This week leaned toward protein–ligand work, with a mix of practical benchmarking, efficient adaptation, and search methods, plus a few useful signals from protein language models and peptide design.

1. A Two-Stage ESM-Based Machine Learning Pipeline for Robust Hierarchical Enzyme Function Prediction
   Builds a simple enzyme annotation pipeline by feeding ESM-2 protein embeddings into a two-stage random forest that predicts enzyme class from broad to specific levels. On 25,778 sequences, it reports weighted F1 from 0.94 at the top EC level to 0.90 at EC4, which makes it a strong practical baseline for teams that want accuracy without end-to-end model training.

2. MULTIMAP TARGETED FREE ENERGY ESTIMATION: A MECHANICAL EMBEDDING QM/MM SCHEME FOR PROTEIN-LIGAND AFFINITIES
   Tests a targeted free energy correction scheme that adds quantum-level detail to protein–ligand affinity estimates without fully re-running the simulation stack. On p38 inhibitors, it improves overlap between force-field and QM/MM states, which matters because better overlap usually means more stable and affordable quantum corrections.

3. PepXPro: a framework for curating, generating, and optimizing structure-affinity protein-peptide datasets
   Provides a reproducible pipeline for building cleaner protein–peptide structure-affinity datasets from messy source collections. The main practical output is PepXPro Benchmark v1, a 70-complex nonredundant set that includes harder chemistries like cyclic peptides, making it useful for method comparison and dataset hygiene.

4. Benchmark Averages Hide the Failures That Matter: Quantizing ESM-2 for Protein Variant-Effect Prediction
   Looks at whether compressed ESM-2 models still work for protein variant-effect prediction. Average correlation changed very little after quantization, but some individual assays dropped sharply, so the takeaway is simple: smaller models may be fine for deployment, but assay-level checks matter more than headline averages.

5. Systematic Benchmarking of AI-Based Molecular Generation Models for Structure-Based Drug Design
   Benchmarks 12 structure-based molecular generation methods across 176 protein–ligand systems instead of relying on a few showcase targets. No single model wins across all metrics, but receptor-conditioned methods make better use of pocket geometry, which is a useful selection signal for teams choosing a generation stack.

6. PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening
   Introduces a test-time adaptation method for virtual screening that updates only a tiny fraction of model parameters for each pocket. It beats both pretrained and fully retrained baselines while changing about 0.03% of parameters, suggesting a practical route for pocket-specific tuning without full retraining costs.

7. FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
   Improves cyclic peptide generation by training on preference signals that account for both feasibility and hard-case robustness. On PepGLAD, CPSea LNR success rises from 46.89% to 57.79%, which is a meaningful gain if your main bottleneck is producing candidates that are not just novel but buildable.

8. LEN-Seek: Fast and scalable ligand binding-site similarity search in the latent space of an SE(3)-invariant graph VAE
   Builds a fast binding-site similarity search tool that maps pockets into a latent space rather than comparing structures directly each time. The reported per-comparison cost is about 3400-fold lower than ProBiS while still retrieving many similar sites, making it interesting for large template or target scouting workflows.

9. Stable Models, Unstable Candidates: Target Transferability in MOF Machine Learning for Gas Uptake Prediction
   Studies whether models trained for one metal-organic framework gas uptake target transfer well to another. Model rankings stay fairly stable across targets, but the actual top candidates often do not, which is a good reminder that transferable benchmarks do not automatically mean transferable discovery lists.

10. Domain-Adapted Molecular Language Models for Efficient Search of Make-on-Demand Libraries
    Shows that fine-tuning molecular language models on the chemistry of a specific make-on-demand library improves downstream search performance. The practical signal is better sample efficiency with adapted encoders, which could help teams screen large vendor spaces with fewer labels.

See you next week.

1. RAVEN: Frozen Random Graph Reservoirs with Physics-Informed Interaction Fingerprints for Protein-Ligand Binding Affinity Prediction builds a binding affinity model from 3D protein–ligand complexes using frozen atom-level graph features plus interaction fingerprints. It reports strong results on similarity-isolated PDBbind 2020R1 and CASF-2016, which is useful if you care about generalization beyond close training analogs.

2. Evaluating Lightweight and Full Fine-Tuning Strategies Against Classical Machine Learning for Protein Function Prediction compares frozen protein language model embeddings, full fine-tuning, and LoRA-style lightweight tuning against classical baselines. The practical signal is that the best tuning strategy depends on the task: LoRA beat full fine-tuning on VHH, while full fine-tuning led on fluorescence datasets.

3. Multi-Granular Rationale-Guided Molecular LLM for Property Prediction combines graph-derived rationales with a molecular language model to focus prediction on relevant substructures. It was the best overall generalist model across eight MoleculeNet tasks, suggesting a practical path for teams that want one model family across varied property endpoints.

4. A Modular Agentic Framework for Synthetically Constrained Multi-Objective Hit-to-Lead Optimization assembles an optimization workflow that links analog enumeration, property scoring, and Bayesian search under synthesis constraints. The main takeaway is efficiency: it enriches candidate sets while scoring only part of the full search space. As usual, the value will depend on the quality of the scoring models plugged into the loop.

5. ChemMoB: Contribution-Aware Dynamic Multi-Modal Balance for Chemical Representation Learning fuses 2D graphs, 3D conformations, SMILES, and spectra while rebalancing each view during training. It reports a 12.0% lower QM9 error than fixed fusion and solid BBBP and hERG results, making it relevant for multimodal property pipelines with uneven data quality.

6. NDST1 as a substrate-reduction target in Mucopolysaccharidosis type IIIC: virtual screening, microsecond molecular dynamics, and peptide design runs a large screening workflow for NDST1, starting from 4.1 million compounds and narrowing with docking, molecular dynamics, and free-energy estimates. It also adds diffusion-based peptide design, so the paper is most useful as an end-to-end example of mixing small-molecule and peptide follow-up around one target.

7. Tetra-Fuse: Late Fusion Preserves Task-Specific Strengths of Heterogeneous Molecular Encoders for ADMET Prediction combines four molecular encoders in a late-fusion setup rather than forcing one shared representation. On 21 Therapeutics Data Commons ADMET tasks, it outperformed three single-model baselines, which is a practical signal for teams deciding whether ensembles still beat picking one foundation model.

8. PARNET: A CLIP-SEQ-BASED FOUNDATION MODEL FOR RNA SEQUENCE REPRESENTATION LEARNING pretrains an RNA model on 223 eCLIP assays to learn sequence representations tied to protein binding. It outperformed RBPNet and matched or beat larger RNA and genomic language models on downstream tasks, which makes it a notable resource for protein–RNA interaction work.

9. A ligand-property-guided computational framework for prioritizing de novo protein binders for small molecules proposes a staged filter for de novo small-molecule binder design using ligand properties and pocket-level checks. The practical point is that local pocket validation removed many weak candidates and left manageable sets for capsaicin, limonene, and quercetin.

10. FuncSeek: Multi-PLM contrastive learning for protein functional similarity search combines several protein language models into a shared embedding space for nearest-neighbor function search. It reached 64.6% EC4 nearest-neighbor accuracy on CARE ood30, ahead of CLEAN and Diamond BLASTp, which is a strong signal for annotation and retrieval settings where homology alone is not enough.

That’s it for this week.

Subscribe now

1. PG-MLD: Physics-Guided Molecular Representation Learning via Dynamic 3D Trajectory Distillation
   Builds a SMILES-only molecular encoder by distilling information from dynamic 3D trajectories into simpler 1D models. The practical signal is straightforward: it improved MoleculeNet property prediction across three student architectures, suggesting a usable path to capture 3D signal without 3D inference.

2. Deep learning-guided identification of bacteriophage receptor-binding protein candidates for foodborne pathogen detection
   Uses an ESM-2 screening pipeline to find phage receptor-binding proteins, then checks candidates with structure prediction, structural search, and docking. It reports 653 high-confidence candidates, which could help teams building phage-based diagnostics or biosensors move from genomes to testable binders faster.

3. Integrated BODIPY photosensitizer discovery: Multimodal property prediction, fingerprint diffusion and human-in-the-loop molecular design
   Combines property prediction with molecule generation to support BODIPY photosensitizer design in a guided workflow. Its fingerprint diffusion model reached 96.88% validity versus 54.79% for the baseline, which matters if you care about generating chemically usable candidates instead of spending cycles on cleanup.

4. Learning to Generalize: Deep Models and Robust Benchmarks for Drug–Target Affinity Prediction
   Revisits drug–target affinity prediction with leakage-aware benchmarks and sequence-based models that take FASTA and SMILES as input. The main takeaway is benchmarking discipline: under a cleaner split, prior best Pearson drops from 0.74 to 0.67, and the new models improve across all reported metrics.

5. Evaluating Protein Language Model Embeddings for Structural Similarity in the Protein-Sequence Twilight Zone
   Tests whether protein language model embeddings can recover structural similarity when sequence identity is too low for standard alignment methods. All tested models beat BLASTP, with ProstT5 giving the strongest rank correlation, making this a useful result for fast remote-homology triage.

6. ActiveFusion: Fused representations improve active learning for molecular property prediction
   Builds active learning around fused molecular representations instead of relying on a single descriptor family. In regression tasks, the best fusion reached an average final R² of 0.71 versus 0.67 for Chemprop alone, a practical gain for teams choosing the next compounds to measure.

7. HDOCK-Multimer: integrating docking and combinatorial assembly for structure prediction of large protein complexes
   Targets large protein complex prediction by combining docking with combinatorial assembly rather than treating the full complex as one hard problem. It outperformed several established methods across three benchmarks, which makes it worth a look for multimer systems that are still difficult for end-to-end predictors.

8. Optimized Multiple Circular Sequence Alignment for Cyclic Peptide Motif Discovery
   Speeds up motif discovery for head-to-tail cyclic peptide libraries with an optimized circular sequence alignment pipeline. It recovered the same MDM2 motif while running more than 650× faster on 1,000 peptides, which is a strong workflow improvement for library-scale screening.

9. Bi-semantic Chemical Embedder for Joint Representation Learning of SMILES and Natural Language
   Learns a shared embedding space for SMILES and scientific text, aiming to connect molecular structure with written chemical meaning. The paper reports competitive performance on both molecular property and language tasks, though the main value will depend on whether your workflow actually needs cross-modal retrieval or search.

10. Boltz-Perturb: Improving Diversity and Accuracy in Protein-Ligand Co-Folding through Training-Free Conditioning Perturbation
    Adds inference-time perturbations to protein–ligand co-folding to explore more binding poses without retraining the model. The strongest variant improved top-20 oracle success by 2.6–7.8× while using over 75% less compute than a high-temperature baseline, making it a practical sampling trick for pose generation.

See you next week.

Subscribe now

1. When Do Models Win? A Learning Curve Benchmark for Molecular Property Prediction in Low-Data Regimes
   This paper benchmarks classical ML, graph networks, transformers, and 3D models across scaffold splits from 50 to 3000 samples. The main signal is practical: Gaussian process regression won on BACE in low data, while deep models pulled ahead on ESOL and Lipophilicity as data grew. A good reminder to choose model class by dataset size, not habit.

2. MATCHA: a toolkit for streamlining molecular property prediction for drug discovery applications
   MATCHA packages molecular property prediction into a more usable workflow with graph models, foundation models, uncertainty estimates, and simple explanations. It ranked 3rd out of 370 in the OpenADMET-ExpansionRx blind challenge, which makes it worth a look for teams that want a working ADMET stack rather than a new model idea.

3. In silico discovery of anticancer candidates targeting PTPN1 and HSP90AA1: molecular docking, ADMET profiling, and quantum chemical analysis
   This is a standard virtual screening workflow combining docking, ADMET filtering, CYP450 checks, and quantum calculations for anticancer candidates. CAT-3 emerged as the best-balanced compound on predicted affinity and drug-like properties. Useful as a compact example of how many groups are still chaining filters before synthesis, though it remains entirely in silico.

4. Structure-free, site-resolved contrastive learning extends small-molecule discovery beyond the reach of structure-based modeling
   Ptarmigan-1 predicts protein–ligand interactions from protein sequence and 2D ligand structure, without docking poses or solved protein structures. It reportedly matches or beats docking and co-folding methods on hard targets like cryptic or disordered sites, while scoring compounds in about 10 ms each. For proteome-scale screening, that speed and structure-free setup are the key signals.

5. Mavchen 1: A Conformational Ensemble Platform for Protein Ligand Pose Prediction That Substantially Outperforms Static Structure Prediction in a Category-Stratified Benchmark
   Mavchen 1 tests a simple but important idea: use receptor conformational ensembles instead of a single static structure for pose prediction. Across 29 targets, the ensemble approach beat the AlphaFold2 static baseline on 21, with mean RMSD improving from 5.60 Å to 3.39 Å. If your targets show induced fit or cryptic pockets, this is a strong argument for modeling flexibility earlier.

6. Histological triage of early-stage mycosis fungoides using a weakly supervised deep learning-based model: a multicentre, external validation, and clinical utility study
   This study uses weakly supervised learning on pathology whole-slide images to triage early-stage mycosis fungoides. The practical result is solid external validation, with AUROC 0.87 and very high sensitivity at the chosen threshold. It is not molecular ML, but it stands out as a careful example of multicentre validation and deployment-minded evaluation.

7. MS-GPT: Rethinking MS/MS De Novo Structure Elucidation as Spectrum-Induced Posterior Querying of a Molecule-Language Model
   MS-GPT reframes de novo structure elucidation from mass spectra as querying a molecule language model with spectrum-derived signals plus molecular formula. It reports state-of-the-art results on NPLIB1 and MassSpecGym, reaching Top-1 exact match rates of 29.8% and 23.9%. That is still far from solved, but it looks like a meaningful step for structure recovery from MS/MS.

8. TriShieldRAG: A Three-Ring Defense-in-Depth Framework Against Knowledge Corruption in Retrieval-Augmented Generation
   TriShieldRAG is not chemistry-specific, but it is relevant for AI-assisted screening and literature workflows that rely on retrieval. Its three-stage defense cuts attack success from about 91% to about 13% on a poisoned retrieval benchmark. For teams building internal scientific copilots, this is a useful reminder that retrieval quality and trust controls matter as much as the model.

9. MEGA-CL: A Molecular Foundation Model for Generalizable ADMET Prediction through Graph External Attention and Contrastive Learning
   MEGA-CL builds a graph-based foundation model for ADMET prediction using contrastive pretraining. The most readable result is that more than 75% of predictions fall within 3-fold error, suggesting useful calibration for practical property screens. Worth watching if you need one model family across several ADMET endpoints.

10. Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling
    Chamaileon focuses on binder design when the same sequence must work across multiple targets or protein states. It improves results on this harder cross-context setting, which is closer to many real engineering problems than single-target design. The caveat is that the main value depends on how often your program truly faces multi-state or multi-target constraints.

See you next week.

Subscribe now

Paper #10 is actually my work from ICML finally in preprint! I promise there was no bias in selection here. This is what fell out of my keywords, which makes sense, as I wrote the paper. It would be odd if it didn’t show up.

Otherwise, this week had some protein-ligand modeling, with a practical mix of uncertainty-aware scoring, faster screening, low-data benchmarking, and a few lightweight alternatives to larger foundation-style models.

1. Trustworthy Protein-Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion builds a binding affinity predictor that combines outputs from multiple docking engines while estimating how reliable each one is for a given case. The main practical signal is better-calibrated confidence and up to 25% lower error on high-confidence protein-ligand pairs. It looks most useful when ranking candidates and deciding which predictions are safe to trust.

2. When Do Models Win? A Learning Curve Benchmark for Molecular Property Prediction in Low-Data Regimes tests a broad model set across small molecular property datasets from 50 to 3000 samples. The useful takeaway is that the best model depends strongly on dataset and sample size: Gaussian process regression did best on BACE, while deep models led on ESOL and Lipophilicity. This is a good reminder to benchmark simple baselines before defaulting to larger architectures.

3. Discrete Ricci Curvature on Protein Contact Graphs for Lightweight Fold Classification uses compact graph-shape features from protein contact maps to classify folds without relying on large embeddings. It outperformed mean-pooled ESM-2 baselines on CATH and SCOPe, with the best setup reaching macro-F1 of 0.71 and 0.68. For teams with limited compute, this is a notable lightweight option.

4. CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards applies reinforcement learning to dock ligands into amyloid fibrils while rewarding cooperative stacking interactions. It reports better pose quality and stronger affinity correlation than baseline methods. The niche focus is important, but amyloid-targeted screening is exactly where standard docking often struggles.

5. ChemRAG-Mol: A Retrieval-Augmented Multimodal Framework for Molecular Property Prediction and Reaction Reasoning adds retrieval and multimodal reasoning to molecular prediction tasks. On QM9 HOMO prediction, it reduced mean absolute error by 13.9% versus the best baseline. The broader question is how well this kind of retrieval setup transfers beyond curated benchmark settings.

6. CHIMIYA-1: An Autoselection Foundation Model for ADMET Property Prediction, Rigorously Benchmarked Against the Therapeutics Data Commons ADMET Group presents an ADMET prediction model evaluated with overlap checks and repeated runs across the Therapeutics Data Commons benchmark suite. It ranked first on 4 of 22 endpoints and landed in the top decile on 20 of 22, which suggests solid all-around performance rather than dominance on a few tasks. The careful benchmarking is the main reason to pay attention here.

7. ProteinDock: A physics-informed layer to improve protein-protein docking reliability adds a physics-informed layer on top of protein docking workflows and can also rerank outputs from deep learning systems. On Docking Benchmark 5.5, it reached acceptable-or-better CAPRI quality on 80.2% of targets, a 32.8-point gain over Rosetta. That is a strong practical signal for groups working on docking reliability rather than new model classes.

8. Runtime-steered ultrashort molecular dynamics enables million-pose protein–ligand screening uses very short molecular dynamics runs plus runtime steering to evaluate huge numbers of protein-ligand poses more efficiently. It reached a pose-aware ROC-AUC of 0.868 for the top ten poses while cutting simulated time to 24% of exhaustive evaluation. For large screening campaigns, the compute savings may be the most important result.

9. DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models uses structure prediction models to guide iterative small-molecule design toward a target pocket. The reported gains are better predicted affinity under Boltz-2 and better pocket coverage on held-out AlphaFold-3 style evaluations. This is promising, though the main value will depend on how well these design gains hold up in downstream experimental filters.

10. Graph Learning on Ensembles of Cyclic Peptides: An Investigation of Molecular Ensemble Modeling models cyclic peptides as conformer ensembles instead of single static structures. The best hybrid model improved to R² = 0.538 versus 0.439 for a BERT baseline, suggesting real value from explicit ensemble handling. This seems especially relevant for flexible peptide modalities where one conformation rarely tells the full story.

See you next week.

Subscribe now

I missed last week due to the travel. I had hoped to sift through the papers and post a list of the bio-models from the conference. I may still do this at some point, but haven’t yet found the time.

Let’s jump into it. This week leaned toward protein–ligand work, with a mix of practical docking improvements, sober screening audits, and a few useful representation and design papers for peptides and immune receptors.

1. MInt-HDX: Leveraging Hydrogen-Deuterium Exchange Mass Spectrometry and Machine-Learning to Improve Protein-Ligand Docking.
   This paper combines hydrogen-deuterium exchange mass spectrometry with XGBoost to narrow docking regions and rerank poses. On three test systems, it recovered poses within 3 Å ligand RMSD, which makes it a practical option when structural data alone is not enough. Small validation set, so generality is still open.

2. A Standard Machine-Learning Virtual Screening Pipeline Produces Candidates Indistinguishable From Random Drug-Like Compounds: A Quantitative Audit Using COX-2 As Case Study
   This is a useful negative result: a standard random-forest virtual screening workflow for COX-2 did not beat random drug-like molecules after docking. The key signal is the null result itself, with no detectable difference from random in docking scores, which is a good reminder to benchmark against simple baselines before trusting hit lists.

3. PEPstrMOD2: Next-generation tertiary structure prediction of chemically modified and non-natural peptides
   PEPstrMOD2 predicts structures for modified and non-natural peptides by combining AlphaFold2 or ESMFold with molecular dynamics refinement. It reports 2.50 Å mean RMSD on ModPep16 and 4.49 Å on ModPep433, outperforming AlphaFold3 on that larger modified-peptide set. Useful if your pipeline goes beyond standard amino acids.

4. A control-validated pan-proteome deep-learning pipeline nominates GPR35 as a candidate target of the orphan bacterial metabolite ligiamycin A
   This work uses a proteome-wide drug-target interaction model, then corrects for model bias with controls and follows up with docking. After that filtering, GPR35 emerged as the top target candidate, giving a concrete template for target deconvolution on hard natural products. The practical lesson is that control-based correction seems as important as the model itself.

5. MolMAE: A Surface-Centric Multimodal Masked Autoencoder for Molecular Representation Learning
   MolMAE learns molecular embeddings from surface, graph, and fragment views rather than relying on one representation alone. It shows strong ESOL scaffold-split performance and stays competitive across property tasks, which suggests surface information can add useful signal for general-purpose molecular features. Worth watching for downstream QSAR and multimodal pretraining.

6. IgGM2: An All-Atom Foundation Model for Adaptive Immune Receptor Design
   IgGM2 is an all-atom model that jointly designs immune receptor sequence and full 3D structure. It outperformed AlphaFold3 on FoldBench spatial modeling and improved Rosetta interface preference, making it one of the more practical design papers this week for antibody, nanobody, and T-cell receptor work.

7. Deep-AGAPE: An Open-source Web-Based Machine Learning Tool for Predicting G-Quadruplex Ligand Stabilization
   Deep-AGAPE predicts whether small molecules stabilize G-quadruplexes using standard molecular descriptors with XGBoost and neural networks. The best model reached an F1 of 0.848 on an independent test set, and the web tool makes it easy to try on screening ideas without rebuilding the stack. Descriptor-based models can still be very serviceable in focused settings.

8. HEDGEHOG: Hierarchical Evaluation of Drug Generators Through Rigorous Filtration
   HEDGEHOG is an evaluation pipeline for generated molecules that applies staged filters from basic chemistry checks through docking and synthesis-related criteria. Only 0.65% of 230k generated molecules passed all stages, which is a strong practical reminder that raw generator output can look much better than truly usable candidates. Good benchmark framing for teams comparing generative models.

9. Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization
   This paper generates ligands directly from pocket context, then pushes candidates toward better developability with a property-aware optimization step. It reports an average binding score of -8.85 kcal/mol with ADMET gains up to 73%, which is the kind of combined potency-plus-properties target many teams actually care about. As usual, external validation matters more than docking alone.

10. MolecuAR: A Multilingual WebAR Molecular Visualization Tool Connected to PubChem for Chemical Education
    MolecuAR is a browser-based augmented reality viewer tied to PubChem with multilingual name lookup. This is more tool than model, but it looks useful for teaching, demos, and quick mobile visualization without install friction. Less relevant for core molecular ML, though handy for communication and onboarding.

That’s it for this week—see you next issue.

Subscribe now

I tuned my pipeline because it was chem heavy with minimal bio models, but I may have gone too far. We’ll run a few weeks with the current keywords, and then evaluate.

Anyway, preprints this week were geared toward protein language models, with a few useful papers on screening workflows, benchmark quality, and low-data prediction. The theme is not bigger models, but better problem setup: structure-aware conditioning, careful splits, and lightweight pipelines that are easier to use in real projects.

1. StructureSAFE: A structure-aware chemical language model for unified hit identification and lead optimization builds a chemical language model that conditions generation on protein structure and evolutionary context for both hit finding and lead optimization. It reports state-of-the-art results on MolGenBench, with notably better chemical plausibility, which makes it relevant for teams trying to reduce invalid or unrealistic proposals.

2. PolyDiT: Masked Diffusion Transformers for Polymer Generation and Property Prediction tests one diffusion-style model across several polymer representations, including SMILES, SELFIES, and graphs, for both generation and property prediction. The practical signal is that representation choice matters: SELFIES maxed out validity, while graph and SMILES variants gave the best held-out prediction across 8 targets.

3. EpiESM-GA: Resource-Efficient Protein Foundation Model Features for Equitable B-Cell Epitope Prediction uses frozen ESM-2 embeddings plus feature selection, then trains small downstream models for linear B-cell epitope prediction. It reaches an AUC-ROC of 0.880 on the IEDB benchmark, suggesting a solid low-compute recipe for sequence-only immunology tasks.

4. Peptide:MHC Binding Stability Prediction Using Protein Language Models predicts peptide–major histocompatibility complex stability by pretraining on affinity data and then fine-tuning on half-life measurements. It outperforms ESM-2 and earlier baselines, and the careful split design makes the result more useful for realistic antigen presentation work.

5. Prosculpt: Lowering the Barrier to Computational Protein Design packages common protein design steps into one configurable pipeline, combining backbone generation, sequence design, structure prediction, and scoring tools. The main value is workflow simplification rather than a new model, which could help smaller teams run modern design stacks with less engineering overhead.

6. RAGMODEX: Retrieval-Augmented Generation for Molecular Design and Explainable AI - Interpretable QSAR and Guided Molecular Design of GLUT-1 Inhibitors adds retrieval-based explanations and applicability tracking to a QSAR workflow for GLUT-1 inhibitors. The headline ROC-AUC of 0.97 looks strong, but the drop to 0.71 under leave-one-cluster-out testing is the more useful signal: generalization depends heavily on chemistry novelty.

7. AI-guided discovery for low-resource peptide engineering using evolutionary scale modeling shows a low-data peptide engineering setup using ESM-2 embeddings with Gaussian processes and tree models. A useful takeaway is that cross-validation performance from roughly 50 labels predicted downstream active-learning success, which may help teams decide early whether a campaign is worth scaling.

8. EnzyKAN: Protein Language Model Embeddings and Kolmogorov–Arnold Network Variants for Enzyme Commission Classification with a Proposed Electron-Transfer Physics Feature Framework evaluates ESM-2 embeddings with simple classifiers for enzyme commission prediction. The strongest result came from a standard multilayer perceptron, with macro F1 of 0.750, which is a reminder that simple heads can still be the best default on protein embedding tasks.

9. Identifying and Addressing Systematic Data Leakage in Protein-Ligand Affinity Benchmarks examines leakage in affinity benchmarks and proposes novelty-tiered splits based on ligand similarity. The key signal is the collapse of a ligand-only baseline from r=0.66 to r=0.14 on the hardest tier, making this a useful paper for anyone evaluating protein–ligand models and worried about inflated benchmark results.

10. Transformer-Based Active Learning for Data-Efficient Vaccine Epitope Selection in PRRS applies a small transformer in an active-learning loop for epitope selection with limited labels. It reaches 86.8% accuracy with just 60 samples, which makes it a practical example of how modest sequence models can support low-data vaccine design workflows.

That’s the week in molecular ML.

Subscribe now

If you haven’t seen biohub’s new agentic exploration of ESM Atlas, please go play with the agent. It’s quite incredible. You can find it here.

With that, this week we have new papers on protein–ligand modeling, with a useful mix of new datasets, structure-aware generators, compact molecular encoders, a pipeline validation, and workflow tools that make cheminformatics and simulation systems easier to run.

1. DockBind: a quality-controlled dataset of modelled protein–ligand complexes associated with affinity data builds a curated set of docked protein–ligand complexes paired with affinity labels using free docking, template docking, and pose filtering. The practical value is scale with structure checks: 18,332 complexes across 612 targets, which could make it useful for training or benchmarking binding models. As with any modeled complex set, downstream results will still depend on how well docking assumptions match the target class.

2. Agentic, Explainable, and Generative Artificial Intelligence for Structure-Aware Small-Molecule Drug Design combines activity prediction, interaction-aware features, and molecule generation in one drug design workflow. It reports ROC AUC 0.868 for activity classification and generates 479 unique molecules with high novelty, making it a practical template for teams that want ranking plus design in the same loop. The tradeoff is that this is a broad workflow paper rather than a narrowly validated discovery study.

3. ModernMolBERT: A Compact SELFIES Encoder for Molecular Representation Learning adapts a compact encoder model to SELFIES strings for molecular representation learning. On 25 binary classification tasks, it reaches 77.9 mean ROC-AUC, about 5 points above comparable string encoders, which is a useful signal for teams choosing a lightweight pretrained molecular backbone. The main question is how well that advantage carries to regression and low-data settings.

4. OracleScreen-LILRB4: Machine Learning-Guided Discovery of Myeloid Immune Checkpoint Binders Validated in Patient-Derived Cells uses an ensemble regression model to rank binders for the immune checkpoint target LILRB4. In prospective testing, the top 200 predictions gave a 28.5% hit rate and 15-fold enrichment, which is the kind of result applied screening teams look for. The patient-derived cell validation also makes this more than a pure virtual screening exercise.

5. MoMA-Chem: Parameter-Efficient Multimodal Transfer Learning for Molecular Property Prediction fuses SMILES, 2D molecular images, and infrared spectra while keeping the base encoders frozen. With less than 1% of parameters tuned, it outperforms several standard baselines and full multimodal fine-tuning on five benchmarks, which is a strong practical signal for resource-constrained teams. It looks especially relevant where multiple assay or characterization modalities are already available.

6. GROMACS-Agent: Text-to-Workflow Generation for Molecular Dynamics A multi-agent framework for topology-aware GROMACS workflow generation, evaluation, correction, and selected safe execution turns text prompts into GROMACS workflows and adds evaluation, correction, and guarded execution steps. In a small five-task pilot, evaluator scores improved from 6.6/10 to 10/10, suggesting a usable path toward automating routine simulation setup. The evidence is still early, but the safety and correction layer is the main practical idea to watch.

7. MultiMolecule: a modular ecosystem for biomolecular sequence-model workflows presents a software ecosystem for running RNA, DNA, and protein sequence-model workflows in a more standardized way. It packages 53 model families, 112 checkpoints, and 10 prediction pipelines, which could reduce setup friction for labs comparing sequence models across tasks. This is more infrastructure than method paper, but that may be exactly the point for production users.

8. Schema-Constrained LLM Planning for Executable Molecular Workflows: An Intent-to-Execution Infrastructure for Cheminformatics uses schema constraints and an RDKit execution layer to turn natural-language requests into runnable molecular workflows. The key result is a jump in executable workflow rate from 1.8% to 69.1%, which matters if you want language models to produce chemistry steps that actually run. It is a good reminder that structured output design can matter more than larger models alone.

9. Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design builds a generative model that conditions molecule design on inputs such as protein sequence or binding pocket information. It reports 100% validity in generation and competitive target-conditioned results, making it relevant for groups exploring unified design interfaces instead of separate task-specific models. The main open question is how much it improves real hit quality beyond standard validity and benchmark scores.

10. Residue-Level Attributions in Protein Language Models Do Not Recover Allergen Epitopes tests whether residue-level explanations from protein language models line up with known allergen epitopes. The answer is mostly no: protein-level predictions stay useful, but attribution maps are near-random against epitope labels. For anyone using saliency as a biological explanation layer, this is a useful caution.

That’s it for this week—see you next issue.

1. TRPM8 Virtual Screening: The Essential Role of Target-Specific, Inactive-Enriched Machine-Learning Scoring Functions builds a virtual screening benchmark for TRPM8 and tests target-specific scoring models trained with extra inactive compounds. The practical signal is straightforward: inactive-enriched training reduced false positives and improved enrichment on more dissimilar test sets, which matters if you care about triaging fewer dead-end hits.

2. Retrieval-Augmented Generation with Parallel Graph Reasoning and Chemical Constraint Refinement for Molecular Property Prediction and Reaction Analysis combines retrieved chemical knowledge with graph-based reasoning for property prediction and reaction analysis. It reports a ROC-AUC of 0.893 on MoleculeNet, suggesting retrieval may be useful when standard predictors miss context from related molecules and reactions.

3. SupraBench: A Benchmark for Supramolecular Chemistry introduces a benchmark for host–guest and related supramolecular tasks, plus a domain text corpus for adaptation. The useful takeaway is that current language models still have clear headroom here, and domain adaptation helps some regression-style tasks but can hurt strict formatted outputs.

4. GLACIER: A Multimodal Student-Teacher Foundation Model for Molecular Property Prediction combines graphs, SMILES, and standard descriptors in a student–teacher setup for property prediction. It reports stronger prediction with better efficiency, making it relevant for teams that want multimodal gains without paying the full cost at inference time.

5. Molecular Representation Learning by Integrating SMILES and Three-Dimensional Structures learns molecular representations by aligning SMILES tokens with 3D geometry during pretraining. It improved ESOL and FreeSolv and also showed strong atom-level 13C NMR prediction, which is a practical sign that the representation transfers across both whole-molecule and local tasks.

6. MDForge: Agentic Molecular Dynamics Pipeline Design under Sparse Simulator Feedback uses an LLM-based agent to write and refine molecular dynamics pipelines under sparse simulator feedback. It was competitive on three SAMPL benchmarks and also surfaced a picomolar CB[7] binder confirmed by NMR, though this is still a pipeline-design story more than a broadly validated discovery engine.

7. Network Pharmacology and Molecular Docking Reveal Convergent PI3K/AKT and MMP9 Mechanisms of Ginkgo biloba Bioactive Compounds in Aging and Cancer combines target-network analysis with docking to map possible mechanisms for Ginkgo compounds. The main value is hypothesis generation around PI3K/AKT and MMP9, but the evidence is still docking-led and should be read as prioritization rather than validation.

8. De novo molecular generation with optical property preconditioning at the token level conditions an autoregressive molecular generator on optical-property tokens for OLED-oriented design. It shows directional control over generated properties, but the reported calibration errors suggest it is better suited for guided exploration than precise property targeting.

9. Integrated Single-Cell Transcriptomics, Cell-Cell Communication Inference, and Network Pharmacology Identify CXCR4 as a Therapeutic Target in Idiopathic Pulmonary Fibrosis links single-cell analysis, communication inference, and docking to prioritize CXCR4-related compounds in pulmonary fibrosis. Myricetin scored better than the reference ligand in docking, but this remains an early-stage ranking result rather than strong binding evidence.

That’s the week in BioML, though it feels more like ChemML today.

Subscribe now

1. Virtual Library Generation, Drug-Likeness Filtering, and Docking of Taloside-Based Galectin-3 Inhibitors builds an open-source screening workflow for taloside regioisomers, combining library generation, simple drug-likeness filters, and AutoDock Vina docking against Galectin-3. In a small practical test, 14 of 16 products passed filters and docked, and the top redocked pose was close to the reference with a 1.39 Å centroid distance. This looks most useful as a reproducible template rather than a full validation study.

2. An accurate nucleic acid-small molecule docking framework via geometric deep learning with large-scale pretraining targets RNA and DNA docking with a geometric deep learning model pretrained on synthetic complexes and refined on co-crystal data. On 125 complexes, it reached 56% top-1 success at 2.0 Å RMSD versus 29% for rDock, which is a meaningful jump for nucleic acid screening. The main caveat is that performance still depends on curated structural data.

3. No Pose Left Behind: Integrating Activity and Structural Data with Uncertainty-Aware Multi-objective Learning for Kinase Inhibitor Prediction combines kinase activity prediction with pose quality and uncertainty in one structure-aware model. The practical value is that it reduces activity error versus a baseline that ignores pose quality, while also flagging more reliable complexes with lower uncertainty. This could be useful for teams mixing assay and docking data of uneven quality.

4. A Replicate-Aware Biophysical Triage Workflow for De-Risking Protein-Ligand Hypotheses Before Experimental Validation uses replicate-aware molecular dynamics to check whether proposed protein–ligand poses stay stable before moving to experiments. In the examples here, X77 and wild-type BCL2 complexes were retained while lopinavir was not, showing the workflow can separate stable from weak or inconsistent hypotheses. This is more of a decision filter than a generator of new candidates.

5. Molecular Structure and Toxicity Predictor (MSTP) V2: A Mobile-Developed Cheminformatics Web Application for 2D/3D Visualization and Machine Learning-Based Toxicity Prediction packages a standard toxicity model into a lightweight web app using Morgan fingerprints and a Random Forest trained on Tox21. The reported test accuracy is 73.23%, which makes it a reasonable baseline tool for quick screening and visualization. The method is straightforward, so the main appeal is accessibility rather than model novelty.

6. Hantavirus ANDV Glycoprotein Gc: Comprehensive Analysis and Molecular Docking of 35 FDA-Approved Antiviral Ligands screens 35 approved antivirals against the ANDV glycoprotein using docking and contact-based scoring. Remdesivir ranked highest at -9.1 kcal/mol, supported by strong contact and disruption scores, which gives a short list for follow-up rather than a direct efficacy claim. As with most docking-only studies, experimental confirmation will matter.

7. ChemDistill: Hierarchical Graph Attention Distillation with Parameter-Efficient Multi-Task Adapters for Molecular Property Prediction and Chemical Reaction Analysis aims to keep molecular property models small while still handling multiple chemistry tasks. The paper reports a Tox21 ROC-AUC of 0.879, beating ChemBERTa with a 1.6M-parameter model, which is a useful signal for teams that care about compact deployment. It looks especially relevant where compute or fine-tuning budget is limited.

8. Physics-Based vs AI-Based Free Energy Prediction for Protein-Ligand Potency: Public Benchmarks and Internal Project Evidence compares a physics-based relative binding free energy workflow with an AI potency predictor on public and internal projects. The physics-based setup reached 1.38 kcal/mol RMSE on a public benchmark and preserved ranking separation better on prospective sets, while the AI baseline struggled more. For applied teams, the message is that careful free energy workflows still appear stronger for potency ranking in hard, real project settings.

9. ChemCorr: A Unified Framework for Molecular Property Prediction and Chemical Knowledge Graph Reasoning via Correlation-Aware Pre-training tries to connect molecular pretraining with downstream tasks such as reaction prediction, retrieval, and chemical knowledge graph question answering. The practical result is broad gains across tasks, including a 3.3% lift in reaction accuracy and 3.8% in Hits@1 for knowledge graph QA over strong baselines. This is promising for groups that want one pretrained backbone across several chemistry workflows.

Thanks for reading — see you next week.

Subscribe now

1. PEARL: Parameter-Efficient Adaptation with Retrieval-Augmented Learning for Molecular Property Prediction
   PEARL tests lightweight fine-tuning across ChemBERTa, MolFormer, and Uni-Mol, then adds retrieval from ZINC-250k to enrich molecular features. The practical signal is that the 3D Uni-Mol setup reached MCC 0.623 on BACE, an 8% absolute gain over the best SMILES-based model.

2. AnnotateMissense: a genome-wide annotation and benchmarking framework for missense pathogenicity prediction
   This paper builds a missense variant predictor from a broad annotation table that combines ClinVar, dbNSFP, AlphaMissense, and protein language model features. It reports very strong benchmark performance, with mean MCC 0.941, though the main evidence is cross-validation rather than a clearly separated real-world deployment test.

3. TIGER: Text-Informed Generalized Enzyme-Reaction Retrieval
   TIGER links enzyme sequences to reaction descriptions by aligning protein and text representations in a shared retrieval space. It outperforms prior bidirectional retrieval baselines across evaluation splits, which makes it a useful signal for enzyme search and reaction annotation workflows.

4. What drives performance in molecular MPNNs? An operator-level factorial benchmark
   This benchmark breaks molecular message passing networks into a small set of operator choices and tests 84 combinations. The clearest result is practical: fusion choice and seed initialization matter, while more complex update functions do not show a significant benefit.

5. PROTOCOL: Late Interaction Retrieval for Protein Homolog Search
   PROTOCOL applies late-interaction retrieval to residue-level protein embeddings instead of compressing each protein into one vector. It beats alignment-based, pooled embedding, and single-vector baselines on SCOPe and Pfam tasks, suggesting a stronger option for remote homology search.

6. DriftingMol: Decoder-Coupled Drift for One-Pass Property-Conditional Molecular Generation
   DriftingMol generates molecules in one pass by steering a latent SELFIES representation toward target properties. On ZINC250K it reaches 94.7% uniqueness and shows a useful property-control signal, though the reported correlation leaves room for stronger optimization accuracy.

7. MolDiff-PEFT: Conditional Molecular Graph Diffusion with Parameter-Efficient Fine-Tuning for Molecular Generation and Property Prediction
   This work combines conditional graph diffusion with parameter-efficient fine-tuning so one model can support both generation and property prediction. The practical tradeoff looks solid: 96% validity and 93% novelty on MOSES, with AUROC 0.872 while using 62.4% less training.

8. Uncertainty-Aware Protein-Ligand Binding Mode Prediction with Counterfactual Pocket Perturbation
   This paper ranks docked poses with a geometry-aware model and estimates confidence by masking pocket features to create counterfactual perturbations. The useful result is selective reliability: keeping the lowest-uncertainty half of predictions raises top-1 success to 88.79%, with failure detection AUROC 0.895.

9. Machine Learning-Guided Virtual Screening Identifies Candidate COX-2 Inhibitors: A Computational Prioritization Study
   This is a straightforward screening pipeline: a random forest on Morgan fingerprints trained on ChEMBL COX-2 data, followed by standard filtering on a large ZINC set. It reports held-out Pearson R of 0.817 and narrows 249,455 compounds to 127 candidates, but the output is still computational prioritization rather than experimental validation.

10. Sequence Scanning Improves Accuracy for Antibody-Antigen and Molecular Glue Ternary Complex Co-Folding
    This study adds a sequence-scanning constraint workflow to co-folding systems for antibody-antigen complexes and molecular glue ternary complexes. The main takeaway is better contact recovery and pose accuracy on larger benchmarks, which could make these co-folding pipelines more useful in early complex triage.

That’s the short list for this week.

Subscribe now

1. SCOPE: A Chemically-Aware Workflow Automation Software for Molecules and Molecular Crystals builds a Python workflow layer for molecular and crystal computations with structured state tracking and high-performance cluster execution. The practical value is reproducible end-to-end automation, including spectra and thermodynamics analysis, which should help teams scale routine computational chemistry jobs with less custom glue code.

2. Retrieval-Augmented Cross-Modal Molecular Reasoning Network for Property Prediction and Reaction Analysis combines retrieved examples with molecular and reaction representations to handle both property prediction and reaction analysis in one framework. It reports strong benchmark results, including 0.912 ROC-AUC on BBBP and 0.935 top-1 on USPTO-50k, useful for teams working across multiple chemistry tasks. As usual, broad benchmark performance still needs testing in production settings.

3. MSAlign: Aligning Molecule and Mass Spectra Foundation Models for Metabolite Identification links molecular and mass spectra models so spectra can retrieve likely metabolite structures more accurately. It consistently beats prior retrieval methods across benchmarks, which makes it relevant for metabolomics groups trying to improve candidate ranking without rebuilding foundation models from scratch.

4. Hierarchical Contrastive Learning for Multi-Domain Protein-Ligand Binding targets binding prediction for multi-domain proteins, where standard single-view approaches often struggle. The paper shows better binding inference and uncertainty estimates on PDBBind, suggesting a practical path for more reliable modeling on harder protein targets. The main caveat is that gains are reported on benchmark complexes rather than prospective campaigns.

5. EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation adapts a protein language model with structural information to design antibody complementarity-determining regions. It reports a 16% gain in sequence recovery and a 43% drop in perplexity versus the best graph baseline. Great for antibody teams balancing sequence plausibility with structural context.

6. Machine Learning-Guided Virtual Screening Identifies Candidate COX-2 Inhibitors: A Computational Prioritization Study uses a straightforward random forest on Morgan fingerprints to rank likely COX-2 inhibitors. The model reaches a held-out Pearson correlation of 0.817, showing that simple ligand-based screening can still be competitive for focused target programs. This looks most useful as a prioritization filter rather than a full discovery engine.

7. QuantumPioneer: Scalable generation of quantum chemical data for solution-phase hydrogen transfer reactions presents a high-throughput quantum chemistry pipeline for building reaction data in solution phase. Its benchmark errors are 1.45 kcal/mol for hydrogen atom transfer barriers and 0.57 kcal/mol for solvation free energies, which makes it a practical dataset and workflow contribution for reaction modeling teams.

8. Fine-tuning Pocket-Aware Diffusion Models via Denoising Policy Optimization fine-tunes a pocket-conditioned diffusion generator with reinforcement learning framed around the denoising process. It reports a best Vina score of -8.5 kcal/mol on CrossDocked2020, suggesting a practical route to push structure-based molecule generation toward better docking outcomes. Docking-based gains should still be treated as an early screen rather than proof of downstream activity.

9. Manifold-Guided Attention Steering steers generation at inference time by nudging attention toward directions associated with more correct SMILES trajectories. It improves over unsteered and static steering baselines, which may interest teams looking for lightweight ways to improve molecular language model output quality without retraining.

10. Generating Physically Consistent Molecules with Energy-Based Models uses an energy-based model to generate 3D molecules while enforcing more physically consistent structures. It reports state-of-the-art results on QM9 and GEOM-Drugs.

See you next week.

Subscribe now

1. VibeProteinBench: An Evaluation Benchmark for Language-interfaced Vibe Protein Design
   This paper builds a three-stage benchmark for language-guided protein design, covering recognition, engineering, and generation tasks. The practical takeaway is that current models do not perform well across all three stages, which makes this a useful stress test for teams building general-purpose protein design agents.

2. CatIF-RL: Activity-Oriented Enzyme Sequence Design by Steered Inverse Protein Folding
   This method designs enzyme sequences from structure while steering toward higher predicted catalytic activity. On its benchmark, it reports about a 4-fold gain in predicted kcat, suggesting a practical route for prioritizing activity-focused variants before wet-lab testing.

3. SCOPE: Siamese Contrastive Operon Pair Embeddings for Functional Sequence Representation and Classification
   This work uses paired operon embeddings to classify functional relationships between sequences. It reaches ROC-AUC 0.71, but average precision did not improve over a simple cosine baseline.

4. PDBMINER: RANK AGGREGATION-BASED SELECTION OF QUALIFIED PROTEIN DATA BANK STRUCTURES
   This paper automates the tedious step of choosing the best Protein Data Bank structures linked to a target protein. For docking and structure-based workflows, the value is speed and consistency: it turns a manual review that can take days into a ranked shortlist produced in minutes.

5. Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
   This method adds a group-level diversity signal during post-training so generative models do not collapse onto a few high-scoring designs. The useful result is a better utility-diversity tradeoff across both protein and small-molecule design settings, which matters when you need varied candidates rather than near-duplicates.

6. MetaGEM: Bottom-Up Reconstruction of Genome-Scale Metabolic Networks via Deep Enzyme-Metabolite Anchoring
   MetaGEM predicts enzyme-metabolite links by combining protein representations with 3D molecular features. It reports AUROC 0.97 on a de-homologized benchmark, making it a practical candidate for faster metabolic network reconstruction when annotation is sparse.

7. QT-Net: Rethinking Evaluation of AI Models in Atomic Chemical Space
   This paper revisits how models are evaluated in atomic chemical space and proposes a stronger held-out test setup. The main value is methodological: it shows that more careful out-of-distribution evaluation can change which models look reliable for downstream molecular property prediction.

8. Machine-learning–enabled virtual screening delivers scaffold-diverse, experimentally validated, c-MET inhibitors
   This study combines a support vector machine with docking and molecular dynamics to triage c-MET inhibitor candidates. The practical signal is experimental validation: two hits were confirmed, with the best reaching 1.2 μM IC50, and the series appears scaffold-diverse rather than narrowly optimized.

9. Learning the Interaction Prior for Protein-Protein Interaction Prediction: A Model-Agnostic Approach
   This work adds an interaction prior on top of existing protein-protein interaction predictors instead of replacing them. That plug-in design is appealing for applied teams, and the paper reports consistent gains over strong baseline predictors, though the benefit will depend on the base model and graph quality.

10. Gradient-Free Noise Optimization for Reward Alignment in Generative Models
    This method improves generative outputs by optimizing noise at inference time without backpropagation. For protein structure generation, that offers a lightweight control mechanism when training-time reward alignment is expensive or unavailable.

See you next week.

1. ProtDBench: A Unified Benchmark of Protein Binder Design and Evaluation
   This paper builds a common test bed for protein binder design, with shared tasks, verifier models, and throughput-aware metrics. The practical takeaway is that the same binder method can look strong or weak depending on the verifier and filtering setup, which makes this useful for anyone comparing design stacks. A good reminder that benchmark design is shaping reported success.

2. Benchmarking open-source tools for in silico antiviral drug discovery
   This study compares 15 open-source protein-ligand affinity tools for antiviral discovery and also fine-tunes DrugFormDTA on a curated viral binding set. The clearest signal is that task-specific tuning helped materially, improving correlation from 0.5 to 0.7, while Boltz-2 and GNINA ranked best overall. Useful if you need a practical open-source starting point for viral targets.

3. AgenticPosesRanker: An Agentic AI Framework for Physically Grounded Ranking of Protein-Ligand Docking Poses
   This paper uses a language-model agent to combine six physics-based checks and rank docking poses. It reached 50.0% best-pose accuracy, roughly matching Smina and beating random selection, which suggests agent-style orchestration can be competitive without replacing standard tools outright. The caveat is that the gain looks modest rather than decisive.

4. Benchmarking Single-Pose Docking, Consensus Rescoring, and Supervised ML on the LIT-PCBA Library: A Critical Evaluation of DiffDock, AutoDock-GPU, GNINA, and DiffDock-NMDN
   This benchmark compares docking, rescoring, and supervised re-ranking for virtual screening on LIT-PCBA. The strongest practical result is that simple docking plus rescoring was solid, but supervised ML re-ranking improved enrichment further, reaching median EF1% of 4.49. For applied screening teams, this supports treating ML rerankers as an add-on stage rather than a full replacement.

5. A Closed-Form Persistence-Landmark Pipeline for Certified Point-Cloud and Graph Classification
   This method classifies graphs and point clouds using topological summaries without learned parameters. It performed strongly against other topology-based baselines, which makes it interesting as a lightweight, interpretable option for structural data workflows. It is less a molecular foundation model paper than a useful pipeline component.

6. MP2D: Constrained Monte Carlo Tree-Guided Diffusion for Multi-Objective Protein Sequence Design
   This paper combines diffusion with tree search to design protein sequences under several competing goals at once. The main value is balanced improvement across 4 to 5 properties, which is closer to how real design campaigns are run than single-score optimization. Worth watching for teams juggling potency, stability, and developability in one loop.

7. Protein-Conditioned Multi-Objective Reinforcement Learning for Full-Length mRNA Design
   This work generates full-length mRNA sequences conditioned on the target protein and optimizes for multiple objectives at once. It improved the in silico Pareto frontier for half-life and translation efficiency versus supervised baselines, pointing to a more practical design setup than one-metric codon optimization. The main caveat is that the reported gains are computational, not experimental.

8. Tree-Conditioned Edit Flows for Ancestral Sequence Reconstruction
   This paper reconstructs ancestral protein sequences using descendant sequences plus tree information, while handling insertions and deletions directly. It performed best on natural homolog sets with indels, which matters because many real evolutionary problems are not substitution-only. On cleaner classical benchmarks, though, standard methods still held an edge.

9. PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction
   This benchmark standardizes peptide MS/MS prediction with tougher data splits and shared evaluation across six models. The practical message is that earlier results may have been helped by leakage and narrow experimental settings, and this benchmark exposes those weaknesses more clearly. Useful for proteomics teams choosing models that need to generalize across labs or species.

10. Edge-specific signal propagation on mature chromophore-region 3D mechanism graphs for fluorescent protein quantum-yield prediction
    This paper predicts fluorescent protein quantum yield from chromophore-centered 3D graphs and edge-specific signals. On 531 proteins, it reported R = 0.772 with strong top-hit precision, making it a practical signal for protein engineering workflows that need rank-ordering more than mechanistic detail. As usual, random cross-validation is encouraging but not the hardest generalization test.

See you next week.

Subscribe now

1. Advancing Ligand-based Virtual Screening and Molecular Generation with Pretrained Molecular Embedding Distance
   Builds a molecular similarity score from pretrained molecular embeddings, without task-specific retraining. The practical signal is strong: the same distance works for ligand-based virtual screening and as a reward for guided molecule generation. This looks useful if you want a simple plug-in scoring function across multiple discovery steps.

2. Computational Design and Experimental Validation of Photoactive PARP1 Inhibitors
   Combines docking, machine-learned force fields, quantum chemistry, and free-energy calculations to design light-responsive PARP1 inhibitors. The key result is experimental: one compound showed a 15-fold gain in PARP1 inhibition under green light. It is a narrow case study, but a good example of a full computational-to-experimental loop.

3. h-MINT: Modeling Pocket-Ligand Binding with Hierarchical Molecular Interaction Network
   Models pocket–ligand binding with both atom-level and fragment-level interactions instead of using a single flat representation. It reports modest but consistent gains of about 2–4% on PDBBind and LBA ranking tasks. For teams working on affinity prediction, the main takeaway is that hierarchical interaction structure may still offer measurable headroom.

4. HBGSA: Hydrogen Bond Graph with Self-Attention for Drug-Target Binding Affinity Prediction
   Focuses binding affinity prediction on hydrogen-bond interaction graphs, then uses attention to weight the most informative contacts. It outperformed baseline models on PDBbind Core and CSAR-HiQ, suggesting that explicit interaction features can still beat more generic molecular graphs in some settings. The tradeoff is likely more dependence on good structural inputs.

5. Simple Self-Conditioning Adaptation for Masked Diffusion Models
   Adds a lightweight self-conditioning step to masked diffusion models by reusing earlier clean-state predictions during generation. The benefit is practical rather than flashy: improved small-molecule generation without redesigning the whole model. This is the kind of post-training tweak that could be easy to test in existing diffusion pipelines.

6. Do Larger Models Really Win in Drug Discovery? A Benchmark Assessment of Model Scaling in AI-Driven Molecular Property and Activity Prediction
   Benchmarks random forests, graph neural networks, and pretrained sequence models across 22 property and activity tasks with scaffold splits. The main result is a useful reality check: classical ML won 10 tasks, GNNs won 9, and pretrained sequence models won 3. If you are choosing a baseline stack, this argues for starting simple before paying for scale.

7. COMO: Closed-Loop Optical Molecule Recognition with Minimum Risk Training
   Improves optical chemical structure recognition by training the model against the quality of its own sampled outputs, rather than only next-token error. It beat both rule-based and learned baselines on 10 benchmarks, including lower-data settings. For chemistry data pipelines, this could matter anywhere structure images still need to be converted into machine-readable molecules.

8. CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation
   Generates dual-target molecules with a tree search over synthesizable fragments while balancing multiple objectives at once. It reports better docking scores with improved diversity, which is a practical combination for early hit generation. As usual for docking-led generation, the value will depend on how well the scoring setup matches real binding.

9. DPRM: A Plug-in Doob h transform-induced Token-Ordering Module for Diffusion Language Models
   Introduces a plug-in token ordering module for diffusion language models so the model reveals easier or more useful tokens first. The paper reports better performance than confidence-based baselines, especially on harder subsets. It is more general than chemistry, but the main applied idea is clear: decoding order can be another lever for molecular sequence generation.

See you next week.

This week there are a few new infrastructure tools like ProDock and MDAgent that automate complex docking and simulation pipelines to improve reproducibility. On the generative side, models are becoming more deliberate, with RosettaSearch applying inference-time steering for protein design and FRIGID generating molecules directly from mass spectra. There is also evidence that fundamentals remain highly competitive, from traditional graph descriptors outperforming heavier neural networks to structured prompts improving model accuracy.

1. ProDock: From multi-target consensus docking into database-backed storage
   ProDock builds a Python workflow for receptor and ligand preparation, batch docking, and structured result storage in SQLite. The practical value is less about a new scoring model and more about making multi-target docking campaigns reproducible, queryable, and easier to audit. Best fit for teams running many docking jobs and struggling with provenance.

2. AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling
   AROMA combines text, graph structure, and protein sequence to model how genetic perturbations affect a virtual cell. It outperformed baselines and stayed robust in zero-shot and sparse-data settings, which matters if your perturbation data are incomplete. The main appeal is broader context integration rather than a single new data type.

3. FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time
   FRIGID generates candidate molecules directly from mass spectra using a diffusion language model with formula and fingerprint guidance. It reports over 18% Top-1 accuracy on MassSpecGym and about 3× higher Top-1 than leading methods on NPLIB1. This is a useful signal for de novo structure elucidation, especially when retrieval methods fall short.

4. MDAgent: A Multi-Agent Framework for End-to-End Molecular Dynamics Research
   MDAgent organizes molecular dynamics work into multiple agents that handle steps across the simulation workflow and reuse prior cases through a memory system. The reported result is more stable end-to-end performance with better adaptability across tasks. The paper is most relevant for teams trying to reduce manual orchestration in MD pipelines.

5. Structure-guided molecular design with contrastive 3D protein-ligand learning
   This paper learns joint 3D representations of protein pockets and ligands, then uses them to generate molecules conditioned on structure. The main practical signal is competitive zero-shot virtual screening, suggesting the learned representations transfer beyond the training set. As usual, screening performance is encouraging, but downstream synthesis and assay results will matter most.

6. Multi-Objective Reinforcement Learning for Generating Covalent Inhibitor Candidates
   This work uses a SMILES-based recurrent model with reinforcement learning to generate covalent inhibitor candidates while balancing multiple design goals. It can rediscover known inhibitors, reaching 0.74% in 10,000 runs, which is modest but shows the search can recover relevant chemistry. Useful if you care about multi-constraint design more than novelty alone.

7. How Creative Are Large Language Models in Generating Molecules?
   This paper tests how well large language models generate molecules under chemistry-related constraints and frames the question as molecular “creativity.” A notable result is that constraint satisfaction improved as more constraints were added, which is a practical reminder that prompt structure can help rather than hurt. More of an evaluation paper than a new generation method.

8. An Integrated Deep-Learning Framework for Peptide-Protein Interaction Prediction and Target-Conditioned Peptide Generation with ConGA-PePPI and TC-PepGen
   This framework pairs peptide-protein interaction prediction with target-conditioned peptide generation in one pipeline. On the prediction side, ConGA-PePPI reached 0.921 AUROC, and on the generation side 40.39% of peptides scored higher than native templates by AlphaFold 3 ipTM. That makes it one of the more directly actionable peptide design papers this week, though experimental follow-up remains the key test.

9. Graph-Theoretic Models for the Prediction of Molecular Measurements
   This paper revisits molecular property prediction with graph-theoretic descriptors plus simple regression models instead of heavier neural architectures. Average R² improved from 0.24 to 0.79 and matched or beat graph convolutional networks on the tested benchmarks. The practical message is straightforward: strong handcrafted features can still be very competitive.

10. RosettaSearch: Multi-Objective Inference-Time Search for Protein Sequence Design
    RosettaSearch applies inference-time search to protein sequence design, using structure-based rewards to steer candidate selection after generation. It reports a 2.5× gain in design success rate over LigandMPNN and 18%–68% better structural fidelity. This is a strong signal for teams who already have sequence generators and want better candidates without retraining.

See you next week.

Subscribe now

Key protein updates include a glycosylation-aware design filter, a standardized peptide benchmark, and multi-encoder architectures for predicting antibody developability. On the small-molecule side, new methods focus on condition-aware generation and scaffold preservation, rounded out by broader advances in single-cell perturbation modeling and structural imaging.

1. Integrating glycosylation in de novo protein design with ReGlyco Binder Design Filter
   This paper adds a glycosylation-aware filter to de novo binder design by checking whether predicted glycans would block binding. The main practical value is simple: it catches designs that look fine in silico but are likely to fail once glycans are present, flagging 11% of non-binders in about 3 hours on a dual-core CPU. A useful downstream filter for teams already running protein binder generation.

2. PepBenchmark: A Standardized Benchmark for Peptide Machine Learning
   PepBenchmark packages 35 peptide datasets into a common preprocessing and evaluation setup across several model families. For researchers, this is mostly about cleaner comparisons and fewer hidden data-processing differences when testing peptide models. It looks especially useful for teams working across canonical and non-canonical peptides.

3. PUFFIN: Protein Unit Discovery with Functional Supervision
   PUFFIN tries to learn reusable multi-residue protein units from structure graphs, guided by functional labels. The learned units line up with known InterPro annotations, which suggests they may be useful building blocks for protein representation or generation pipelines. The caveat is that this is more of a representation paper than a direct design result.

4. A robust workflow for 3D imaging of human mitochondria using cryo-electron tomography
   This work presents an end-to-end workflow for molecular-resolution 3D imaging of isolated human mitochondria using cryo-electron tomography. While not a molecular ML paper in the narrow sense, it matters because better structural imaging workflows can feed cleaner training and validation data into downstream models. Useful for groups linking imaging, segmentation, and structure-aware learning.

5. CAGenMol: Condition-Aware Diffusion Language Model for Goal-Directed Molecular Generation
   CAGenMol builds a small-molecule generator that handles multiple design constraints at once, such as binding compatibility and drug-like properties. The practical result is better success on dual-conditioned generation tasks, with gains in affinity and drug-likeness compared with baselines. Worth watching for teams that need controllable generation rather than open-ended molecule sampling.

6. Biologically-Grounded Multi-Encoder Architectures as Developability Oracles for Antibody Design
   This paper builds antibody developability predictors by combining frozen protein language model encoders with assay-specific prediction heads. On a 242-IgG benchmark, it reports 12–20% improvement on 3 of 5 assays versus baselines, which makes it a practical screening tool for antibody design funnels. Best fit seems to be ranking or triage before wet-lab testing.

7. PRiMeFlow: Capturing Complex Expression Heterogeneity in Perturbation Response Modelling
   PRiMeFlow models how single-cell gene expression distributions shift after genetic or small-molecule perturbations. The key result is the best distribution fit on PerturBench, which matters if your use case depends on capturing heterogeneous cell states rather than average response. More relevant to perturbation modeling teams than to generation-focused readers.

8. Counterfactual Peptide Editing for Causal TCR--pMHC Binding Inference
   This paper uses counterfactual peptide edits to reduce shortcut learning in TCR–pMHC binding prediction. It reports family-held-out AUROC of 0.831 and a 39.7% drop in shortcut index, suggesting better out-of-distribution behavior. That makes it interesting for immunology models where spurious sequence cues often break generalization.

9. Scaffold-Conditioned Preference Triplets for Controllable Molecular Optimization with Large Language Models
   This work tunes a molecular language model to edit compounds while preserving a desired scaffold. The practical signal is improved optimization success with better scaffold similarity than baseline methods, which is often the real constraint in medicinal chemistry workflows. A useful direction for teams optimizing leads without drifting too far from known chemotypes.

10. Learning Parameterized Nonlinear Elasticity on Curved Surfaces
    This paper uses a physics-informed neural network to model nonlinear elasticity on curved surfaces. It matches exact or numerical solutions and extrapolates to unseen parameters, which could be relevant for groups modeling shells, membranes, or protein-like elastic surfaces. Interesting method paper, but less directly tied to current molecular ML production work.

See you next week.

1. General Multimodal Protein Design Enables DNA-Encoding of Chemistry
   This paper builds a diffusion model that designs protein sequence and structure together around reactive chemical intermediates. The practical result is strong: designed heme enzymes carried out new-to-nature carbene-transfer reactions with activity above engineered enzyme baselines. The main caveat is that this is a specialized enzyme design setting, not a general protein benchmark.

2. LeukemIA: A Machine Learning Pipeline for Patient-Specific Molecular Generation in Acute Myeloid Leukemia via Latent Diffusion and Multi-Target Docking
   This work builds a patient-specific small-molecule generation pipeline for acute myeloid leukemia using latent diffusion followed by multi-target docking. The useful signal is the workflow design: it narrows generated compounds by activity across several disease-relevant targets rather than a single screen. As presented, the evidence is prioritization-focused, so experimental follow-up will matter.

3. Making Room for AI: Multi-GPU Molecular Dynamics with Deep Potentials in GROMACS
   This paper integrates deep-potential inference into GROMACS for multi-GPU molecular dynamics on protein systems. It scales to 32 GPUs, but efficiency drops to about 40%, with most runtime spent in model inference, which is a useful reality check for teams planning production deployment. This is more of an infrastructure paper than a modeling advance.

4. Predicting Activity Cliffs for Autonomous Medicinal Chemistry
   This study predicts where small chemical changes cause large potency shifts, using matched molecular pairs plus simple molecular and 3D context features. For medicinal chemistry workflows, the signal is strong ranking performance: NDCG@3 reaches 0.910 versus 0.839 for random, helping surface high-risk or high-reward edits earlier. It looks especially relevant for closed-loop design systems.

5. Reinforcement Learning with LLM-Guided Action Spaces for Synthesizable Lead Optimization
   This paper uses reinforcement learning with an LLM-guided action space, but keeps molecule edits grounded in validated reaction templates. The practical benefit is better top-ranked molecules under synthesizability constraints, with a 10.4% gain over the strongest baseline on the reported top-10 score. That makes it more useful than unconstrained generation for teams that need compounds they can actually make.

6. Structure-Based and Stability-Validated Prioritization of BACE1 Inhibitors Integrating Meta-Ensemble QSAR and Molecular Dynamics
   This work combines QSAR, docking, and molecular dynamics to prioritize BACE1 inhibitors. The main practical signal is solid triage performance, with ROC-AUC of 0.920 and a final shortlist checked for binding stability. It is a classic screening pipeline rather than a new model class, but that may be exactly what some discovery teams want.

See you next week.

1. Q-BIOLAT: Binary Latent Protein Fitness Landscapes for QUBO-Based Optimization
   This paper builds a binary latent space for protein fitness search from protein language model embeddings, then uses combinatorial optimization methods to explore it. The practical takeaway is that simple PCA-derived binary codes worked better than collapsed autoencoder latents for search, which may matter if you want a compact surrogate before running expensive wet-lab rounds.

2. Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
   This work builds a fully atomistic protein binder generator and combines pretraining with extra optimization at inference time. It reports higher in silico success rates than prior generative and hallucination-based baselines, making it a useful signal for teams designing binders directly against structure.

3. A Generative Neuro-Symbolic AI for Protein Sequence Design
   This paper combines a learned protein fitness landscape with explicit rule-based reasoning to design sequences under constraints. The notable result is a de novo nanobody with nanomolar affinity against an immune-evasive SARS-CoV-2 variant, though the validation here is still focused on a specific case.

4. ImmSET: Sequence-Based Predictor of TCR-pMHC Specificity at Scale
   ImmSET predicts T cell receptor and peptide-MHC binding specificity using sequence inputs rather than full structure pipelines. It outperformed ESM2 and AlphaFold-based workflows under stricter evaluation settings, which is a practical sign that sequence-only models may be enough for large-scale screening.

5. Efficient and Principled Scientific Discovery through Bayesian Optimization: A Tutorial
   This is a tutorial paper on how to use surrogate models and acquisition rules to choose the next experiment efficiently. There is no new benchmark result, but it is a clear refresher for teams building active learning loops in molecule or protein discovery.

6. Context-Guided Equivariant Diffusion Model for Efficient and Physics-Aware 3D Molecular Generation and Editing
   This work generates and edits 3D molecules with an equivariant diffusion model that uses context to keep structures physically consistent. The main practical message is improved efficiency while preserving 3D realism, though the summary provides less benchmark detail than some of the other generation papers this week.

7. Hierarchy-Guided Topology Latent Flow for Molecular Graph Generation
   This paper separates high-level topology planning from 3D coordinate generation for molecular graphs. It reports 94.0% PoseBusters validity and strong QM9 stability, which makes it a useful result for researchers who care about valid 3D outputs rather than just graph novelty.

8. Property-Guided Molecular Generation and Optimization via Latent Flows
   MoltenFlow organizes latent space by molecular properties, then uses guidance to generate or optimize candidates. The practical signal is better unconditional generation and stronger multi-objective optimization under fixed oracle budgets, which is relevant if assay calls are your main bottleneck.

9. Data Sieving for Scalable Real-Time Multichannel Nanopore Sensing
   This paper is not a generative model, but it tackles a common experimental bottleneck by detecting useful nanopore events in real time on GPU. It cuts stored data volume by up to 98% while keeping translocation signatures, which could make closed-loop sensing workflows much easier to run.

10. Latent-Y: A Lab-Validated Autonomous Agent for De Novo Drug Design
    This work packages target understanding, epitope selection, design, and lab follow-up into an autonomous antibody design agent. The strongest signal is real experimental output: binders for 6 of 9 targets, with single-digit nanomolar affinity reported for successful cases.

See you next week.

Subscribe now

This week centers on generation and representation: two papers on protein-related modeling, two on small-molecule generation, one on protein–ligand affinity prediction, and one on RNA structure under force. The clearest practical signals are better binding-affinity benchmarks, more efficient molecular generation, and useful gains from combining sequence representations at different biological scales.

1. Persistent local Laplacian prediction of protein-ligand binding affinities
   This paper builds protein–ligand binding affinity predictors from persistent local Laplacian descriptors and standard machine learning models. The practical signal is simple: it outperformed prior methods on three benchmarks, which makes it worth a look for teams doing structure-based scoring. Caveat: the gain depends on descriptor quality and benchmark fit.

2. Computational modeling of RNA-protein binding interactions under an external force
   This work extends ViennaRNA to model RNA secondary structure while also accounting for protein binding and external force. It shows that predicted force–extension behavior changes at biologically relevant protein concentrations, which could be useful for interpreting single-molecule experiments. This is more of a mechanistic modeling tool than a general ML benchmark paper.

3. SIGMA: Structure-Invariant Generative Molecular Alignment for Chemical Language Models via Autoregressive Contrastive Learning
   SIGMA trains a chemical language model to treat different tokenizations of the same molecule more consistently during generation. The main practical result is better sample efficiency and stronger structural diversity on multi-parameter optimization tasks, which matters if you are using autoregressive models for design loops. The value here is cleaner search, not just more samples.

4. Permutation-Symmetrized Diffusion for Unconditional Molecular Generation
   This paper tackles a common 3D generation problem: the same molecule can be represented by many atom orderings. By explicitly symmetrizing over permutations, it reaches competitive QM9 generation quality with better efficiency. For practitioners, that suggests a cleaner path to unconditional 3D molecular generation without wasting model capacity on arbitrary atom indexing.

5. End-to-End Training for Unified Tokenization and Latent Denoising
   This paper studies joint training of tokenization and latent denoising in one end-to-end setup rather than separate stages. It reports strong image-generation results, but the practical relevance here for molecular ML is mainly methodological: unified latent training may simplify future protein or molecule generators. Treat this as a transferable modeling idea rather than a domain result.

6. Cross-Granularity Representations for Biological Sequences: Insights from ESM and BiGCARP
   This study compares and combines sequence representations from models that operate at different biological scales. The useful takeaway is that mixing granularities improves intermediate-level prediction tasks, and deeper-layer embeddings appear most faithful. For applied teams, it is a reminder to test representation fusion before building a new model from scratch.

See you next week.