Peldom/papers_for_protein_design_using_DL: A Categorised Reading List, Not a Tool
List of papers about Proteins Design using Deep Learning
At a glance
- What is it?
- This repository is a curated bibliography of deep learning papers for protein design, organised by task family and by model architecture. It solves a discovery problem for researchers, not an engineering one, and its value depends entirely on how the list is maintained.
- Who is it for?
- Adopt this list if you are entering deep learning protein design and need a structured starting point rather than a keyword search, or if you want a single place to check what appeared in the last few weeks. Do not adopt it if you need executable code, benchmark numbers, or per-paper quality judgements, because the repository provides none of those.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Yes. The repository last received commits 32 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem is discovery, not computation
Deep learning for protein design spans at least four distinct problem framings: generating a scaffold from a desired function, generating a sequence from a fixed scaffold, generating a sequence directly from a function, and generating a structure from a function. Papers addressing these framings appear in machine learning venues, structural biology journals, and preprint servers, and they rarely cite each other across those boundaries. A researcher who wants to know what has been tried for binder design has to search several literatures and reconcile incompatible vocabulary.
This repository addresses that by acting as a single index. It is aimed at graduate students, postdocs, and research engineers entering the area, plus anyone who wants a periodic digest of new work. It is not aimed at someone who needs to run a model. There is no installable package, no inference code, and no benchmark harness. The repository name says papers, and that is literally what it holds.
The README credits Kevin Kaichuang Yang's Machine-learning-for-proteins as the inspiration and describes this repository as a specialised and focused platform for deep learning for protein design. That framing is accurate: the scope is narrower than the parent project, which matters if you work on protein property prediction or folding in general rather than design specifically.
The taxonomy is the actual content
The value here is not the paper titles. It is the classification scheme, which the README exposes as a menu of anchor links. Two organising principles run in parallel.
The first is the generator-predictor-optimizer paradigm, which the README attributes to section 2, Model-based design. The second is the inside-out paradigm from RosettaCommons, described in the README as function-scaffold-sequence, which governs sections 3, 4, and 6. Sections 5 and 7 are described as following other ML and DL strategies. That is a deliberate mixing of two incompatible axes, and the README says so rather than hiding it.
The architecture axis is where the list gets granular. Section 4, Scaffold to Sequence, is subdivided into MLP-based, VAE-based, LSTM-based, CNN-based, GNN-based, GAN-based, Transformer-based, ResNet-based, Diffusion-based, Bayesian method, Flow-based, RL-based, and a Train method category. Section 5, Function to Sequence, is split even further, adding RNN-based, Autoregressive, Boltzmann machine, Score-based, and GNN-based. That granularity is useful when you are trying to find prior art for a specific architecture choice, and it is the main reason to prefer this list over a general search.
Section 0 collects benchmarks and datasets, split into sequence datasets and benchmarks, structure datasets and benchmarks, public databases, similar lists, and guides. Section 1 collects reviews and surveys, split by application area: de novo design, antibody design, peptide design, binder design, enzyme design. Section 7 covers effects of mutations and fitness landscape, protein language models and representation learning, molecular design models, frameworks, and an unclassified bucket.
Reading the list, and the papers-last-week block
There is nothing to install. The repository is consumed through its README on GitHub, and every entry is a link to a paper page, a preprint, or occasionally a code repository and project website.
The one structural element worth understanding is the papers-last-week block near the top. The README shows it with a date, and the example in the supplied material is dated 2026.08.02 with the heading Papers last week, updated on 2026.08.02. Entries in that block follow a two-level bullet format: a paper title, then an indented link to the venue or preprint server, sometimes with additional code and website links. The example entries are Generative artificial intelligence for enzyme design and biocatalysis from Current Opinion in Chemical Biology, and The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates, which carries a bioRxiv identifier plus a GitHub code link and a website link.
That block is the only part of the repository with a timestamp attached to its content, and the timestamp is maintained by hand. If the date is stale, the digest function has lapsed even if the rest of the list is intact. This is the first thing to check before treating the repository as an active source.
Contribution is handled through a CONTRIBUTING.md file in the repository root, and suggestions go through the GitHub issues page. The README also points to responsiblebiodesign.ai for community values and guiding principles on responsible AI development for protein design. Neither the contribution format nor the review process for submitted papers is described in the material available, so it is not possible to say how entries are vetted.
What a static list cannot do
The main limitation is that the repository carries no assessment of the papers it lists. There are no benchmark comparisons, no notes on whether a method was reproduced, no indication of which entries have been superseded. A reader who finds three diffusion-based entries in section 3 has no signal from the repository about which one to read first.
The README does point to a Zhihu column described as containing notes on these papers in simplified Chinese and English, and to a separate column called RosettAI for further suggested notes. Those links are commented out in the README markup, which means they are not visible on the rendered page. If you want commentary rather than titles, you have to read the raw README to find the pointers.
The second limitation is scope drift. The README notes that more de novo protein design papers are collected in a separate repository by Wangchentong, and that mini protein, binders, metalloprotein, antibody, peptide and molecule designs are included here. That inclusion list is broad, and the unclassified bucket in section 7 is an admission that the taxonomy does not always fit. Papers that straddle two categories can end up in the wrong one, and there is no cross-referencing mechanism visible in the material.
The third limitation is the wrong-tool case. If you need to reproduce a result, evaluate a model on your own sequences, or compare throughput across methods, this repository gives you nothing to run. It is a bibliography. Treating it as a survey with conclusions would be a category error.
How it differs from a general protein ML list
The obvious comparison is the repository this one credits: Kevin Kaichuang Yang's Machine-learning-for-proteins. The difference is scope, and it cuts both ways.
The parent list covers machine learning for proteins broadly, which includes folding, property prediction, and representation learning alongside design. This repository restricts itself to design and then subdivides aggressively within that restriction. The result is that a reader looking for scaffold-to-sequence methods gets a thirteen-way architecture split here, whereas a broader list would likely place those papers in one or two categories.
That trade-off has a cost. A paper that advances protein language models without addressing design directly may be listed here under section 7.2 but would sit in a more prominent position in a general list. Conversely, a design paper that relies on a folding model as a component will be indexed here by its design contribution, and the folding work it builds on may not appear at all.
If your question is what architectures have been applied to a specific design subproblem, this list is the better starting point. If your question is how design relates to the wider protein ML literature, the parent list is. The README links to both, plus a third list focused specifically on de novo design, which suggests the maintainers see these as complementary rather than competing.
Licence and maintenance cost
The repository is licensed GPL-3.0. That is unusual for a paper list, where CC0 or CC-BY is more common, and it has practical consequences worth understanding before you reuse the content. GPL-3.0 is a copyleft software licence, and applying it to a collection of links and titles creates ambiguity about what the licence actually covers: the curation effort, the README text, or the linked papers, which have their own separate terms. The linked papers are not covered by this licence, and the README does not attempt to clarify the boundary. This is a description of the licence as stated in the repository metadata, not legal advice; if you plan to mirror the list or embed it in a product, get your own reading.
Maintenance cost falls on whoever keeps the list current. The work is manual: reading new preprints, deciding which section an entry belongs in, and updating the papers-last-week date. There are no releases retrieved for this repository, so there is no versioned snapshot to pin against. The last push date in the repository metadata is 2026-08-15, which is consistent with an actively edited list, but a push date tells you nothing about whether the content in a given section was reviewed.
The upgrade path is the same as the contribution path: open a pull request following CONTRIBUTING.md, or file an issue. There is no migration concern, because there is nothing to migrate. If you fork the list to add your own annotations, you inherit the GPL-3.0 terms on your fork.
Editorial conclusion
Adopt this list if you are entering deep learning protein design and need a structured starting point rather than a keyword search, or if you want a single place to check what appeared in the last few weeks. Do not adopt it if you need executable code, benchmark numbers, or per-paper quality judgements, because the repository provides none of those. Before relying on it, open the README and check two things: whether the papers-last-week block has a recent date, and whether the section you care about lists methods that have since been superseded by newer entries in the same section.
Community notes