Novel protein folding predictor that shows that using a simpler model architecture one can get quite far.
Architecture/training: SimpleFold swaps AF2/RF-style pair reps, triangle updates, MSAs, and equivariant blocks for plain Transformer layers trained with a flow-matching objective to generate full-atom structures; rotational symmetry is handled via SO(3) augmentation.
Training data: It is not crystals-only like previous predictors, the model mixes ~160k PDB experimental structures with large distilled sets from AFDB SwissProt (~270k) and AFESM (≈1.9M; 8.6M for the 3B model), then finetunes on PDB + SwissProt. So practically this is not a head-to-head comparison with other methods as they started from the smaller x-al dataset.
Performance: It’s competitive but generally below AlphaFold2/RoseTTAFold2/ESMFold on CAMEO22, while on CASP14 the 3B model beats ESMFold but does not surpass AlphaFold2; overall they claim ~95% of AF2/RF2 on most metrics, with especially strong results for ensemble generation.
Benchmarking of computational models for predicting antibody aggregation propensity (developability) using size-exclusion chromatography (SEC) readouts.
Developed an experimental dataset of ~1,200 IgG1 antibodies, measured for monomer percentage and ΔRT (difference in retention time) relative to a reference.
Evaluated four main prediction pipelines: Sequence + structure-based features (hand-crafted biophysical features from Schrödinger, using AlphaFold2 or ImmuneBuilder for structure). PLM (protein language model) pipeline (e.g., ESM2-8M, fine-tuned or LoRA-adapted). GNN (graph neural network) pipeline using residue graphs from predicted structures. PLM + GNN hybrid pipeline combining sequence embeddings with structural graphs.
Two structure prediction tools were benchmarked: AlphaFold2 (high accuracy, slow) and ImmuneBuilder (faster, antibody-optimized, slightly less accurate).
The sequence + structure feature model achieved the highest accuracy overall, but low sensitivity (missed many problematic antibodies).
The PLM-only pipeline performed nearly as well and offered a much faster, high-throughput solution, making it attractive for early screening.
The GNN and PLM + GNN approaches performed comparably, with GNN slightly better for ΔRT predictions but more variable.
Using ImmuneBuilder instead of AlphaFold2 reduced sensitivity slightly but greatly improved speed without major loss of accuracy.
So all pipelines performed similarly within a narrow performance range, but faster, less resource-intensive approaches (PLM and ImmuneBuilder-based pipelines) offer strong trade-offs for early-stage developability screening.
Investigation how biases in the Observed Antibody Space (OAS) database, such as overrepresentation of a few donors and limited species or chain diversity, affect the performance and generalizability of antibody language models.
The authors developed OAS-explore, an open-source pipeline to analyze, filter, balance, and sample OAS data by donor, species, chain type, and publication, enabling systematic assessment of data biases.
By training 17 RoBERTa models on datasets with different compositions, they found that models struggle to generalize across chain types, species, individuals, and batches, and that even increased donor diversity alone does not guarantee better performance.
They recommend systematic preprocessing, inclusion of more diverse data, and open sharing of datasets and pipelines to mitigate biases and improve antibody LM robustness.
Review of currently available large scale software for antibody analysis.
Today’s biologics R&D is slowed by fragmented tools and manual data wrangling; the paper proposes a unified, open-architecture platform that spans registration, tracking, analysis, and decisions from discovery through developability.
Key components are end-to-end registration of molecules/materials/assays; a harmonized data schema with normalized outputs; automated analytics with consistent QC; complete metadata capture and “data integrity by design.”
The platform should natively interface with AI, enable multimodal foundation models and continuous “lab-in-the-loop” learning, and support federated approaches to counter data scarcity while preserving privacy.
Dotmatics, Genedata, and Schrödinger each cover pieces (e.g., LiveDesign lacks end-to-end registration), and the authors stress regulatory-ready features.
Trained on SAbDab with a time split—6,448 heavy+light complexes + 1,907 single-chain (nanobodies), clustered at 95% ID into 2,436 clusters; val/test are 101 and 60 complexes, plus 27 nanobodies.
A two-stage diffusion (structure→seq+structure) followed by consistency distillation, epitope-aware conditioning, frozen ESM-PPI features, and mixed task sampling (CDR-H3 / heavy CDRs / all CDRs / no seq).
Antigen structure (can warm-start from AlphaFold3) + VH/VL framework sequences; you pick which CDRs (and lengths) to design; model outputs CDR sequences and the full complex.
Runs without an epitope but docking drops (DockQ ~0.246 → 0.069, SR 0.433 → 0.050); AF3 initialization lifts success to 0.627 (≈+0.19 vs baseline).
Novel open nanobody design method with experimental validation.
On the surface it might appear like a lot of methods stitched together. The magic sauce appears to be in the joint, gradient-based co-optimization: AF-Multimer and IgLM gradients are merged through a 3-phase schedule (logits → softmax → semi-greedy), with CDR-masking/framework bias and custom losses that force CDR-mediated, loop-like interfaces; then AbMPNN edits only non-contact CDR residues, and designs are filtered independently with AF3 + PyRosetta.
All this is actually not a ‘trained’ model but rather a filtering pipeline that WAS NOT trained (using previous methods, gradients, weights etc.) Just validated experimentally.
Experimental benchmark was ran on four targets: PD-L1, IL-3, IL-20, and BHRF1.
Authors measured how different their designs weren’t just ‘regurgitations’ of known abs. CDR identities were computed against SAbDab and OAS (via MMseqs); many designs show <50% CDR identity to any public sequence.
Novel de novo antibody design method with massive experimental testing.
The computational method involves integration, not retraining, of existing tools. It combines AlphaFold-Multimer, protein language models (ESM2/AbLang2), and NanoBodyBuilder2 with templating/sequence priors to design/filter antibody-format binders.
They perform massive testing. >1.1 million VHH binders designed across 436 targets (145 tested); ~330k experimentally screened.
Hit rates look low per binder (~0.5–1%) but that’s ~50× above random libraries, and still yields thousands of validated binders.
Target-level success is 45%, for how many targets we got binders; some epitopes reached 30–38% hit rates after filtering.
The big caveat is the specificity of epitopes- it really makes a difference, with some epitopes producing nought.
Novel model to predict the heavy/light chain compatibility
Data: H/L with the same single-cell barcode; negatives = swap L chains between pairs but only if CDRL3 length matches; balanced set of 233,880 pairs with a 90/10 train–test split.
Training: Full VH+VL into AntiBERTa2 with a classification head; fine-tuned 3 epochs, lr 2×10⁻⁵, weight decay 0.01; κ/λ-specific variants trained identically. Final AUC-ROC 0.75 (withheld) and 0.66 (external); κ/λ models: 0.885/0.831.
Baselines: (i) V/J gene-usage → logistic reg. & XGBoost ≈ 0.50–0.52 acc.; (ii) CDRH3+CDRL3 CNNs → moderate; (iii) ESM-2 improves with fine-tuning but AntiBERTa2 FT is best.
It seems to do better than just ‘matching to the database’. Weak gene-usage baselines, explicit control of CDRL3 length in negatives, external generalisation, and sensitivity to interface residues (CDRH1/2 & framework) in therapeutic-antibody tests argue the model learns sequence-level pairing rules, not just V/L distributions.
Introduces TNP, a nanobody-specific developability profiler inspired by TAP.
Uses six metrics: total CDR length, CDR3 length, CDR3 compactness, and patch scores for hydrophobicity, positive charge, and negative charge.
Thresholds are calibrated to 36 clinical-stage nanobodies.
In vitro assays on 108 nanobodies (36 clinical-stage + 72 proprietary) show partial agreement with TNP flags, indicating complementary—but not perfectly correlated—assessments.