mAbs

Antibody developability is increasingly predictable from sequence, yet software and trained models are rarely made available. We present DELPHI, open software for training developability predictors from labelled antibody assay data, together with ready-to-run, retrainable models. DELPHI compares 25 language-model and classifier combinations under CDR H3-cluster cross-validation that reduces sequence-similarity leakage, measures how performance changes with labelled training-set size, and reports residue-level model attributions. Applied to in-house polyreactivity and size-exclusion (SEC) data, it reaches mean AUC 0.959 and 0.933. Trained on those data alone, it transfers to a 246,293-antibody public library (AUC 0.950 with our deployed model) and ranks polyreactivity at a level similar to the best reported Ginkgo competition point estimate, without training on its data. Any laboratory can screen candidates before running assays, generate residue-level engineering hypotheses, and retrain DELPHI for a new assay.

Antibody developability is determined by multiple physicochemical properties, and the relevance of individual properties varies among experimental assays. Existing computational approaches, however, often provide fixed assessments or rely on computationally intensive descriptors. We present kitAb, an open-source framework for scalable physicochemical characterization of antibody Fv regions and assay-specific developability modeling. kitAb calculates 59 interpretable sequence- and structure-based descriptors and integrates them with automated feature selection and regression. We evaluated kitAb across ten public datasets comprising 46 experimental endpoints, using sequence-aware or predefined splits where feasible. Predictive performance was strongly assay-dependent. In nested evaluations with sequence-aware or predefined data splits, models using kitAb descriptors achieved predictive performance comparable on average to those using PROPERMAB descriptors, while outperforming models based on TAP or sequence-only features. In a single-structure CPU timing benchmark, kitAb descriptor calculation was approximately 11-fold faster than PROPERMAB. Predictive performance was similar across the three structure-generation backends, and broader AutoML searches did not substantially improve mean performance. We further applied the same descriptor representation to approximately 3.4 million paired natural antibodies and 706 Phase II or later therapeutic antibodies. Repertoire-scale analysis revealed systematic differences in surface charge, hydrophobic and aromatic organization, CDR length and germline-associated physicochemical profiles. The kitAb code and analysis workflows are openly available at https://github.com/katyachemistry/kitAb/.

Structure-guided antibody CDR design constrains residue selection using backbone geometry, but structural evidence alone may leave multiple amino acids plausible. We investigate whether the fixed antibody framework provides complementary sequence information for resolving this ambiguity. We introduce AbRefine, an information-aware fusion framework that combines backbone-conditioned structural predictions with framework-conditioned sequence distributions derived from a pretrained antibody language model with all designed CDRs masked. Its Information-Aware Fusion (IAF) module learns residue-specific fusion weights from the uncertainty, disagreement, and information content of the two predictive distributions, while keeping both pretrained experts frozen. Controlled interventions show that the signal is target specific: matched frameworks outperform shuffled frameworks by 7.37 percentage points, and the largest gain occurs at positions with high structural uncertainty and informative framework evidence (+5.77$ points). Compared with static fusion, IAF further improves AbMPNN recovery by +1.53 points (95\% CI +0.93$ to +2.14$) and significantly reduces native negative log-likelihood for both AbMPNN and AntiFold, without detectable degradation in fold-back structural compatibility. These results show that framework context provides information beyond backbone geometry and that learning its residue-specific utility enables effective adaptive fusion for antibody CDR design.

De novo antibody design promises to transform therapeutic antibody and nanobody discovery. Yet, success remains inconsistent across targets, with generative models often yielding no experimentally validated binders. Here we introduce MIMOSA (MIMic-Oriented Structural Antibody generation), a model-agnostic, inference-time framework that constrains diffusion-based antibody design pipelines to reproduce the interaction geometry and chemistry of a target's known cognate binder interface. Rather than searching for a productive interface from scratch, MIMOSA constrains generation around residue identities and geometries already known to support binding, while allowing the generative model to complete the surrounding antibody interface. We test MIMOSA on different interface topologies: contiguous motifs that fit within a single CDR loop and spatially dispersed hotspots distributed across the paratope. Applied without retraining to two architecturally distinct generators (RFAntibody and BoltzGen), this framework delivers experimentally validated binders-achieving sub-micromolar to single-digit nanomolar affinities-at 20-26% hit rates on three targets on which unconstrained methods produce zero or near-zero binders: KEAP1, uPA, and IL-8. By turning any target with a structurally characterised cognate binder into an accessible de novo design problem, this framework broadens the practical reach of generative antibody design to targets where unconstrained approaches currently fail.

Accurate modelling of biomolecular interactions is fundamental to drug discovery, yet current artificial intelligence (AI) workflows remain fragmented across structure prediction, affinity estimation, molecular design, and experimental decision-making. We introduce AnewDDE, an agentic Drug Discovery Engine that connects these capabilities into a closed-loop system for biomolecular interaction modelling and design. We demonstrate several components of this system: AnewFold delivers superior performance on challenging targets, including antibody-antigen complexes and molecular glues, providing more reliable structural hypotheses together with pocket identification and conformational analysis; AnewDesign achieves a 10.7% success rate in identifying binders with single-digit-nanomolar affinities measured by surface plasmon resonance (SPR) in a representative nanobody discovery campaign, substantially reducing the experimental search burden; AnewAffinity combines accuracy approaching that of free-energy perturbation (FEP) methods with the speed required for high-throughput screening and prioritizes promising transformations; and AnewMind, a large language model (LLM) at the hundred-billion-parameter scale that underwent in-house full-parameter post-training, provides scientific reasoning and absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction for pharmaceutical research and development (R&D). To systematically evaluate AnewMind, we introduce PharmBench, an internal benchmark designed by pharmaceutical R&D experts to evaluate long-horizon decision-making in drug discovery. Evaluations on ADMET tasks and PharmBench show that AnewMind performs competitively with leading frontier models. Together, these capabilities establish AnewDDE as an integrated engine for addressing challenging therapeutic targets and translating molecular insights into experimentally validated candidates at scale.

We describe the concept for a new modality to create multi-specific therapeutics that utilizes modified nucleic acids as a scaffold to concatenate multiple warheads (hence called 'catamers'). Catamers are assembled via Watson-Crick interactions into multi-specific molecules from individual components such as ligand-binding modules, protease-cleavage sites, and toxins. Each warhead is covalently bound to an oligonucleotide, which in turn is connected to the catamer scaffold. Post-synthesis of the individual catamer components, they are annealed together such that Watson-Crick base pairs on the oligonucleotides dictate the organization of the warheads on the complete catamer. Notably, catamers are not aptamers, though they can contain aptamers as warheads, along with small molecules and peptides. By design, catamers are intended to address three key problems that have hamstrung the field of multi-specifics: (i) optimization rate; (ii) manufacturing cost; and (iii) unpredictable immunogenicity (Amash et al., 2024 PMID: 39189686). Catamers are purely synthetic and do not require cultured cells to produce. These advantages are based on the underlying design of the catamer and will require experimental validation in future studies. As a proof-of-concept catamer, we focus on a prostate-specific membrane antigen (PSMA)-targeted toxin conjugate. In molecular simulations, this catamer displays structural characteristics to support its intended pharmacology. Importantly, the design addresses one of the key shortcomings of antibody drug-conjugates (ADCs): off-target toxicity caused by release of the toxin after pinocytosis by healthy cells in contact with the blood.

Directed evolution enables the discovery of protein mutants with improved fitness through iterative rounds of selection and has been widely applied to adeno-associated virus (AAV) capsid engineering. Machine learning (ML) can augment this process by prioritising mutants for experimental validation, but whether its benefits outweigh the added cost of ML-designed library construction remains unclear. Here, we address this question by considering improvements in manufacturing efficiency of AAV capsids in the context of exosomal encapsulation, which is a technique for improving AAV immunogenicity. We generated a directed evolution dataset comprising 53,974 mutants across three rounds of selection and an independent assessment dataset of 472 ML-designed mutants with detailed profiling. Directed evolution alone yielded a 3-fold improvement, while ML-assisted directed evolution yielded a 41-fold improvement in manufacturing efficiency, demonstrating that ML-assisted directed evolution outperforms directed evolution alone. We show that data quality - specifically the number of selection rounds and the statistical power of mutant counts - has a greater impact on model performance than model architecture. Finally, we provide practical guidance on experimental design and implementation of ML tuning strategies that can augment model performance in protein engineering applications.

Designed binders are routinely filtered on predicted interface confidence, which reports how self-consistently a model places an interface, not where it lies. We designed 200 nanobody loops against the survivin BIR-domain groove, co-folded 196, and checked epitope contact in every predicted complex. Confidence showed no detectable association with epitope engagement ({rho} = +0.100, 95% CI -0.04 to +0.24, n = 196); its highest and lowest quartiles failed alike (odds ratio 0.97, 95% CI 0.41 to 2.26). It tracked instead the run's other confidence estimates ({rho} = +0.61). Twelve-seed replication separated two candidates the criteria had not: one reached the complete epitope in 10 of 12 runs, the other never. Across 155 predictions the top-ranked model was no better than one drawn at random from its five (P = 0.610), though only 3 of 31 failed on every model. Nothing was assayed; nb_cand_057 is an in silico candidate only.

Constrained macrocyclic peptides can engage protein surfaces that resist both small molecules and biologics but producing them at library scale still depends on chemical synthesis and macrocyclisation workflows that are slow, specialised and hard to parallelise. We describe a recombinant pipeline for disulphide-stapled cyclic peptides, compatible with 96 well plates, that lowers this barrier. Peptide-encoding sequences are introduced by extension PCR and in vivo assembly (IVA) cloning at about $2 per variant, expressed as small ubiquitin-like modifier (SUMO) fusions secreted to the oxidising periplasm, where the intramolecular disulphide forms spontaneously, and purified by immobilised metal-affinity chromatography (IMAC) with analytical size-exclusion chromatography (SEC). The route from a list of designs to characterised material takes about one week, much of it unattended. Applied to a 60-design library across nine protein targets, the pipeline gave purified products for all 60 designs, at a median total soluble yield of approximately 135 micrograms per 6 mL culture (range 9 to 261) that varied by target. Ellman's assay across 38 designs showed disulphide-formation failure to be sequence-driven and uncorrelated with yield, detecting a failure mode invisible to chromatography. Surface plasmon resonance confirmed target binding and showed that the recombinant SUMO fusion constructs retained affinities comparable to their corresponding peptide-only forms. Complementary to general-purpose recombinant platforms, the pipeline enables higher-throughput disulphide macrocycle screening and is amenable to partially or fully automated workflows.

The resurgence of the Mpox virus (MPXV) highlights the urgent need for scalable, mutation-resistant countermeasures. Slow development timelines hinder traditional empirical vaccine discovery, while conventional vaccinology often yields high false-positive rates because it relies on static, linear sequence-based screening. These classical informatics filters consistently fail to predict physical structural stability or the dynamic presentation of immune cells under physiological conditions, creating a distinct translational gap between computational design and in vitro efficacy. To address these limitations, here we engineered an advanced, AI-augmented vaccine discovery pipeline targeting the most virulent components of the MPXV proteome. We systematically screened eight critical viral targets essential for replication, membrane morphogenesis, and host-cell entry to map highly conserved T-cell and B-cell epitopes. Moving beyond standard heuristic filtering, we implemented a high-fidelity deep learning validation layer: ESM-2 protein language models (pLMs) were deployed to calculate pseudo-perplexity scores across chimeric linker junctions (EAAAK, GPGPG) to guarantee structural naturalness, MHCflurry ensemble models successfully verified active intracellular antigen presentation, and structure-informed graph neural networks (GNNs) simulated the physical node-attention dynamics of the immune synapse. Finally, recurrent neural networks (RNNs)-driven mRNA kinetic modeling optimized translation initiation rates and minimum free energy profiles during in silico cloning. This multi-layered pipeline successfully yielded three structurally refined, multi-epitope vaccine constructs optimized for maximum global population coverage. ESM-2 structural validation confirmed exceptionally low perplexity scores at structural junctions, indicating that the engineered chimaeras have native-like stability and a minimized risk of proteolysis. GNN-based immune synapse modeling demonstrated highly stable binding kinetics and robust binding free energies between the target epitopes and human HLA alleles. Furthermore, RNN kinetic optimizations established that the cDNA sequences are optimized for high-yield expression in heterologous expression hosts without risk of ribosomal stalling or translational bottlenecks. This study establishes a deep learning-augmented framework that bridges the gap between theoretical sequence optimization and biopharmaceutical scalability. By validating the structural naturalness, real-world immunogenicity, and manufacturing kinetics of our vaccine chimeras, this framework presents a robust, proactive paradigm for rapid vaccine architecture against emerging orthopoxvirus threats.

Relating the drivers of binding affinity to structural features remains difficult because affinity emerges from many weak, context-dependent interactions, while experimental affinity and mutation data are limited. Here, we ask whether a machine-learning model can learn energetically relevant interaction preferences directly from native structures, without affinity or mutation labels. We decompose interfaces into local structural motifs and learn the compatibility of each motif with its surrounding molecular environment. Applied across a protein surface, these models yield target preference maps (TPMs), spatial fields of local interaction compatibility. Trained exclusively on native antibody-antigen structures without affinity or mutation labels, TPMs recover charged, aromatic and backbone-mediated recognition across peptide-protein and other non-antibody interfaces. On an out-of-domain SKEMPI benchmark, wild-type TPM scores discriminated mutation-sensitive positions (AUC 0.643, n=1323), while TPM-based substitution scores ranked alternative amino acids comparably to FoldX and Rosetta Flex {Delta}{Delta}G. These results show that native interface structures can supervise learning of transferable, energetically relevant molecular preferences without direct thermodynamic labels.

Recent advances in deep learning-based structure prediction, including AlphaFold3 and the open-source Boltz model family, have extended biomolecular modeling to joint prediction of protein-ligand, protein-nucleic acid, and multi-chain complexes with binding-affinity estimation. Boltz-2 is among the most feature-complete of these open models, but its practical use requires a local CUDA-capable GPU, command-line execution, and manually authored YAML configuration files, limiting accessibility for researchers without dedicated computational infrastructure. We developed Boltz2-Notebook, a Colab-native interface comprising four integrated stages - automated environment setup, interactive parameter-to-YAML generation, execution management, and automated confidence and affinity visualization - together with a manifest-driven batch mode for multi-target screening. All modelling capabilities are inherited unmodified from Boltz-2; Boltz2-Notebook's contributions are limited to accessibility, input construction, and workflow automation. Independent of the software, we curated a benchmark of 317 protein-ligand pairs (122 proteins, 277 ligands) from BindingDB and predicted binding affinity in triplicate using the Boltz-2 command-line engine on high-performance computing infrastructure. Predicted and experimental pIC50 values showed moderate correlation (Pearson r = 0.609 [95% CI 0.540-0.675]; Spearman {rho} = 0.625; R2 = 0.371; MAE = 0.968 pIC50 units), with high triplicate reproducibility (pairwise r = 0.97) but a systematic compression of the predicted affinity range and no measurable relationship between Boltz-2's self-reported confidence metrics and prediction accuracy. Boltz2-Notebook is freely available as open-source software and provides external, reproducible evidence - including a specific confidence-calibration limitation - relevant to interpreting Boltz-2 affinity predictions responsibly.

Accurate and fast prediction of protein binding sites remains essential for understanding molecular interactions and facilitating protein engineering. Here, we present SurfGraphPro, a geometric deep learning approach that bridges protein language models with surface-based structural representations for a rapid binding interface prediction. Our method operates on coarse-graphed triangulated protein surfaces, utilizing solvent-excluded surface meshes downsampled into amino acid residue centered patches. Unlike existing approaches that rely on extensive physicochemical feature engineering, we leverage evolutionary information directly from protein language model embeddings, eliminating the need for computationally expensive multiple sequence alignments and hand-crafted features. This integration achieves on average $\sim18-28 \times$ speedup for proteins of $\sim$100 to a few 1000s of amino acids over current state-of-the-art surface-based model while maintaining comparable accuracy on diverse binding interfaces, including challenging antibody-antigen complexes. To our knowledge, this represents the first approach to integrate protein language model embeddings with coarse geometric surface representations for binding site prediction, demonstrating that learned evolutionary features coupled with geometric transformers can replace traditional feature engineering without sacrificing performance.

Cyclic peptides can engage extended protein interfaces, but sequence changes that improve developability often disrupt the selected target-bound geometry. Existing property-guided design methods lack a principled way to preserve this geometry during local multiproperty optimization. Here we present RF-PepTune, which separates structural eligibility from property prioritization. Before any property score is queried, RF-PepTune evaluates single substitutions and all position-distinct pairs induced by admitted substitutions to build a parent-specific, pair-complete admission graph. Pareto Monte Carlo tree search then optimizes predicted permeability, solubility, non-fouling, and low-haemolysis scores exclusively within this graph, and selected sequences undergo full structural re-evaluation. In a matched three-arm comparison of 334 parents across MDM2, GABARAP, and MCL1, joint property-structure success rose from 10.2% without admission to 51.5% with single-substitution admission and 77.5% with pair-complete admission; pair-complete gating improved success by 26.0 percentage points over single-substitution gating (95% CI, 20.7-31.4). Property-positive but structure-failing endpoints fell from 88.3% to 9.3%. In a separate target-balanced set, all 90 prioritized outputs improved all four model-native scores while satisfying prespecified structural criteria. This work establishes pairwise structural gating as a general strategy for constrained multi-objective molecular optimization.

Antidiabetic peptides (ADPs) are short bioactive sequences of therapeutic interest, and sequence-based classifiers prioritise experimental candidates. A classifier is useful only if its accuracy transfers to unseen sequences, which depends on how the benchmark was as sembled and partitioned. We term the distance between what such a classifier is scored on and the function it is meant to predict the provenance-to-function gap, and we quantify it. We re-evaluate the two-layer ADP benchmark of Basith et al. under a protocol that groups homologous peptides rather than splitting them at random. Of the 877 ADPs, 218 attribute to a precursor protein by exact substring containment; within that subset the second-layer label coincides exactly with precursor identity, all 140 human-insulin fragments carrying the type-1 label and all 76 bovine milk-protein fragments the type-2 label, without exception. Accuracy tracks identity to the training set, rising from a Matthews correlation coefficient (MCC) of 0.39-0.43 below 50% identity to 0.92-0.96 between 70% and 90%, and peptide length alone reaches MCC 0.619 on held-out data. Auditing sixteen further peptide bench marks shows the coupling is not confined to this resource: fragment families share a label more often than chance in eleven of fourteen testable datasets, and not in three, so the prop erty is common rather than universal. Rebuilding the published architecture on identical folds shows its advantage over a single classifier is a function of the split: present under random partitioning, absent once homologues are separated. ADP-Hybrid, one tree-ensemble classifier per layer, reaches MCC 0.843 on layer 1 against a published 0.841 at 21-38 times the inference throughput, and 0.801 against 0.858 on layer 2. We release the protocol and the controls that expose these properties.

Physiological compartments span a wide pH range, from blood plasma near pH 7.4 through the mildly acidic tumor microenvironment to the acidified endosome and lysosome, and binders whose affinity responds to these differences underpin recycling antibodies, catalytic extracellular degraders and tumor-conditional therapeutics. Such binders are still obtained mainly by combinatorial histidine scanning and display selection, which is laborious and yields few transferable design rules. Recent computational approaches have enabled the design of pH-sensitive binders that engage at physiological pH and release under acidic conditions, by introducing histidines at key positions so that protonation generates electrostatic repulsion, evaluated through physics-based calculations. However, because the design model has no explicit representation of protonation state, it cannot shape the residues around a titratable site to favor one state over the other, leaving pH sensitivity to be screened for rather than specified. Here we present Proton-PottsMPNN, a sequence-design model that explicitly represents the protonated and deprotonated states of histidine, aspartate and glutamate as distinct sequence tokens. Because protonation state is specified rather than inferred, the model redesigns the microenvironment around a titratable residue so that the chosen protonation state is energetically favored over its alternative. Using Proton-PottsMPNN we designed a library of 8,407 de novo PD-L1 candidate binders in which sequence-structure compatibility and protonation state preference were designed jointly under a single energy function, with no structure-prediction filtering applied. The full library was synthesized and displayed on yeast, then screened against PD-L1 across a series of pH values by fluorescence- and magnetic-activated cell sorting, recovering 237 and 288 unique pH-dependent designs respectively. Transition pH values spanned 4.0 to 5.8, and switches were centered on protonated aspartate and glutamate as well as histidine. Recovered designs showed a stronger model preference for their designed protonation state than designs from the same backbone that were not recovered, and this preference was strongest in the designs that required the most acidic conditions to bind. Proton-PottsMPNN therefore designs binders that engage PD-L1 under acidic conditions fully de novo, directly from a specified protonation state.

N-linked glycans are ubiquitous post-translational modifications that mediate diverse biological processes, including cell signaling and adhesion, molecular recognition, and host-pathogen interactions. Despite their importance, N-linked glycans have largely remained beyond the reach of biomolecular design due to a lack of methods capable of jointly modeling glycans and protein structure. Here, we develop a general all-atom computational pipeline for creating de novo glycoproteins in which the amino acid sequence encodes both the location and composition of N-linked glycans. The method uses three-dimensional protein structure to tune glycan accessibility to processing enzymes in the eukaryotic secretory pathway. Systematic increases in steric restriction progressively reduced processing, shifting glycans from heterogeneous complex-type structures toward increasingly underprocessed oligomannose-type species resembling those found on many viral glycoproteins. Despite this programmed restriction, the resulting oligomannose glycans remained accessible for biological recognition, inhibiting uropathogenic Escherichia coli adhesion to human bladder cells and binding mannose-binding lectin. These results establish steric control of N-glycan processing as a genetically encodable design principle and extend computational protein design to functional glycoproteins with programmed composition.

Scientific design requires reasoning in scientific spaces that represent the objects being designed. Building on Scinetics' Chain of Science Tokens, we present chain-of-structures (COS) as a concrete protein-design case study that connects reasoning and generation through explicit structural states. In this technical report, we apply COS to enzyme design, where catalytic function requires precise local geometry within a compatible global fold, making it a demanding test of reasoning from chemical requirements to protein structure. Starting from a reaction specification and target length, COS couples a language reasoning expert with an all-atom generation expert across four rounds: catalytic core, active site, global backbone, and complete sequence and structure. Each reasoning state directly conditions generation, and the resulting structure becomes context for the next decision. Discrete residue constraints and continuous latents connect the two experts, enabling structural supervision of the reasoning interface. On the evaluated benchmark, COS leads the compared baselines on five of six metrics, achieving 57.0% EC3 agreement and a TM-score of 0.729.

Structure-guided mutation design requires relating a proposed substitution to its molecular environment to anticipate its effects. Existing evaluations of large language models (LLMs) in protein engineering emphasize sequence-based prediction or tool-integrated workflows, leaving unclear whether LLMs themselves can infer mutation effects from structural evidence without dedicated predictors. We evaluate physical intuition-qualitative reasoning that connects structural observations with molecular interaction principles-using a benchmark that tests the same substitution in the same protein across different binding partners, with predictions based on local atomic coordinates and no dedicated mutation-effect predictors. In the tested comparisons, the best-performing LLM achieved group accuracy comparable to the evaluated energy predictors and Biomni, an LLM-driven tool workflow, while the methods differed in their prediction patterns. Most evaluated LLMs achieved higher group accuracy for larger affinity improvements. Structural interventions support selective dependence on mutation-site and partner information, while assessable structural claims largely agree with the supplied coordinates. These findings support direct LLM structural reasoning as a basis for developing rational mutation-design tools that connect molecular observations with partner-specific effects.

Cyclic peptides can engage targets beyond the reach of small molecules, but their discovery typically relies on libraries reused across targets, diversification around a known interaction motif, or computational methods limited to a handful of non-canonical building blocks. Here, we present nCycle-Forge, a simulated annealing framework that designs cyclic peptide binders by iteratively substituting building blocks and scoring each proposal by cofolding with the target. We designed, synthesized, and screened 1,472 cyclic peptides against the active-site pocket of Thrombin and the shallow p53-binding surface of MDM2 from the target sequence alone and without fixing binding mo-tifs. The library yielded chemically diverse hits against both targets in a single round without experimental optimization, with hit rates comparable to or up to 24-fold higher than in larger experimental libraries. Purified designs reached sub- to low-micromolar potency, with the best Thrombin inhibitor at Ki 0.14 M and the best MDM2 binder at EC50 2.4 M. Surface plasmon resonance confirmed direct binding, competition with or-thosteric inhibitors at both sites, and weaker binding under conditions expected to linearize the designs. nCycle-Forge requires no building block-specific parameterization, conformer generation, or retraining, and provides a target-focused route to functional cyclic peptides beyond the canonical amino acid alphabet.