CBM methods like Flux Balance Analysis (FBA) are computationally efficient for genome-scale models, they often rely on idealized assumptions; hybrid frameworks that integrate CBM with ML (e.g., using Random Forest or Deep Neural Networks to predict amino acid consumption) improve dynamic accuracy in industrial settings (Direct, High; PMID: 39571767) «✓ PMID:39571767».
• BioStructNet Benchmarking: This structure-based network for enzyme function prediction was validated by comparing its attention weights to key residue sites identified through molecular dynamics (MD) simulations (Direct, High; PMID: 39705058) «✓ PMID:39705058». BioStructNet demonstrated that capturing local protein-ligand interaction patterns via bilinear attention was superior to global self-attention for small, high-similarity datasets like Candida antarctica lipase B variants (Direct, High; PMID: 39705058) «✓ PMID:39705058».

Generative Modeling and Data Augmentation

AI evaluations include the ability of generative models to produce realistic synthetic data, which serves to augment small biological datasets.

• DeepImmuno-GAN: For T-cell immunity, a Generative Adversarial Network (GAN) was benchmarked on its ability to produce synthetic immunogenic peptides (Direct, High; PMID: 33398286) «✓ PMID:33398286». Results showed that 87% of the generated pseudo-sequences achieved >60% similarity to real validated peptides, demonstrating that immunogenic motifs are learnable and can be used to expand training repertoires for prediction models (Direct, High; PMID: 33398286) «✓ PMID:33398286».
• Single-Cell Augmentation: Generative models such as ACTIVA (introspective VAE) and scGAN (Wasserstein-GAN) have been evaluated for their capacity to generate rare cell populations in scRNA-seq data (Direct, High; PMID: 37393865) «✓ PMID:37393865». ACTIVA was reported to train up to 17 times faster than scGAN while maintaining comparable generation quality, facilitating the benchmarking of downstream classifiers and marker gene detection (Direct, High; PMID: 37393865) «✓ PMID:37393865».

Interpretability and Bias Detection

Evaluations of medical AI emphasize the detection of "shortcut learning" and the use of explainability tools to build clinical trust.

• Explainable AI (XAI) Methods: Systematic reviews of medical ML identify SHAP (SHapley Additive exPlanation), LIME (Local Interpretable Model-agnostic Explanation), and Grad-CAM as essential for de-black-boxing models (Direct, High; PMID: 38227273) «✓ PMID:38227273». These techniques help clinicians identify if a model is relying on clinically irrelevant features, such as hospital metal tokens in pneumonia scans (Direct, High; DOI: 10.1038/s42256-020-00257-z).
• Morgan’s Canon for ML: To prevent over-interpreting AI results, researchers recommend applying "Morgan's Canon": never attribute high-level biological "understanding" to a model if its performance can be adequately explained by lower-level shortcut learning or simple correlations (Direct, High; DOI: 10.1038/s42256-020-00257-z) «✓ DOI:10.1038/s42256-020-00257-z».

How does the DOME score calculation in the registry quantify adherence to best practices in supervised machine learning?

What are the specific "Be on the lookout for" (BOLO) criteria for training and test data partitioning in proteomics studies?

Which deep learning architectures are most effective for generating realistic synthetic single-cell RNA-sequencing data?

Evaluation reports for artificial intelligence (AI) in biological studies further detail performance benchmarks in spatial cellular mapping, genomic variant accuracy in nanopore sequencing, and the sensitivity of models to specific study design parameters. These reports emphasize the role of adversarial training and contrastive learning in improving the fidelity of computational reconstructions.

Spatial Recovery and Deconvolution Benchmarks

Specialized models have been evaluated for their ability to recover spatial information and cell type compositions in low-resolution transcriptomic data.

• Cell Location Recovery (CeLEry): Benchmarking of the CeLEry framework against Tangram and novoSpaRc demonstrated superior performance in 2D coordinate recovery in mouse brain datasets (Direct, High; PMID: 37422469). In a "Scenario 1" test, CeLEry achieved a Top-1 accuracy of 53.8%, which increased after incorporating data augmentation via a variational autoencoder (VAE).
• Adversarial Mitigation: The inclusion of a Domain-Adversarial Neural Network (DANN) module in ST-deconv significantly reduced deconvolution errors (RMSE) from 0.0526 to 0.0507, highlighting its efficacy in bridging distributional gaps between simulated and real biological data (Direct, High; PMID: 40896262).

Genomic Variant and Binding Site Accuracy

Benchmarking efforts in genomics focus on identifying single nucleotide polymorphisms (SNPs), insertions/deletions (indels), and transcription factor binding sites (TFBS).

• Bacterial Variant Calling: Deep learning variant callers (Clair3 and DeepVariant) were benchmarked using Oxford Nanopore R10.4.1 data across 14 bacterial species. Clair3 achieved a median F1 score of 99.99% for SNPs and 99.53% for indels (Direct, High; PMID: 39388235). These tools remained robust at low read depths, with 10x "super-accuracy" (sup) data matching the performance of full-depth Illumina sequencing (Direct, High; PMID: 39388235).
• TFBS Models: A comparative analysis of Position Weight Matrix (PWM), SVM, and Deep Learning models showed that SVM-based models (LS-GKM) are highly sensitive to sequence width, requiring full-length sequences for optimal performance on biological backgrounds (AUPRC = 0.96) (Direct, High; PMID: 40702706). Transformers like Nucleotide Transformer (NT) similarly showed consistent performance gains as input sequence length increased (Direct, High; PMID: 40702706).

Interpretable Systems Biology and Clinical Modeling

Evaluations in systems biology characterize how models identify biological mechanisms and clinical phenotypes.

• Feature Contribution Analysis: Models predicting schizophrenia and lung cancer utilized "Integrated Gradients" and "Conductance" methods to identify specific SNPs and genes influencing model output (Direct, High; PMID: 36387282). This approach allowed researchers to evaluate the importance of biologically meaningful connections in a sparse deep neural network (Direct, High; PMID: 36387282).
• Clinical Protocol Comparison: A systematic review of chronic kidney disease (CKD) prediction found that machine learning methods often showed no distinct performance advantage over conventional logistic regression for clinical prediction models (Direct, High; PMID: 36821539).

Sensitivity to Study Design and Data Constraints

Research on the reliability of biological AI highlights the impact of hyperparameters and data partitioning.

• Training Proportion Impact: In immunological signaling models, increasing the training data proportion significantly improved accuracy across all classifiers (RF, GLM, NN, SVM, NB) (Direct, High; PMID: 40360553). Specifically, RF required over 80% of data for training to reach high accuracy on protein-level datasets, while GLM and NN were more robust to smaller splits (Direct, High; PMID: 40360553).
• Missing Data Management: In protein-level datasets, pruning features with more than 70% "missingness" was a critical step for ensuring the stability of downstream machine learning modeling (Direct, High; PMID: 40360553).

How does the integration of Domain-Adversarial Neural Networks (DANN) improve the accuracy of spatial transcriptomic deconvolution in the ST-deconv framework?

What are the observed differences in performance between deep learning-based variant callers and traditional methods in repetitive genomic regions?

What role do "Integrated Gradients" and "Conductance" play in identifying high-impact SNPs and genes in the interpretable machine learning models discussed?

Unverified Citations

The following sources failed to support their assigned claims after 3 verification rounds designed to ensure only high-confidence, relevant references are retained:

• PMID:37422469 — 2% after incorporating data augmentation via a variational autoencoder (VAE)
  Failed: conclusion — The claim text is fragmented and makes a quantitative assertion ('2%') that is not supported by the paper's reported accuracy improvements (which were 9.4% and 7.2% for Scenarios 1 and 2, respectively).
  Possible alternatives (unverified): PMID:37393865 (100% topic match); PMID:41332459 (100% topic match)

Evaluation reports for artificial intelligence (AI) and machine learning (ML) in biological studies provide performance benchmarks across structural biology, genomics, clinical diagnostics, and drug discovery. These reports emphasize that while deep learning (DL) often achieves superior statistical accuracy, the validity of these models depends heavily on data quality, reporting transparency, and the mitigation of "shortcut learning" (Direct, High; DOI: 10.1038/s42256-020-00257-z, PMID: 38691601).

Structural Biology and Biomolecular Interaction Benchmarks

Evaluation of AI in structural biology highlights the capabilities of deep learning in folding and docking across diverse molecular classes.

• AlphaFold2 (AF2) Community Assessment: AF2 demonstrated high accuracy in protein domain prediction (median backbone accuracy of 0.96 Å RMSD95) but showed performance declines in intrinsically disordered regions (IDRs), which correlate with low predicted local-distance difference test (pLDDT) scores (Direct, High; PMID: 36344848, PMID: 34265844). AF2 added ~25% novel high-confidence residue coverage to 11 model proteomes compared to existing homology models (Direct, High; PMID: 36344848).
• Boltz-1 Benchmarking: This open-source model achieved AlphaFold3-level accuracy, accurately reproducing protein folding (RMSD < 1 Å) and outperforming AutoDock Vina in predicting ligand-binding poses across diverse heterocyclic and macrocyclic molecules (Direct, High; DOI: 10.26565/2220-637x-2025-44-06).
• Biocatalyst Function Prediction: The BioStructNet framework, using bilinear attention mechanisms, outperformed global self-attention models for small, high-similarity datasets like Candida antarctica lipase B variants (Direct, High; DOI: 10.26565/2220-637x-2025-44-06). Reliability was validated by comparing attention weights to residue sites identified via molecular dynamics (MD) simulations (Direct, High; PMID: 39705058).
• RNA Structure Prediction: Benchmarking for RNA remains limited by the relative scarcity of RNA chains in the Protein Data Bank (~11,176 remaining after filtering) compared to proteins (Direct, High; PMID: 38552946). The RNA3DB dataset was developed to prevent data leakage by clustering RNA 3D structures into non-redundant components based on structural homology (Direct, High; PMID: 38552946).

Genomics, Transcriptomics, and Variant Calling Benchmarks

Evaluations in genomics focus on identifying sequence motifs, recovering spatial data, and assessing variant calling accuracy.

• Bacterial Variant Calling: Deep learning callers (Clair3, DeepVariant) achieved Top F1 scores (99.99% for SNPs; 99.53% for indels) on Oxford Nanopore R10.4.1 data, significantly outperforming traditional Illumina-based methods in repetitive or variant-dense regions (Direct, High; PMID: 39388235).
• Transcription Factor Binding Sites (TFBS): A comparison of Position Weight Matrix (PWM), SVM, and DL models showed SVM-based models (LS-GKM) achieve the best performance on long sequences (AUPRC = 0.96) and imbalanced datasets (Direct, High; PMID: 40702706).
• Spatial Transcriptomics (ST):
  • CeLEry: This framework achieved Top-1 accuracies (53.8% in brain tissue) for 2D location recovery, outperforming Tangram and novoSpaRc (Direct, High; PMID: 37422469).
  • ST-deconv: Utilizing contrastive learning and Domain-Adversarial Neural Networks (DANN) reduced deconvolution errors (RMSE) by aligning simulated and real data distributions (Direct, High; PMID: 40896262).
• RNA-Binding Proteins (RBPs): The RBP-TSTL framework, which uses two-stage transfer learning (ProtT5-XL embeddings followed by annotated RBP knowledge), outperformed nine sequence-based predictors across four species (Direct, High; PMID: 35649392).

Clinical Diagnostic and Risk Modeling Reports

Systematic reviews and meta-analyses provide quantitative assessments of AI diagnostic performance.

• Bone Metastases Detection: AI models achieved a pooled sensitivity of 0.88, specificity of 0.89, and AUC of 0.95 for detecting metastases on CT and MRI, showing comparable or superior performance to radiologists (Direct, High; PMID: 39966724).
• Breast Tumor Classification: In ultrasound elastography, deep learning models showed a higher pooled sensitivity (0.94) than machine learning (0.87) for differentiating benign and malignant lesions (Direct, High; PMID: 41184781).
• Ovarian Hyperstimulation Syndrome (OHSS): Evaluation of 29 risk prediction models revealed high risks of bias due to retrospective designs and small sample sizes (events per variable < 10 in many cases) (Direct, High; PMID: 40826343).
• Surgical Case Duration: Tree-based models (XGBoost, CatBoost) were found to be significantly more accurate than industry standards (historical averaging), reducing scheduling inaccuracies by up to 70% in some centers (Direct, High; PMID: 37931236).

Reporting Quality and Reproducibility Frameworks

Standardized guidelines address the "black-box" nature of AI and identifying "shortcut learning" (relying on irrelevant dataset artifacts).

• DOME Recommendations: A community-driven set of recommendations for reporting Data, Optimization, Model, and Evaluation in supervised biological ML (Direct, High; PMID: 34316068).
• REFORMS Checklist: A 32-item consensus-based framework for ML-based science focusing on study goals, computational reproducibility, data quality, and leakage (Direct, High; PMID: 38691601).
• MI-CLAIM: Minimum information checklist for clinical AI modeling; evaluation of 21 pediatric diabetes studies showed generally low reporting quality, particularly in model examination and cohort representativeness (Direct, High; PMID: 38241075).
• Shortcut Learning Mitigation: Reports suggest applying "Morgan's Canon" to prevent over-attributing high-level abilities to models whose performance can be explained by simple statistical correlations (Direct, High; DOI: 10.1038/s42256-020-00257-z).

What quantitative differences exist between deep learning and traditional machine learning models in the diagnosis of tumor bone metastases?

How do the DOME and REFORMS frameworks differ in their requirements for reporting data leakage and model validation?

What specific biological and technical factors are identified as the primary causes of performance declines in AlphaFold2 predictions?

Unverified Citations

The following sources failed to support their assigned claims after 3 verification rounds designed to ensure only high-confidence, relevant references are retained:

• PMID:39705058 — ** Boltz-1 Benchmarking: This open-source model achieved AlphaFold3-level accuracy, accurately reproducing protein ...*
  Failed: entities,conclusion — The paper describes the BioStructNet model and uses CalB as a case study, but does not benchmark or provide data for the Boltz-1 model.