Journal of Cheminformatics

Drug-target affinity (DTA) prediction is a key task in drug discovery, enabling the estimation of the interaction strength between candidate compounds and biological targets. However, current models rely on connectivity-based molecular representations and do not explicitly account for the spatial organization, also known as stereochemistry. This limitation becomes evident when considering chirality, where a drug can exist as enantiomers, i.e., molecules that share the same atoms and bonds but differ in their three-dimensional arrangement. Despite their chemical similarity, they can interact differently with the same target, leading to variations in binding affinity and biological activity. In this paper, we propose a stereochemistry-aware DTA prediction framework that incorporates this information into molecular representations. Drug representations are learned from chemical structure using a directed-bond message passing graph neural network that captures enantiomers configurations, while protein targets are represented through sequence-based embeddings. Experiments on the Davis dataset demonstrate that our model can improve affinity prediction. Importantly, a case study on a manually curated dataset of enantiomers with different biological action shows that the model is able to distinguish the affinities in the two forms consistent with their experimentally observed biological activity. These findings support the relevance of stereochemistry-aware molecular representation for more accurate and chemically faithful DTA prediction.

Protein structure prediction models (such as AlphaFold, Chai, Boltz) have transformed structural biology and are increasingly explored for drug discovery; however, their utility for large-scale screening of antibody-antigen (AB-AG) interactions remains unclear, particularly for distinguishing true binding from non-binding pairs at scale. To our knowledge, there has not been an exhaustive exploration of Boltz-2 inference settings on this high impact problem, and in this paper we set out to describe and implement a novel benchmarking framework that can accelerate progress in the field. We evaluated Boltz-2 (NVIDIA NIM implementation) on 519 therapeutic monoclonal antibodies from Thera-SAbDab, pairing each antibody with its cognate target and a randomly assigned non-cognate antigen. We developed a novel evaluation framework that systematically captures variability across stochastic seeds while benchmarking different inference settings, including datasets with and without crystallographically resolved antibody structures. Across settings, Boltz-2-derived confidence metrics showed weak, though above-chance, discrimination (0.5 < ROC-AUC < 0.60). Among evaluated metrics, the minimum value of the interface predicted TM-score (ipTM-min) across seed-samples, captured the strongest signal. Interestingly, additional feature aggregation and multivariate modelling provided little to no improvement. Increasing the number of stochastic predictions yielded front-loaded gains, with diminishing returns beyond [~]15-20 seed-samples, suggesting limited value of extensive sampling in practical workflows. Notably, inference without multiple sequence alignments (MSAs) slightly improved performance on non-crystallized antibodies ({Delta}AUROC {approx} +0.027) while reducing runtime by [~]8 seconds per prediction compared to shallow MSA settings. Overall, these results indicate that off-the-shelf confidence metrics from general-purpose structure prediction models may be insufficient for reliable target-antibody screening and highlight the need for task-specific optimization, while confirming that modest amounts of sampling can be helpful, but not in itself sufficient to improve performance significantly as gains plateau relatively quickly.

BackgroundThe function of a protein is intrinsically linked to its three-dimensional fold, and deep learning has revolutionized the field by enabling high-accuracy structure prediction at an unprecedented scale. Nevertheless, the growing deployment of these predictive pipelines in drug discovery and structural biology reveals a critical bottleneck that lies in the lack of independent and rigorous estimation of model accuracy (EMA) methodologies. ResultsHere we present DeepUMQA-Global, a single-model deep learning framework for estimating accuracy of protein structure models. Our method employs a structure-sequence cross-consistency mechanism to evaluate the bidirectional compatibility between the predicted structure and the input sequence, enabling comprehensive characterization of fold accuracy. DeepUMQA-Global outperforms the self-assessment confidence scores of AlphaFold3, achieving improvements of 57.8% in Pearson correlation and 49.0% in Spearman correlation. With respect to the CASP16 retrospective benchmark, DeepUMQA-Global outperforms all single-model accuracy estimation methods that participated in CASP16 and achieves performance comparable to that of the top consensusObased methods. A lightweight consensus strategy built upon DeepUMQA-Global ranks first among all CASP16 participants, surpassing all other methods, including consensus approaches, and highlighting the strength of our method. Remarkably, DeepUMQA-Global demonstrates a strong ability to discriminate between alternative conformational states of proteins, as evidenced in the CASP unique alternative conformation protein complex target and the CoDNaS benchmark. ConclusionsOur results indicate that DeepUMQA-Global can be extended to broader protein modeling tasks, moving beyond static evaluation to offer a foundation for dynamic conformation EMA, where it accurately discriminates alternative conformational states and demonstrates reliable predictive fidelity in model accuracy estimation.

We screened a 5 receptor x 7 aptamer = 35-cell cross-target matrix with HADDOCK3 [1] under blind ambiguous-interaction-restraint (AIR) protocols on AlphaFold-modelled receptors. The screen surfaced 12 operationally distinct failure modes (collapsing to [~]8 conceptual classes; [&#167;]3.1). The K_D-calibration subset is n = 4 cells with literature K_D records under matched assay conditions; the broader cohort includes [&ge;] 6 biological cognate or intended-cognate cells. The principal case study is P01031 (complement C5, 1676 aa, [&ge;] 12 structural domains): all 7 panel members produced positive HADDOCK3 top-1 scores under a scale-adaptive AIR. Score-term decomposition locates the anomaly in the AIR term (+217 to +268 to top-1 score). With AIR zeroed, scores fall to -131 to -74 -- the small-receptor regime. Boltz-2 cofolding chain-pair ipTM (cpi_AB) is an independent channel: P01031 shows the lowest median cpi_AB (0.211; 0/7 above the 0.5 confident-interface threshold). To our knowledge, this is the first reported case study of a 1676 aa multi-domain receptor exhibiting this signature under blind scale-adaptive AIR -- an n = 1 mechanistic case, not a statistical generalisation. We adapt the QSAR applicability domain concept [14-16] to in silico aptamer screening. [&#167;]3.7 reports an empirical Mode 1 mitigation (pLDDT-aware AIR prefilter; cohort Jaccard recovery [~]10x).

Recent deep learning models for L1000 chemical perturbation prediction incorporate dedicated drug molecular encoders. We retrained seven such models from scratch with zeroed or shuffled drug inputs, and compared them with a multilayer perceptron that uses only cell-line basal expression. Under drug-blind evaluation, ablation caused negligible performance changes and the drug-free baseline matched all models. Current architectures do not yet utilise drug molecular features for generalisation to unseen compounds.

Machine learning continues to accelerate peptide and protein design through the rapid prediction and generation of sequences with desired characteristics. Many applications focus on predicting properties, functions, and structures, as well as generating point mutations and de novo designs. Nevertheless, many models prove less generalizable than initially claimed. Most predictors and generators are trained on sequential datasets, where imbalances can be addressed during preprocessing. In contrast, structural bias, a subtype of algorithmic bias arising from uneven representation of structural classes in training datasets, and the limitations of early protein structure predictors have frequently remained undetected and uncorrected. The recent surge in powerful protein structure prediction tools, such as the AlphaFold and RosettaFold series and their variants, now presents opportunities to mitigate this issue. We hypothesize that such structural sampling biases influence the downstream performance of ML models. Using antimicrobial peptides as a case study, we audited the structural biases in 16 state-of-the-art predictors for antimicrobial activity and tested whether structural information constrains their predictions. Our analysis revealed that models explicitly trained on sequential data still produce predictions biased by uneven fold representations and data leakage. These findings highlight the importance of integrating balanced structural data or implementing bias-mitigating strategies to develop agnostic models that maximize bioactive protein discovery and multi-objective optimization.

Ab initio prediction of side effect frequencies is important for assessing the risk-benefit profile of drugs and for identifying potential adverse effects early in development. A key challenge is chirality: many drugs exist as enantiomers, pairs of molecules with the same atoms and bond connectivity but different three-dimensional arrangements. Although chemically similar, enantiomers can interact differently with biological targets and therefore exhibit distinct efficacy and adverse-effect profiles. Here we introduce F2S (Features to Signatures), a method to predict the frequencies of drug side effects while explicitly accounting for chirality. Drug representations are learned directly from chemical structure using a directed-bond message-passing graph neural network that captures stereochemical configurations. Side effect representations are derived from curated textual descriptions encoded with a frozen PubMedBERT model. Side effect frequencies are predicted from the dot product between drug and side effect signatures together with biases for drugs and side effects. We evaluated F2S extensively across multiple settings, including cold-start and warm-start prediction, prospective evaluation, and scenarios controlling for chemical similarity between training and test drugs. Across these evaluations, F2S achieves performance comparable to state-of-the-art methods for general side-effect frequency prediction while producing fewer false positives and substantially improves the prediction of frequency differences between enantiomer pairs. Finally, F2S learns compact 10-dimensional signatures that support interpretability: drug signatures reflect therapeutic class and shared targets, side-effect signatures capture phenotype similarity, and the learned bias terms correlate with the popularity of drugs and side effects.

Despite the breakneck pace of progress in protein design methodology, frontier problems remain challenging, with leading methods struggling to design high-affinity binders, scaffold multiple functional motifs, or stabilize large multi-domain proteins. Recent research efforts have focused on two areas: improving model reasoning when generating active sites or binding interfaces, and improving concordance between the design process and the in silico oracle used to select promising designs. In addressing the first, the field has shifted towards all-atom models that capture sidechain conformations in atomistic detail by eschewing data-efficient SE(3)-equivariance, mirroring the evolution of AlphaFold2 to AlphaFold3. In addressing the second, recent work has focused on replacing generative models employing diffusion or flow-matching with hallucination approaches that directly optimize the oracle in sequence space; this improves success rates but reduces computational efficiency. Here, we close and surpass the generation-hallucination gap by revisiting SE(3)-equivariance using a branched polymer treatment of protein structures. The resulting diffusion model, Genie 3, achieves state-of-the-art performance on binder design, motif scaffolding, and unconditional generation, while being significantly faster than the best existing methods. We use Genie 3 to design a nanomolar binder of Nipah Glycoprotein G, a tetramer with minimal structural or biophysical characterization, as part of the Adaptyv Bio Nipah Competition, achieving a 12.5% success rate. Taken together, our results present a new frontier in protein design capability and a reexamination of the role of SE(3)-equivariance in molecular modeling.

Molecular similarity searching is a workhorse of cheminformatics, but the dominant Tanimoto/topological-fingerprint paradigm has well-known blind spots. It is highly sensitive to molecular size, suffers from steep activity cliffs, and frequently fails to retrieve scaffold-hopping bioisosteres. A complementary descriptor that has received comparatively little attention is global elemental composition. Despite the conceptual simplicity of comparing molecules by their elemental ratios, no widely deployed method exists for the statistically rigorous identification of "chemical twins" defined by stoichiometric proximity. We address this gap with TwinSAR (Stoichiometric Analysis and Retrieval), an adaptive kernel-based algorithm that combines three methodological innovations: (i) binary fingerprint blocking that partitions molecule by element-presence patterns and bounds the cost of all-pairs comparison from O(NM) to O({sum}nimi) enabling million/billion-scale searches; (ii) a per-block adaptive radial basis function (RBF) kernel whose precision parameter is calibrated independently for each fingerprint block via the median heuristic, providing fair similarity comparison across chemical sub-spaces of vastly different density; and (iii) a logit-transformed Z-score filter that maps bounded RBF scores onto an unbounded scale, allowing high-similarity pairs to be prioritized relative to the empirical score distribution of their own fingerprint block. TwinSAR is offered in two operating modes: (i) a deterministic BULK mode for exact reproducibility; and (ii) a stochastic FAST mode that achieved a 3.29x wall-clock speed-up in the present benchmark while preserving the similar unique-query and unique-target coverage. Statistical validation showed that detected twin pairs are 12.7x more similar in absolute ratio space than block-matched random pairs (p < 0.001), while a column-permutation negative control returned a median of zero spurious twins across three independent permutations. A controlled benchmark further established that an 8-element representation (single-element heavy-atom ratios) is sensitivity-equivalent to a comprehensive 254-element representation while running 3.55x faster. As a case study, TwinSAR was deployed in an end-to-end virtual screening pipeline against the BCL-2 target protein, where it reduced a 327,071-compound commercial library to a 390 focused candidate panel. The chemical interpretability of the retrieved twins is illustrated by their structural diversity around conserved heavy-atom skeletons. TwinSAR therefore provides a fast, conformation-free, and statistically principled prefilter that is fully orthogonal to topological fingerprints.

In any field, unquestioningly accepting artificial intelligence (AI) results should be considered bad practise. Here, we devised a comparative modelling-based strategy for validating protein structures that exploits the well-known observation that protein folds are far more conserved than protein sequences. We identify proteins with a similar fold to the AlphaFold-generated query protein and determine their structural alignment to the query. The hypothesis is that if the sequence alignment coincides with the structural alignment, then the structure is validated. The strategy is implemented on a helix-by-helix and strand-by-strand basis using a multi-template pairwise local profile alignment method that works well into the twilight zone. The method is illustrated by application to the transmembrane transporter PEPT1, for which the structure is known, and the S-deacylases ABHD13 and ABHD16A, for which only AI-generated models exist. ABHD16A is particularly challenging because a sequence alignment search with BLASTp does not reveal any structural homologues and therefore requires work with extremely remote homologues; however, both models are validated through this strategy and are stable during classical molecular dynamics simulations. The ability of the strategy to identify errors is assessed with reference to misaligned ABHD13 models and misfolded decoy proteins.

Understanding protein proximities in the theoretical ligand space is essential for developing therapeutics with desirable polypharmacology, predicting off-targets, and discovering surrogate ligands for poorly characterized proteins. This is especially important for G protein-coupled receptors (GPCRs) - a major class of drug targets, many of which still lack known ligands. Circumventing this limitation, we present GPCR-CoINPocket v2, a contact-informed metric for detecting GPCR pharmacological similarities from amino-acid sequences alone. We first establish a "gold standard" of pharmacological relatedness using ChEMBL-derived ligand sets. We then replace traditional evolutionary amino acid similarity matrices with a chemically-informed matrix derived from protein:ligand interaction patterns across 3,306 structures, significantly improving early detection of shared pharmacology between distantly homologous receptors. An additional unconstrained, contact-informed matrix further enhances predictive performance. Pilot application of the method revealed previously unrecognized similarities between the {beta}2 adrenoceptor and three Class A peptide GPCRs, which we confirmed experimentally by demonstrating the binding of select ligands of these receptors to the {beta}2. Dimensionality reduction of similarity scores recapitulates known receptor relationships and predicts neighbors of orphan GPCRs later confirmed experimentally. Overall, GPCR-CoINPocket v2 provides a powerful sequence-based framework to prioritize ligand space, predict polypharmacology, and accelerate GPCR drug discovery and deorphanization.

Abstract SummaryReliable structure-based prediction of small-molecule druggability is hindered by a fundamental labeling problem. Experimentally confirmed liganded sites (positives) are observable, but credible "undruggable" pockets (negatives) are almost impossible to define. Standard supervised machine learning consequently relies on arbitrary definitions of undruggable, leading to bias and false negatives. Here we introduce PocketBagger, a positive-unlabeled (PU) learning framework for pocket druggability prediction trained exclusively on experimentally determined Protein Data Bank1 (PDB) structures. PocketBagger uses PU bagging to learn key features associated with reliable druggable pockets and considers all remaining pockets in the structurally characterized proteome as unlabeled. We demonstrate the capability of PocketBagger through the training of a simple Random Forest classifier and demonstrate its power in recall (0.804), even when challenged with increasingly difficult generalizability assessments and entire protein-family hold outs. We benchmark and demonstrate the added value of PU learning by comparing PocketBagger to a leading deep-learning predictor. However, PocketBagger is intended to be used as a framework for any model architecture. Along with the code, the data generated by PocketBagger are deployed in canSAR.ai, providing scalable, generalizable pocket druggability predictions to the drug discovery community.

Structure Integration with Function, Taxonomy and Sequences (SIFTS) provides residue-level mappings between UniProt Knowledgebase sequences and Protein Data Bank structures and has historically been generated through internal Protein Data Bank in Europe (PDBe) pipelines. Here, PDBe-SIFTS is presented as a fully open-source, locally deployable implementation of this mapping framework. The pipeline combines fast, scalable sequence search using MMseqs2, an improved bounded scoring scheme for ranking candidate mappings, and residue-level mapping refinement based on backbone connectivity. PDBe-SIFTS is distributed as a Python package with command-line tools for 1) building a sequence search database, 2) identifying the best sequence-structure match, 3) one-to-one mapping at the residue level, and 4) generating SIFTS annotations in PDBx/mmCIF format. Benchmarking on the complete Protein Data Bank archive showed that MMseqs2 reduced archive-scale UniProtKB searches from hours with BLASTP to minutes, approximately 22-36 times faster, while curated mappings were recovered at top rank in 93.1% of cases. The remaining discrepancies mainly involved biologically ambiguous cases such as highly conserved proteins, chimeric constructs, or closely related orthologs. These results show that PDBe-SIFTS enables fast mapping, improving structural coherence in residue-level alignments while delivering the most up-to-date and accurate mappings, comparable to expert curation. Tool: https://github.com/PDBeurope/SIFTS Quick start notebook with example: https://github.com/PDBeurope/SIFTS/tree/master/notebooks Broader audience statementMatching protein sequences to their three-dimensional structures, and mapping annotations across both, is essential for understanding protein function, interactions, and molecular mechanisms. This integrated view enables richer interpretation of biological data and underpins advances in drug discovery, disease research, and protein engineering. PDBe-SIFTS provides an open and functional framework for structure-sequence mapping, allowing researchers and databases to run, inspect, and extend these mappings locally, while benefiting from faster searches, transparent scoring, and structurally informed residue-level alignments. Graphical abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=110 SRC="FIGDIR/small/721839v1_ufig1.gif" ALT="Figure 1"> View larger version (25K): org.highwire.dtl.DTLVardef@5e6ea6org.highwire.dtl.DTLVardef@1b2754dorg.highwire.dtl.DTLVardef@1334f9forg.highwire.dtl.DTLVardef@1b083a1_HPS_FORMAT_FIGEXP M_FIG C_FIG

Molecular representation determines which aspects of chemical structure can be learned, compared, and interpreted in computational drug discovery. Existing encodings typically emphasize either compact string description, as in SMILES and SELFIES, or efficient similarity search, as in circular fingerprints, but they may not simultaneously provide deterministic sequence structure, graph-level interpretability, pharmacophore annotation, and high-fidelity molecular reconstruction. Here, we introduce MolCodon, a codon-based molecular language that represents small molecules as deterministic sequences of fixed-width three-character tokens over a five-symbol alphabet, C, N, O, S, and X. Inspired by the triplet organization of the genetic code, MolCodon assigns chemically defined codon families to atoms, bonds, ring and branch topology, fused-ring references, pharmacophore features, bond mobility, charge, and stereochemistry. A deterministic graph traversal with ring-contiguity preservation produces sequences in which chemically meaningful substructures remain locally organized and traceable to the underlying molecular graph. Across around 2,9 million molecules from six commercial screening libraries, MolCodon achieved 98.93% InChIKey-level round-trip fidelity, supporting its use as a high-fidelity sequence representation for drug-like chemistry. MolCodon-derived sparse sequence and trace features further outperformed SELFIES and Group SELFIES across ten QSAR tasks and exceeded classical fingerprint baselines in six out of ten tasks. As an application of the representation, MolCodon BLAST similarity engine decomposes molecular similarity into ring topology, branch context, attachment architecture, and pharmacophore correspondence, enabling interpretable scaffold-hopping searches. In a PARP1 virtual screening study, MolCodon retrieved scaffold-diverse candidates to a known PARP-1 inhibitor Olaparib. Together, these results establish MolCodon as a new molecular representation paradigm that transforms chemical graphs into high-fidelity, interpretable, and alignment-compatible codon sequences, opening a direct path for bioinformatics-inspired analysis of small-molecule chemical space. The MolCodon encoder, decoder, and BLAST similarity engine are freely available as open-source software at https://github.com/DurdagiLab/MolCodon

Drug-induced arrhythmias, particularly Torsades de Pointes (TdP), pose a significant risk to patient safety and can sometimes have life-threatening outcomes. They remain a major concern in drug development and regulation. Machine learning (ML) has become a powerful tool for analyzing complex biological and chemical datasets, enabling researchers to identify subtle patterns that differentiate safe compounds from those likely to cause dangerous cardiac effects. However, most existing in silico approaches do not sufficiently incorporate biological elements, relying heavily on chemical and structural properties or on computationally expensive simulations. Here, we introduce BioMADE, a novel ML framework that harnesses small-molecule-protein activity profiles from publicly available datasets to predict TdP risk without requiring exhaustive mechanistic annotation. Activity data from ChEMBL were used to train individual models for each gene, which predict activity values for any given compound. A curated set of arrhythmia-relevant genes was then used to construct a latent biological embedding (BioMADE embedding) for each molecule. We validated the performance of these features in distinguishing biological elements such as ATC3 class, showing superior classification performance compared with representations such as Molformer (lacks biological information) and MACCS (limited chemical properties) (0.85 AUROC vs 0.81 and 0.73, respectively). BioMADE representations served as input to a support vector machine classifier to discriminate TdP-inducing drugs from safe compounds. BioMADE achieved an AUROC of 0.89 in internal validation, indicating strong predictive performance. Against state-of-the-art models such as ADMEThyst, BioMADE achieved an AUROC of 0.74 on ADMEThysts validation set (vs. 0.72 for ADMEThyst). When we combined both approaches, the AUROC reached 0.77. These results demonstrate that BioMADE provides a scalable, biology-informed, and generalizable approach for predicting drug-induced toxicities. By integrating protein activity profiles into toxicology modeling, our framework highlights the critical role of human biology in adverse drug reaction prediction, an aspect often overshadowed by purely chemical or structural descriptors.

Antimicrobial resistance (AMR) has a profound impact on animal and human health and is associated with substantial morbidity, mortality and public health costs. There is a clear need to develop novel, effective antibiotic agents, which can overcome the current AMR crisis. Antimicrobial peptides (AMPs) may offer such a solution and have attracted growing attention for their potential to combat AMR. In parallel, the growing availability of peptide sequences in public databases has stimulated the development of numerous machine learning and deep learning tools to predict antimicrobial activity computationally. However, it remains unclear how reliably these tools can be compared, as existing studies often rely on heterogeneous datasets and inconsistent evaluation protocols that may lead to data leakage and inflated performance estimates. This raises a central question: what evaluation criteria and benchmark resources are needed to enable fair, reproducible, and biologically meaningful assessment of AMP prediction tools? We address this question by focusing specifically on antibacterial peptides (ABPs). We first provide an overview of AMP databases relevant to antibacterial activity and compare their content, redundancy, and experimental metadata. We then critically assess existing computational tools for ABP prediction, highlighting key limitations related to dataset construction, affinity to certain sequences, data leakage, and inconsistent performance reporting. Based on these limitations, we propose a reference evaluation framework designed to improve comparability, reproducibility, and practical utility in ABP prediction. Finally, we provide targeted recommendations for AMP databases and future tool development to support more robust progress in the computational discovery of ABPs.

Computational protein design demands generally applicable models that reliably predict or generate unmeasured variants with superior functional properties. Although protein language models (pLMs) have been used in zero-shot and transfer-learning design studies, they have generally not been assessed in benchmarks that explicitly test combinatorial extrapolation from lower- to higher-order variants. Here we benchmark widely used pLMs against conventional baseline methods in recently described dense, experimentally validated multi-mutant landscapes. We find that regardless of architecture and parameter count, pLMs are statistically similar to one another, and none consistently outperforms conventional baseline methods. Furthermore, their ability to distinguish functional from non-functional variants in zero-shot prediction is comparable to that of conventional homology-based methods. We suggest that to contribute significantly to the design of protein function, pLMs may need to encode biophysical and structural priors or be combined with structure-based approaches.

The rapid growth of molecular foundation models and large language models has encouraged a scale centred view of AI in drug discovery, in which larger pretrained models are expected to supersede compact cheminformatics models and graph neural networks (GNNs) trained for individual tasks. We test this assumption across 26 endpoints for molecular properties, toxicity, safety liabilities and biological activity, grouped into ADME, toxicity and bioactivity classes. The benchmark contains 78 endpoint and split entries spanning random, Murcko scaffold and structure separated 5-fold CV. Ordered from easiest to hardest, these splits approximate retrospective evaluation on a closed library, scaffold expansion in hit to lead, and library expansion on novel chemotypes. Each entry includes ML, GNN, pretrained molecular sequence and LLM based SAR families. Across 156 fold mean comparisons, classical ML such as RF(ECFP4) and ExtraTrees(RDKit) win 116, GNNs such as GIN and Ligandformer win 25, pretrained sequence models such as MoLFormer and ChemBERTa2 win 12, and LLM based SAR baselines win three. ML dominates random split interpolation but loses part of this advantage under harder splits; GNN and sequence models also decline but gain relative ground, whereas LLM based SAR is weaker in absolute terms yet less sensitive to the split axis. Paired bootstrap analyses support family level trends more strongly than individual model rankings. SAR knowledge derived from training folds improves many GPT5.5-SAR and Opus4.7-SAR metrics but does not make rule based reasoning a universal substitute for supervised predictors. Compact specialized models remain highly effective for molecular property and activity prediction. Larger models add value for SAR interpretation and reasoning in low data settings, but predictive performance depends on the fit among model, task and validation scenario, not on scale alone.

MotivationEnzyme-reaction retrieval is increasingly used to prioritize candidate biocatalysts for experimental follow-up, where useful recommendations should indicate not only whether an enzyme can catalyze a target reaction but also under which pH and temperature conditions it should be tested. Existing retrieval models optimize catalytic matching scores, whereas catalytic optima predictors are typically developed as enzyme-level regressors because public pH and temperature annotations are sparse and often available only at the enzyme or EC-associated record level. This separation leaves a practical gap: high-ranking enzyme-reaction pairs are not evaluated for condition suitability, and enzyme-level optima predictions do not use the reaction context being retrieved. ResultsWe present GERO, a multimodal fusion framework that uses feature-gated cross-modal fusion to integrate global enzyme sequence semantics, sequence-derived pocket geometry, and molecular reaction representations for condition-aware enzyme-reaction retrieval and catalytic optima estimation with reaction context. To evaluate this setting, we define the tolerance-restricted hit rate (Hit@k-TR), which requires both top-k retrieval of the correct candidate and condition prediction within predefined tolerances. Across enzyme- and reaction-similarity splits, GERO improves Hit@k-TR over two-stage retrieval-then-prediction baselines. Representative benchmark examples and an iodinin biosynthesis case study further illustrate GEROs ability to provide candidate rankings together with plausible assay-condition estimates for downstream experimental prioritization. Availability and implementationSource code is available at https://github.com/ykxhs/GERO. Contactliujuan@whu.edu.cn Supplementary informationSupplementary data are available at XXXX online.

MotivationFixed-backbone sequence design methods such as ProteinMPNN operate on backbone coordinates alone and cannot represent target side-chains at the binding interface. Their decoding algorithm also lacks a mechanism to balance binding affinity and folding stability or to improve selectivity against structurally similar off-targets. These gaps limit the computational design of protein binders with high affinity and specificity. ResultsWe present RedNet, a multiscale graph neural network that encodes side-chain information of the binding target. We further develop a contrastive decoding algorithm, motivated by the thermodynamic decomposition of binding free energy, that addresses two objectives: (1) balancing binding affinity and folding stability, and (2) improving selectivity against structurally similar off-targets. RedNet reaches 43% native sequence recovery on heterodimers, compared with 37% for ProteinMPNN and 33% for ESM-IF. With contrastive decoding, it matches native-sequence co-folding success (68%) on high-confidence AlphaFold3 targets, exceeding ProteinMPNN (59%) and ESM-IF (61%). On a new benchmark of structurally similar on-/off-target pairs, RedNet with contrastive decoding reaches 64.8% energetic selectivity, ahead of PiFold (55.6%), ProteinMPNN (53.7%), and ESM-IF (53.7%). AvailabilitySource code and datasets are released at https://github.com/zw2x/rednet_public. Contactjinbo.xu@gmail.com