De Novo Protein Design: From Physics to Deep Learning

A review built from a curated 507-paper corpus (Europe PMC, 2026). The full corpus — every paper with its abstract and method/algorithm tag — is a searchable table in the ELN: De Novo Protein Design — Literature (workspace members; sign in required). Prepared with the NextELN AI Lab Assistant; scientifically reviewed by the NextELN science team, Excellgen, Inc.

What "de novo" means

De novo protein design builds proteins that do not exist in nature — new backbones and new sequences — to perform a specified function: bind a target, catalyze a reaction, assemble into a material. It is the inverse of the folding problem: instead of sequence → structure, we ask desired structure/function → sequence. The field has moved through five overlapping eras, and today converges on a single, powerful pipeline.

Era 1 — Physics-based sequence selection (1997–2018)

The founding idea: fix a target backbone, then search sequence space for the amino acids that minimize a physics-based energy. Dahiyat & Mayo (1997) achieved the first fully automated sequence selection for a target fold; DeGrado and colleagues designed and characterized novel proteins and metalloproteins, establishing that rational design of new folds was possible at all. This matured into the Rosetta framework — a physics/statistics energy function plus Monte-Carlo sequence/structure sampling — and a set of "design rules" for ideal, hyperstable folds (Huang, Boyken & Baker, 2016, The coming of age of de novo protein design). Rosetta also delivered the first de novo enzymes by grafting idealized active sites into designed scaffolds (Röthlisberger 2008, Kemp eliminases; Kiss 2013, review), and massively parallel target-binder design screened by high-throughput synthesis (Chevalier 2017).

Strengths: interpretable, no training data. Limits: energy functions are approximate, success rates were low, and sampling is expensive.

Era 2 — Deep-network hallucination (2020–2021)

The first deep-learning wave inverted a structure predictor. Given a network that maps sequence → structure (trRosetta), hallucination (Anishchenko 2021, De novo protein design by deep network hallucination) optimizes a random sequence so that the predictor becomes confident it folds into a well-defined (if unspecified) structure — "activation maximization" on the folding network. Add a constraint (a binding motif, a symmetry) and you steer the hallucination toward function. This showed a learned folding model already contains a designable energy landscape.

Era 3 — Inverse folding / sequence design: ProteinMPNN (2022)

If Era 2 dreamed up backbones, Era 3 solved the complementary problem definitively: given a fixed backbone, what sequence folds into it? ProteinMPNN (Dauparas 2022) — a message-passing graph neural network over the backbone — designs sequences with far higher experimental success and speed than Rosetta, and a soluble-trained variant biases toward expressible proteins. ProteinMPNN became the default "sequence-design" step for every generative backbone method and is the workhorse of the modern pipeline. (This is the exact tool we use in our own protein-engineering work.)

Era 4 — Generative backbones: diffusion (2023)

The breakthrough that made de novo design broadly practical was generative diffusion over protein backbones. RFdiffusion (Watson 2023) fine-tunes the RoseTTAFold structure network as a denoising diffusion model: start from noise, iteratively denoise into a novel backbone, conditioned on a target (a binding site, a symmetry, a scaffold around a motif). It produces experimentally validated binders, symmetric assemblies, and motif scaffolds at unprecedented rates. Chroma (Ingraham 2023, Illuminating protein space with a programmable generative model) is a parallel programmable diffusion model with conditioning on shape, symmetry, and text-like constraints. Newer iterations (RFdiffusion2/3, all-atom and flow-matching variants) extend to ligands and nucleic acids.

The dominant modern recipe is now three steps:

RFdiffusion (backbone) → ProteinMPNN (sequence) → AlphaFold2/ESMFold (in-silico validation).

Only designs whose predicted structure self-consistently matches the intended backbone (low RMSD, high pLDDT/pAE) advance to the bench.

Era 5 — Protein language models (2022– )

In parallel, autoregressive protein language models learned to write protein sequences directly, no explicit structure step. ProtGPT2 (Ferruz 2022) and ProGen (Madani 2023, Large language models generate functional protein sequences across diverse families) generate novel, expressible, sometimes catalytically active proteins by learning the "grammar" of natural sequences, and can be prompted or fine-tuned toward a family or function. Structure-aware language models (ESM-IF for inverse folding; ESMFold for validation) blur the line between the language-model and structure-based branches.

Applications that now work

• Binders / therapeutics: de novo minibinders designed from a target structure alone (Cao 2022), improved by deep learning (Bennett 2023), and de novo antibodies with RFdiffusion.
• Enzymes: from Rosetta grafting to deep-learning designed luciferases (Yeh 2023) and retro-aldolases; active-site scaffolding is now diffusion-conditioned.
• Assemblies / materials: symmetric nanocages, filaments, and shape-programmable assemblies.

How we implement it (NextELN / Excellgen)

Our approach is pipeline-first and evidence-gated, mirroring the field's convergence but tuned for practical, expressible proteins:

1. Generate — conditioned backbones (diffusion) for the target geometry / motif, or start from a natural scaffold for redesign.
2. Design sequences — ProteinMPNN (soluble model), with interface-aware position masking so catalytic and partner-binding residues are never touched (as in our recombinase-solubility work).
3. Validate in-silico — AlphaFold/ESMFold self-consistency (pLDDT ≥ reference, low RMSD/pAE), plus a validated solubility/aggregation oracle and codon-optimization for the intended host — before any $1000s/30-day expression trial.
4. Learn — every wet-lab result updates the ranking models. The moat is the accumulated prediction-to-experiment record, not any single model.

Our own ML work focuses on the parts the public tools do not: a calibrated expressibility/solubility model trained on real expression outcomes, interface- and function-preserving design constraints, and host-specific optimization — the difference between "a design that folds in-silico" and "a protein you can actually make."

Reference table (selected; full 507-paper corpus in the ELN workspace)

Year	First author	Contribution	Key method / algorithm
1997	Dahiyat	First fully automated de novo sequence selection	Physics-based dead-end elimination over a fixed backbone
2008	Röthlisberger	First de novo enzymes (Kemp eliminases)	Rosetta active-site theozyme grafting
2016	Huang	Field-defining review ("coming of age")	Rosetta rules for ideal, hyperstable folds
2017	Chevalier	Massively parallel target-binder design	Rosetta + high-throughput synthesis screening
2021	Anishchenko	De novo design by hallucination	Activation maximization on trRosetta
2022	Dauparas	Robust sequence design (ProteinMPNN)	Message-passing GNN inverse folding; soluble variant
2022	Cao	Binders from the target structure alone	Rosetta/deep-learning target-conditioned design
2022	Ferruz	ProtGPT2 generative language model	Autoregressive transformer on protein sequences
2023	Watson	RFdiffusion — structure & function	Denoising diffusion over backbones (RoseTTAFold fine-tune)
2023	Ingraham	Chroma — programmable generative model	Diffusion with shape/symmetry conditioning
2023	Madani	ProGen — functional sequences across families	Large autoregressive protein language model
2023	Yeh	De novo luciferases	Deep-learning active-site scaffolding
2023	Bennett	Improved binder design	RFdiffusion + ProteinMPNN + AF filtering

The complete, searchable corpus — 507 papers with abstracts and per-paper method/algorithm tags across nine method families (general design, inverse folding, enzyme design, physics/Rosetta, language models, diffusion, binders, assemblies, hallucination) — is maintained as a live ELN table: De Novo Protein Design — Literature ↗ (sign in required).

Status: research review. De novo designs are computational hypotheses; expressibility, activity, and safety require experimental validation.