2. Biomolecular Design Tools

ABACUS-T: The latest in the ABACUS family of protein sequence design models (from the Liu lab at USTC), ABACUS-T performs multimodal inverse folding, generating sequences conditioned jointly on structure and additional functional context. It matches or exceeds many SOTA inverse folding models on standard metrics, and was experimentally validated by redesigning four proteins — an allose-binding protein (improved affinity), endo-1,4-β-xylanase (improved activity), TEM1 β-lactamase (improved activity), and OXA β-lactamase (improved substrate selectivity).
Paper published: November 19, 2025
Paper | Code (Zenodo)

Caliby / SolubleCaliby: Ensemble-conditioned protein sequence design models from Po-Ssu Huang’s group aimed at extending structure-conditioned sequence design to non-idealized protein backbones. Caliby is a Potts-model-based method that conditions on a synthetic ensemble of structures generated from an input backbone, sampling sequences consistent with the ensemble’s structural constraints while averaging out non-structural signals (e.g. phylogenetic bias) that native-sequence-recovery training objectives tend to pick up; SolubleCaliby extends the approach with a solubility design objective.
Preprint posted: October 2, 2025
Preprint | Code (GitHub)

MapDiff: A discrete diffusion probabilistic model from the University of Sheffield, AstraZeneca, and the University of Southampton for inverse protein folding. MapDiff iteratively denoises amino acid sequences conditioned on a target protein backbone using a graph-based denoising network with a mask-prior pretraining strategy, accurately capturing both structural information and residue interactions. Across four challenging sequence design benchmarks, MapDiff substantially outperformed prior state-of-the-art methods, with particular strength in predicting uncertain/flexible regions and in silico foldability of generated sequences.
Preprint posted: December 10, 2024
Paper published: June 16, 2025
Paper | Preprint | Code (GitHub)

ESMC (ESM Cambrian): A family of protein language models from EvolutionaryScale, released alongside (but architecturally distinct from) the generative ESM3 — ESMC focuses on producing high-quality representations of protein biology rather than controllable generation. Trained and scaled across 300M, 600M, and 6B-parameter variants, ESMC defines a new scaling frontier relative to ESM2, with stronger emergent long-range structural understanding as model scale increases; the 300M and 600M models are released as open weights, with the 6B model available via hosted inference.
Preprint posted: December 4, 2024
Blog Post | Code (GitHub) | Model (HuggingFace)

ESM3: A frontier multimodal generative language model from EvolutionaryScale that jointly reasons over and generates protein sequence, structure, and function, represented as discrete token tracks processed by a bidirectional transformer trunk. At its largest scale, ESM3 was trained with 1.07×10²⁴ FLOPs on 2.78 billion proteins (771 billion unique tokens) across 98 billion parameters, and was used to generate a novel green fluorescent protein (esmGFP) with sequence identity comparable to proteins separated by over 500 million years of natural evolution.
Preprint posted: July 1, 2024
Paper published: February 21, 2025
Paper | Preprint | Code (GitHub)

AF2_seq-MPNN: A protocol, based on AF2_seq and ProteinMPNN, for the computational design of topologically complex protein folds and soluble analogues of membrane proteins.
Paper published: June 19, 2024
Paper | Preprint | AF2_seq Code (GitHub) | Colab Notebook 1 | Colab Notebook 2

CarbonDesign: A protein sequence design method that adapts the success ingredients of AF2. CarbonDesign utilizes Inverseformer, a network architecture adapted from AlphaFold’s Evoformer. Based on scTM score, sequence recovery rate, and BLOSUM score, CarbonDesign outperformed other published methods, including versions of ProteinMPNN and ESM-IF, on CAMEO and CASP15 test sets, as well as RFdiffusion-generated backbones. CarbonDesign is also capable of predicting the effects of mutations on protein function.
Paper published: May 23, 2024
Paper | Preprint | Code (GitHub) | Code (Code Ocean)

AntiFold: An inverse folding model for antibody sequence design based on ESM-IF1.
Preprint posted: May 06, 2024
Preprint | Code (GitHub) | Webserver | Colab Notebook

SPDesign: A method that combines structural sequence profile, fast shape recognition, and pre-trained language models for protein sequence design.
Paper published: April 09, 2024
Paper | Webserver | Code

Evo: A long-context foundation model that generalizes across the central dogma of biology: DNA, RNA, and proteins. Evo is a 7 billion parameter model trained to generate DNA sequences and is capable of prediction and generative tasks, from molecules to whole genomes.
Preprint posted: March 06, 2024
Preprint | Code (GitHub) | Code (PyPI) | Blog | Playground | Colab Notebook

PocketGen: A method for generating full-atom ligand-binding pockets to design small molecule-binding proteins. PocketGen uses a co-design strategy that, given the ligand and the scaffold, simultaneously designs the sequence and structure of the protein pocket.
Preprint posted: Feb 28, 2024
Preprint | Code (GitHub) | Blog

CoVES: A “simple” model to design functional and diverse combinatorial protein variants.
Paper published: Feb 22, 2024
Paper published: Jan 06, 2024
Paper | Code (GitHub) | Code (Zenodo)

Multiflow: A generative model for protein sequence and structure co-design based on the DFM and FrameFlow models.
Preprint posted: Feb 07, 2024
Preprint | Code (GitHub)

Proteo-R1: A reasoning-guided protein design framework (accepted at ICML 2026) that decouples molecular understanding from geometric generation via a dual-expert architecture. A multimodal large language model acts as an “understanding expert,” analyzing protein sequence, structure, and textual context to identify functionally critical residues governing binding and specificity; these residue-level decisions are then passed as hard constraints to a diffusion-based “generation expert,” which performs conditional co-design while respecting the fixed interaction anchors — mirroring how human protein engineers reason about critical interactions before optimizing geometry.
Preprint posted: May 2026
Preprint | Code (GitHub)

PPIFlow: A flow-matching framework from Changping Laboratory for de novo protein binder design, paired with an in silico affinity maturation stage. PPIFlow uses a pairformer architecture to reason explicitly over pairwise geometric and chemical interactions, modeling protein backbone rigid-body transformations as continuous flows; the maturation stage then combines interface rotamer enrichment with partial flow refinement to optimize energetic packing. Across seven mini-binder targets (210 designs total), PPIFlow achieved a 36.2% sub-µM hit rate, with best K_D values reaching 1 pM (PDGFR, VEGFA), 1.38 pM (IFNAR2), and 5.14 pM (PD-L1), and also generated pM–nM-affinity VHH nanobodies for 7 of 8 additional targets.
Preprint posted: January 19, 2026
Preprint | Code (GitHub)

BoltzGen: BoltzGen is an all-atom generative model for universal binder design that preserves the Boltz-2 architecture and diffusion process, extended with a design specification language for covalent bonds, structural constraints, and binding sites. By reframing binder design as a structure-prediction task jointly trained on folding and design objectives, it matches state-of-the-art folding performance while generating proteins, peptides, and nanobodies across diverse target modalities. Across 8 wet-lab design campaigns spanning 26 targets, nanobody designs against 10 novel targets dissimilar to any bound structures in the PDB achieved a 60% hit rate testing 15 or fewer designs per target, with a 50% hit rate for protein binder designs on the same target set.
Preprint posted: November 20, 2025
Preprint | Code (GitHub)

RFD2-MI (RoseTTAFold Diffusion 2 for Molecular Interfaces): A fine-tuned, per-residue-conditioned version of RFdiffusion2 from the Baker lab for all-atom generative design of binders to protein, small-molecule, and covalently/post-translationally modified targets, using SE(3) flow matching. The method was validated by designing binders to four phosphotyrosine sites (CD3ε, EGFR, INSR, STAT5) with affinities comparable to native pY-binding domains (EGFR pY1173 Kd as low as 577 nM) and strong phospho-selectivity, with co-crystal structures matching design models within 2.1–2.4 Å RMSD. A fine-tuned RFD2-MI was further used to design de novo zinc proteases built around a minimalist metallohydrolase active site plus an adjacent peptide-binding groove; in one round of 135 designs, 36% showed site-specific proteolytic activity, with the best design accelerating peptide-bond hydrolysis over 10⁸-fold, and designs were reprogrammed to cleave human TDP-43.
Preprint posted: November 21, 2025
Preprint (Zinc Proteases) | Preprint (Phosphotyrosine Binders)

RFdiffusion2: A generative model from the Baker lab, building on RFdiffusion All-Atom, for atom-level enzyme active-site scaffolding directly from sequence-agnostic descriptions of functional-group geometry, without requiring predefined sequence indices or inverse rotamer generation. On a 41-site in silico benchmark it successfully scaffolded all 41 active sites (versus 16/41 for prior deep-learning methods), and enzymes designed for three catalytic sites were experimentally validated, with active catalysts found among fewer than 96 sequences tested per site.
Preprint posted: April 10, 2025
Preprint | Code (GitHub)

GENzyme: A de novo enzyme design model from McGill/Mila and Shanghai Jiao Tong University, conditioned on a target catalytic reaction rather than a fixed binding-pocket geometry — addressing the fact that enzyme-substrate interactions are rarely a perfect lock-and-key fit. It is an end-to-end, three-stage pipeline: catalytic-pocket generation and sequence co-design, pocket inpainting and enzyme inverse folding to build the full enzyme structure, and a binding/screening module that optimizes and predicts the enzyme-substrate binding complex.
Preprint posted: November 10, 2024
Preprint | Code (GitHub)

Protpardelle: A generative model for structure-based protein design that sets a new state-of-the-art by outperforming existing models on various metrics. Genie 2 builds on Genie (version 1) to expand the structure space (size and diversity) captured by the model.
Paper published: Jun 25, 2024
Paper | Code (GitHub) | Preprint | Web Demo (HuggingFace)

Genie 2: A generative model for structure-based protein design that sets a new state-of-the-art by outperforming existing models on various metrics. Genie 2 builds on Genie (version 1) to expand the structure space (size and diversity) captured by the model.
Preprint posted: May 24, 2024
Preprint | Code (GitHub)

RFdiffusion All-Atom (RFdiffusionAA): A diffusion model that generates de novo protein structures around small molecules and other non-protein targets. RFdiffusionAA was developed by fine-tuning RosettaFold All-Atom (RFAA) on structure denoising.
Paper published: March 07, 2024
Paper | Code (GitHub)

Proteus: A diffusion model for de novo protein backbone generation that, based on multiple metrics, surpasses other leading methods, including RFdiffusion, Genie, FrameDiff, and Chroma.
Preprint posted: Feb 12, 2024
Preprint | Code (GitHub)

Multiflow: A generative model for protein sequence and structure co-design based on the DFM and FrameFlow models.
Preprint posted: Feb 07, 2024
Preprint | Code (GitHub)

FrameFlow: FrameFlow is an SE(3) flow matching model for fast protein backbone generation, adapted from the diffusion-based FrameDiff model. FrameFlow has also been extended to the motif-scaffolding task.
Preprint posted: Jan 08, 2024 | Oct 10, 2023
Preprint 1 | Preprint 2 | Code (GitHub)

RFdiffusion: A generative diffusion model for de novo protein backbone design, built by fine-tuning the RoseTTAFold structure-prediction network on a structure-denoising task. RFdiffusion achieves strong performance on unconditional protein monomer design, protein binder design, symmetric oligomer design, and both enzyme active-site and symmetric motif scaffolding, with hundreds of designs experimentally validated.
Paper published: July 11, 2023
Paper | Code (GitHub)

ProteinGenerator: A RoseTTAFold-based sequence-space diffusion model from the Baker lab that jointly generates protein sequence and structure by iterative denoising from random amino-acid sequences, guided by arbitrary sequence and/or structural attributes. Demonstrated applications include designing proteins enriched for specific amino acids, proteins with internal sequence repeats, proteins with masked bioactive peptides (e.g. melittin) embedded in a scaffold, and multistate “parent-child” structures; experimentally validated designs were monomeric, had the intended secondary structure, and were thermostable up to 95°C.
Preprint posted: May 8, 2023
Paper published: September 25, 2024
Paper | Preprint | Code (GitHub) | Demo (HuggingFace)

FrameDiff: An SE(3)-equivariant diffusion model for protein backbone generation, developed without relying on a pretrained structure-prediction network unlike earlier diffusion-based design methods. FrameDiff formalizes SE(3)-invariant diffusion over rigid-body frames and can generate designable monomers up to 500 residues; it was later adapted into the flow-matching model FrameFlow.
Preprint posted: Feb 05, 2023
Preprint

Subtimizer: A computational workflow for structure-guided design of potent and selective kinase peptide substrates. It integrates AlphaFold-Multimer for structural modeling, ProteinMPNN for sequence design, and AlphaFold2-based interface evaluation into an end-to-end pipeline, automating SLURM job submission, sequence analysis, clustering, and validation for high-throughput substrate engineering on HPC systems.
Paper published: February 7, 2026
Paper | Code (GitHub) | Documentation

CleaveNet: CleaveNet is an end-to-end AI pipeline that leverages deep learning for the design of protease substrates. The model generates peptide substrates and incorporates a conditioning tag to steer generation towards desired cleavage profiles. It can design efficient and selective substrates, revealing both well-established and previously uncharacterized cleavage motifs. Applied to matrix metalloproteinases, CleaveNet enhanced the scale, tunability, and efficiency of substrate design. The CleaveNet-generated substrates were experimentally validated through a large-scale in vitro screen, including designing highly selective substrates for MMP13.
Paper published: November 15, 2025
Paper

RFpeptides: A denoising diffusion-based pipeline from the Baker, Bhardwaj, and DiMaio labs for de novo design of macrocyclic peptide binders to protein targets, built on RoseTTAFold2 and RFdiffusion. Macrocycle backbones are generated from randomly initialized atoms via stepwise RFdiffusion-based denoising, followed by fixed-backbone sequence design with ProteinMPNN. Testing 20 or fewer designs against each of four diverse protein targets yielded medium-to-high-affinity binders against all four, with atomic-level structural accuracy confirmed for select designs.
Preprint posted: November 18, 2024
Paper published: October 15, 2025
Paper | Preprint | Code (GitHub)

D-Flow: A full-atom flow-matching framework for de novo D-peptide design (mirror-image peptides resistant to proteolytic degradation), conditioned on receptor binding and incorporating backbone frames, side-chain angles, and discrete amino acid types. D-Flow uses a mirror-image algorithm to leverage existing L-protein training data by converting the chirality of L-receptors, and integrates protein language model embeddings with structural awareness via a lightweight structural adapter, trained in two stages moving from general design to targeted binder design. On the PepMerge benchmark, D-Flow improved sequence identity by 10.2% over the best baseline, with a top affinity score of 24.31%.
Preprint posted: November 15, 2024
Preprint | Code (GitHub)

PepFlow: A method to predict structures of peptides, including single-state and ensembles of conformations. PepFlow is also capable of generating structures of macrocycles.
Paper published: Jun 27, 2024
Paper | Preprint | Code (GitLab)

EvoBind2: A method for designing novel linear and cyclic protein-binding peptides using just the amino acid sequence of the target protein as input. EvoBind2, similar to EvoBind (v1), presents as an in silico directed evolution method using a strategy analogous to the "diversify-screen/select(-amplify)" cycle of experimental directed evolution.
Preprint posted: Jun 22, 2024
Preprint | Code (GitHub) | EvoBind v1 Paper | EvoBind v1 Code (GitHub) | EvoBind v1 Colab Notebook

HELM-GPT: A generative model for de novo design of macrocyclic peptides.
Paper published: Jun 12, 2024
Paper | Code (GitHub)

AfCycDesign: An AlphaFold2-based approach for accurate structure prediction and design of cyclic peptides. AfCycDesign is capable of fixed backbone and de novo design of cyclic peptides, outperforming Rosetta on cyclic peptide design tasks.
Preprint posted: February 25, 2023
Paper published: May 21, 2025
Paper | Preprint | Colab Notebook

NA-MPNN: A message-passing neural network from the Baker lab that extends ProteinMPNN to treat proteins, DNA, and RNA within a unified biopolymer graph representation. NA-MPNN outperforms previous methods on RNA sequence design and fixed-dock protein–DNA specificity prediction, and is intended to be broadly useful for de novo RNA structure design and prediction of DNA-binding specificity.
Preprint posted: October 4, 2025
Preprint | Code (GitHub)

RNAFlow: A flow-matching model for simultaneous design of RNA structure and sequence conditioned on protein structure and sequence. The denoising network of RNAFlow comprises: 1. an RNA inverse folding model, and 2. a pre-trained RosettaFold2NA network. In each design cycle, RNAFlow first designs a RNA sequence for a given noised protein-RNA complex, and then uses RosettaFold2NA to predict a denoised RNA structure.
Preprint posted: May 29, 2024
Preprint | Code (GitHub)

gRNAde: The “ProteinMPNN for RNA sequence design”. gRNAde is a geometric DL-based model for computational design of RNA sequences given 3D RNA backbone structures. gRNAde enables both single- and multi-state fixed-bb sequence design by generating candidates RNA sequences conditioned on one or more bb 3D conformations.
Preprint posted: April 01, 2024
Preprint | Code (GitHub) | Colab Notebook 1 | Colab Notebook 2

Evo: A long-context foundation model that generalizes across the central dogma of biology: DNA, RNA, and proteins. Evo is a 7 billion parameter model trained to generate DNA sequences and is capable of prediction and generative tasks, from molecules to whole genomes.
Preprint posted: March 06, 2024
Preprint | Code (GitHub) | Code (PyPI) | Blog | Playground | Colab Notebook

GenerRNA: A generative pre-trained language model for de novo RNA design.
Preprint posted: Feb 08, 2024
Preprint | Code (GitHub)

RfamGen: A generative model for designing functional RNA family sequences. RfamGen incorporates alignment and consensus secondary structure information and generates novel and functional RNA family sequences.
Paper published: Jan 18, 2024
Paper | Code (GitHub)

Evo (Semantic Design / Evo 1.5): A follow-up work to the original Evo genomic language model, showing that Evo can leverage genomic context to perform function-guided “semantic design” — a DNA prompt encoding genomic context for a function of interest guides generation of novel sequences enriched for related functions. The approach was experimentally validated by testing generated anti-CRISPR proteins and type II/III toxin–antitoxin systems, including de novo genes with no significant sequence similarity to natural proteins, achieving robust activity and high experimental success rates. The authors also released SynGenome, a database of over 120 billion bp of Evo-generated DNA sequences spanning prompts across 9,000 functional terms.
Paper published: January 15, 2026
Paper | Code (GitHub) | Model (HuggingFace) | SynGenome (Database)

Evo: A long-context foundation model that generalizes across the central dogma of biology: DNA, RNA, and proteins. Evo is a 7 billion parameter model trained to generate DNA sequences and is capable of prediction and generative tasks, from molecules to whole genomes.
Preprint posted: March 06, 2024
Preprint | Code (GitHub) | Code (PyPI) | Blog | Playground | Colab Notebook

DNA-Diffusion: A diffusion model to generate context-/cell type-specific DNA regulatory sequences.
Preprint posted: Feb 01, 2024
Preprint | Code (GitHub) | Documentation

EvoStruct: A method addressing vocabulary collapse in equivariant GNNs for Ab complementarity-determining region (CDR) design. It integrates a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN using a cross-attention adapter. The method employs progressive PLM unfreezing and R-Drop consistency regularization to improve amino acid diversity. On the CHIMERA-Bench dataset, EvoStruct achieves higher amino acid recovery and diversity, improving sequence recovery by 16% and reducing perplexity by 43% compared to GNN baselines.
Preprint posted: May 21, 2026
Paper | Preprint

AgForce: An antigen-conditioned generative antibody design model that diagnoses and addresses systematic failures in prior CDR design methods. A benchmark of 11 existing CDR-H3 design methods on CHIMERA-Bench revealed three failure modes — antigen blindness (near-identical CDRs regardless of target), vocabulary collapse (only 3–5 amino acids predicted per position), and a cross-entropy training ceiling that provably prevents antigen-specific predictions. AgForce addresses these with a GNN encoder and specialized sequence-structure co-design decoders, applying framework dropout, gated bottlenecks, hyperbolic cross-attention, a Mixture Density Network sequence head with Potts-like pairwise coupling, annealed Multiple Choice Learning, and an antigen cycle-consistency head.
Preprint posted: May 20, 2026
Preprint | Code (GitHub)

DualGPT-AB: A dual-stage conditional generative pre-trained transformer (GPT) framework for therapeutic antibody design that simultaneously optimizes antigen-binding specificity and developability. It models sequence-property relationships through learnable property embeddings across two stages — Prior GPT Training and reinforcement-learning-driven Enhanced GPT Training — producing antibody candidates with enhanced tumoricidal activity compared with current therapies.
Paper published: April 15, 2026
Paper | Code (GitHub)

Chai-2: A multimodal generative model from Chai Discovery for fully de novo antibody design, achieving a 16% hit rate in generating functional binders across 52 novel targets using 20 or fewer computationally designed candidates per target — over 100-fold higher than prior computational methods — eliminating the need for large-scale experimental screening. Designed antibodies showed nanomolar-range affinities, target specificity, and drug-like developability profiles. A follow-up study extended the approach to design full-length, drug-like monoclonal antibodies against challenging targets with atomic precision.
Preprint posted: July 6, 2025
Preprint | Follow-up Preprint

RFantibody: RFantibody combines a version of RFdiffusion fine-tuned on antibody complex structures with ProteinMPNN for CDR sequence design and a separately fine-tuned RoseTTAFold2 network for in silico filtering, enabling de novo, epitope-specific design of antibody VHHs, scFvs, and full antibodies entirely in silico. Designed VHH binders were confirmed by SPR against four disease-relevant epitopes — RSV site III, influenza hemagglutinin, SARS-CoV-2 RBD, and C. difficile toxin B — with affinities of 1.4 µM, 78 nM, 5.5 µM and 260 nM respectively, out of libraries screened by yeast display and/or E. coli expression. Cryo-EM confirmed the binding pose of the designed anti-hemagglutinin VHH (1.45 Å backbone RMSD, 0.8 Å CDR3 RMSD to the design model) and the designed anti-toxin-B VHH, and further confirmed the fold and CDR loop conformations of a fully de novo, six-CDR-designed scFv against toxin B (72 nM affinity). OrthoRep-based continuous hypermutation and selection affinity-matured designed VHHs by roughly two orders of magnitude to single-digit nanomolar affinity while preserving the originally designed epitope and binding pose.
Preprint posted: March 14, 2024
Paper published: November 5, 2025
Paper | Preprint | Code (GitHub) | Preprint (v1)

GeoAB: A method for computational design and optimization (affinity maturation) of antibody. GeoAB utilizes a co-design strategy, predicting the structure of a CDR and optimized 1D sequences for structure.
Preprint posted: May 17, 2024
Preprint | Code (GitHub)

IgDiff: A model for de novo antibody design (structure generation) that adapted FrameDiff (a diffusion-based model for protein backbone generation) by fine-tuning it on artificial antibody structures.
Preprint posted: May 13, 2024
Preprint | Code (Zenodo)

AntiFold: An inverse folding model for antibody sequence design based on ESM-IF1.
Preprint posted: May 06, 2024
Preprint | Code (GitHub) | Webserver | Colab Notebook

tFold-Ab and tFold-Ag: Methods for antibody and antibody-antigen complex modelling and design by Tencent.
Preprint posted: Feb 08, 2024
Preprint | Code (GitHub)

IgDesign: IgDesign is a deep learning method for antibody CDR design via inverse folding, combining ideas from ProteinMPNN and LM-Design's structure-encoder/sequence-decoder framework: IgMPNN (a ProteinMPNN variant fine-tuned for antibodies) encodes the antibody-antigen complex structure, and its embeddings are cross-attended with an ESM2-3B language model before decoding CDR sequences, using the antigen and antibody framework sequences as added context. It is the first antibody inverse-folding model validated in vitro for binder design, tasked with designing either CDR-H3 alone or all three heavy-chain CDRs (HCDR123) from a reference antibody-antigen complex structure. Across 8 therapeutic antigens, IgDesign screened 1,437 designed antibodies by surface plasmon resonance and identified 278 binders; HCDR3 designs bound significantly more often than an SAbDab-sampled baseline for 7 of 8 antigens, and HCDR123 designs did so for 4 of 8. For 5 of the 8 antigens, IgDesign generated binders with equal or higher affinity than the reference antibody.
Preprint posted: December 8, 2023
Preprint | Code (GitHub)

Subtimizer: A computational workflow for structure-guided design of potent and selective kinase peptide substrates. It integrates AlphaFold-Multimer for structural modeling, ProteinMPNN for sequence design, and AlphaFold2-based interface evaluation into an end-to-end pipeline, automating SLURM job submission, sequence analysis, clustering, and validation for high-throughput substrate engineering on HPC systems.
Paper published: February 7, 2026
Paper | Code (GitHub) | Documentation

Ovo: An open-source de novo protein design ecosystem that consolidates state-of-the-art design models, workflows, data management, and interactive visualization into a single, infrastructure-agnostic platform. Ovo features Nextflow-based workflow orchestration, a SQL-based storage layer, and both command-line and Streamlit-based graphical interfaces that democratize scaffold design, binder design and diversification, and validation workflows. Its novel ProteinQC module computes comprehensive sequence and structure descriptors, contextualizing designs against reference sets, and a plugin system lets the community add new workflows and user interfaces to accelerate adoption of emerging methods and community-driven benchmarking.
Preprint posted: November 28, 2025
Preprint | Code (GitHub) | Documentation | Demo (Web App)

BinderFlow: An open, structured, and parallelized pipeline that automates end-to-end protein binder design, built for SLURM-based HPC clusters. Its batch-based architecture enables live monitoring of design campaigns, coexistence with other GPU-intensive jobs, and minimal user intervention, with a modular design that facilitates integrating new tools as methods emerge. The accompanying BFmonitor is a web-based dashboard for real-time campaign monitoring, design evaluation, and hit selection.
Paper published: November 21, 2025
Paper | Preprint | Code (GitHub)

EvoPro: A protocol paper detailing the Kuhlman lab’s EvoPro platform, which uses a genetic algorithm together with iterative structure prediction (AlphaFold2/AlphaFold3) and sequence design (ProteinMPNN/LigandMPNN) to engineer protein–protein interactions with customizable properties — a computational adaptation of directed evolution, cyclically designing and re-scoring candidates via structure prediction. The protocol describes multistate design objectives (simultaneous positive and negative design goals, e.g. selectivity for one target over a homolog) and its adaptation to binding site targeting, conformational-switch engineering, and de novo symmetric homo-oligomer design.
Preprint posted: October 2025
Protocol Paper | Code (GitHub) | Original EvoPro Paper (2023)

ProtDAT: A unified framework for ab-initio amino acid sequence design directly from free-form protein text descriptions, optionally combined with sequence fragments. ProtDAT combines an ESM1b protein encoder, PubMedBERT text embeddings, and a decoder to generate de novo sequences conditioned jointly on natural-language description and partial sequence context.
Paper published: November 26, 2025
Paper | Code (GitHub)

ProDVa: A protein language model that augments de novo generation with a dynamically retrieved “protein vocabulary” of natural sequence fragments to improve foldability, rather than generating purely de novo. It combines a GPT-2-based text encoder, a ProtGPT2-initialized decoder, and a fragment encoder that retrieves functionally relevant natural fragments (via PubMedBERT description similarity and InterPro annotations) at both training and inference time. Trained on only ~712K description–sequence pairs (under 0.04% of the data used by the comparison baseline Pinal), ProDVa achieves comparable function alignment while substantially improving structural plausibility (pLDDT>70 rate of 76.35% vs. 68.97%; PAE<10 rate of 68.06% vs. 58.44%). Accepted as a Spotlight paper at NeurIPS 2025.
Preprint posted: October 14, 2025
Preprint | Code (GitHub)

MP4: A transformer-based text-to-protein model from 310 AI that generates de novo protein sequences directly from free-form natural-language functional descriptions, rather than fixed backbone structures or predefined ontology terms. Trained on a large proprietary dataset (~3.2 billion data points), MP4 improved sequence realism, predicted fold quality, and functional alignment across 96 diverse functional prompts relative to prior approaches. In wet-lab validation, 84% of 94 selected designs expressed successfully in a cell-free system (versus a typical 20–30% for non-AI baselines), with average melting temperatures exceeding 62°C; a broader generation run produced 6,590 proteins spanning the full top-level EC enzyme class space, including 725 flagged as novel.
Preprint posted: August 18, 2025
Preprint

BioM3: A three-stage pipeline for designing functional proteins from natural-language prompts: PenCL aligns protein-sequence and text embeddings via contrastive learning, a Facilitator network refines text embeddings into the protein representation space, and ProteoScribe, an order-agnostic autoregressive discrete diffusion model, generates the final sequence. Trained on ~45 million text–sequence pairs from SwissProt and Pfam, BioM3 achieves state-of-the-art zero-shot prediction and homology-detection performance and generates proteins with native-like folds; in vivo/in vitro validation of BioM3-designed synthetic analogs of the yeast Sho1 SH3 signaling domain showed wild-type-level function, reported as the first experimentally validated functional proteins designed purely from text prompts.
Preprint posted: November 11, 2024
Preprint | Code (GitHub) | Demo (HuggingFace)

Pinal: A large-scale (up to ~16B-parameter) two-stage generative framework for de novo protein design from natural language, factoring generation through an intermediate structure step: T2Struct maps a text description to a protein backbone structure encoded as discrete VQ tokens, and SaProt-T then generates the amino-acid sequence conditioned jointly on the generated structure and the original text. Trained on ~1.7 billion protein–text pairs, Pinal was experimentally validated across four distinct functional families designed purely from language instructions — a GFP, a PETase, an alcohol dehydrogenase, and a metabolic H-protein — with confirmed activity in all four; the designed H-protein showed a 1.7-fold performance improvement over its natural counterpart.
Preprint posted: August 2, 2024
Preprint | Code (GitHub) | Webserver

ProteinDT: A multi-modal framework coupling natural-language text with protein sequence design via three stages: ProteinCLAP, a CLIP-style contrastive language–protein pretraining step; a facilitator network mapping text embeddings into protein representation space; and an autoregressive decoder generating the final sequence. Trained on SwissProtCLAP, a newly constructed dataset of 441,000 text–protein pairs, ProteinDT achieves over 90% retrieval accuracy for text-guided generation, the best hit ratio among baselines across 12 zero-shot text-guided protein editing tasks, and superior performance on 4 of 6 protein property-prediction benchmarks.
Preprint posted: February 9, 2023
Paper published: March 27, 2025
Paper | Preprint | Code (GitHub)