Sign in

Lars Dornfeld

@ldornfeld.bsky.social
505 followers 235 following 35 posts

PhD student in Milles Lab | Passionate about protein design | 🥼&💻 | Formerly with Sereina Riniker (MD), Noelia Ferruz (ML), Bruno Correia (Protein Design) & Kai Johnsson (Protein Engineering) Webpage: ldornfeld.github.io

PostsRepliesMedia
Reposted by Lars Dornfeld
Lukas Milles @lfmilles.bsky.social · 07/10/2026
First preprint from the lab, design & control your protein's amino acid, mRNA and DNA sequences all at once; driven by @blazejba.bsky.social and @ldornfeld.bsky.social, in great collaboration with Dexiong Chen and Karsten Borgwardt.
0122
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
I really loved the dynamics in this project. Huge thanks to the whole team: @blazejba.bsky.social, Dexiong Chen, @lfmilles.bsky.social and Karsten Borgwardt. We would love to hear feedback and ideas from the community on further applications and extensions of this work. 11/11
020
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Looking forward, we are testing functional RNA elements within functional proteins. Very excited to see where this goes. 10/11
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Taking the Nucleotide sequences to the lab, we obtained monodisperse, soluble proteins comparable to ProteinMPNN designs with post-hoc codon optimization. We could also embed a 29-nt RNA motif into most backbones while retaining protein solubility. 9/11
Experimental protein expression and solubility across six design approaches. Total soluble yield is shown on the x-axis, while point color indicates the estimated molecular weight of the main SEC peak relative to the expected monomer. Most NuCaliby designs are soluble and monomeric, including nucleotide-level designs with tRNA guidance and RNA motif incorporation. Their yields are comparable to amino acid-level NuCaliby and ProteinMPNN designs followed by post-hoc codon optimization. CodonMPNN generally shows lower yields.
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Second, we optimize codon choice for host-specific tRNA availability. Going beyond synonymous optimization, tRNA adaptation can be further improved by trading off amino acid selection and self-consistency. 8/11
Effect of increasing tRNA adaptation guidance on protein designability and codon usage. Stronger guidance increases the tRNA adaptation index (tAI) beyond synonymous optimization, while TM-score decreases and pLDDT remains relatively stable. At high guidance strengths, amino acid distributions increasingly diverge from unguided sampling, indicating non-synonymous adaptation. Codon usage also shifts progressively toward codons with the highest tAI rank.
120
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
We use this for a few cool things. First, we directly incorporate RNA sequence motifs, including aptamers and the HIV-1 frameshifting element, into the coding sequences of different protein backbones. The model jointly optimizes the sequence and where to place the motif. 7/11
Illustration of mRNA motif guidance. A target nucleotide motif is compared against sliding windows along a candidate coding sequence. The Hamming distance between the motif and each window is calculated, producing a distance landscape across sequence positions. Guidance favors sequence changes that reduce this distance, allowing the target motif to be embedded directly into the coding sequence.
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Sequences are sampled from the Potts model using gradient-guided Monte Carlo, allowing us to introduce different biological proxies at inference time and directly balance them against protein design-ability. 6/11
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
To solve this, we build on neural-network-derived Potts models such as Caliby, but construct and sample the Potts model directly at nucleotide-level resolution. We achieve AlphaFold2 self-consistency comparable to amino acid-level models such as ProteinMPNN and Caliby. 5/11
Comparison of conventional sequential protein design with NuCaliby. In the sequential approach, an amino acid sequence is first designed from a protein backbone, after which the nucleotide sequence is optimized only within synonymous codons. NuCaliby instead designs directly at the nucleotide level, allowing amino acid and nucleotide constraints to be considered jointly and enabling both synonymous and non-synonymous sequence changes. The codon table at the top illustrates the mapping of 61 coding codons to 20 amino acids.
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
This is insufficient for tasks that require encoding additional functionality directly into the coding sequence or balancing amino acid and nucleotide-level properties. 4/11
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Most inverse folding approaches first design an amino acid sequence and then optimize its coding sequence synonymously. This sequential process progressively reduces the available degrees of freedom from protein structure → amino acid sequence → DNA. 3/11
Illustration of the increasing sequence space from protein structure to amino acid sequences to nucleotide coding sequences. A single protein fold can be encoded by many amino acid sequences with the same topology, each of which can in turn be encoded by many nucleotide sequences due to synonymous codon choices. The nucleotide sequence space therefore provides additional degrees of freedom for sequence design and optimization.
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Evolution selects mutations at the DNA level, with consequences across the central dogma. Can we design proteins the same way?... 2/11
110
Lars Dornfeld @ldornfeld.bsky.social · 07/10/2026
Our paper, “Joint protein, mRNA and DNA sequence design and optimization with nucleotide-level Potts models,” has been accepted at NeurIPS 2026 and is now available on bioRxiv! Preprint: www.biorxiv.org/content/10.6... Code: github.com/blazejba/NuC... 1/11
biorxiv.org
Joint protein, mRNA and DNA sequence design and optimization with nucleotide-level Potts models
Protein design typically generates amino acid sequences first and then translates them into DNA post-hoc limiting control over the nucleotide sequence via codon choice at design time. We introduce NuC...
1237
Reposted by Lars Dornfeld
Martin Pacesa @martinpacesa.bsky.social · 21/09/2026
ʙɪɴᴅᴄʀᴀꜰᴛ2 is out, and we're not waiting for the paper. The full code drops today, free for academic and industry use. We're releasing it early so you can start designing right now, and bring its full power to the current Adaptyv competition. github.com/PacesaLab/Bi...
4294142
Reposted by Lars Dornfeld
Lucas Diedrich @lucas-diedrich.bsky.social · 15/09/2026
Integrating spatial + multimodal data with proteomics was our key motivation to develop AlphaPeptTools. So happy to work with @brennsteiner.bsky.social, our co-authors + an amazing #scverse community! Let us know your feature requests and thoughts on making your next analysis easier :)
162
Reposted by Lars Dornfeld
Martin Steinegger 🇺🇦 @martinsteinegger.bsky.social · 30/08/2026
Run AlphaFold2 locally in your browser via WebGPU, no installation needed, it's just a static webpage using your own GPU or CPU. Short proteins run in seconds. Larger ones are still slow: a 291-aa protein takes ~7 minutes on my M4 Pro. ❗It's just a PoC. 🌐 martin-steinegger.github.io/alphafold2-w...
117741
Reposted by Lars Dornfeld
Ben Recht @beenwrekt.bsky.social · 30/08/2026
The overreaction to this Hugging Face thing is driving me nuts. Come flay me for being wrong but... 1/x
14436126
Reposted by Lars Dornfeld
Janani Durairaj (Jay) @ninjani.bsky.social · 28/08/2026
Some exciting news: I'm joining the University of Lausanne @unil.bsky.social @dbc-unil.bsky.social as an Assistant Professor next month 🎉 My group will work on context-aware deep learning for protein structure, interaction & design. Postdoc and PhD openings coming soon - keep an eye out!
compare, predict, design
412232
Reposted by Lars Dornfeld
Maik Wolfram-Schauerte @maiktungsten.bsky.social · 18/08/2026
🦠 How do phages interact with their bacterial hosts... across infection phases and across diverse phage-host combinations? You can address such questions with the PhageExpressionAtlas now published. Please spread the word and provide your feedback as a user: academic.oup.com/nargab/artic...
academic.oup.com
The PhageExpressionAtlas reveals shared and unique transcriptional patterns across phage–host interactions
Abstract. Time-resolved transcriptomic profiling has been used to study phage–host interactions for more than a decade. However, the resulting datasets are
196
Reposted by Lars Dornfeld
Bryan Dickinson @chembiobryan.bsky.social · 06/08/2026
1/ Binder ≠ inhibitor. Most PPI inhibitor campaigns find a binder first, then hope it blocks the target protein-protein interaction. We built a system that skips the hoping and selects directly for disruption of a pre-formed PPI. Meet PANCS-Inhibitors. 🧵 www.biorxiv.org/content/10.6...
biorxiv.org
PANCS-Inhibitors: A rapid method to directly select for protein-protein interaction inhibitors
Aberrant protein-protein interactions (PPIs) drive myriad diseases. Inhibiting these PPIs often relies on discovering molecules that bind to one of the proteins and hoping that this binding inhibits t...
12211
Reposted by Lars Dornfeld
Bastian Grossenbacher-Rieck @pseudomanifold.topology.rocks · 20/07/2026
New blog post, this time quite personal: "40 Bits of Advice That, in Murky Light, Might Pass as Wisdom" #AcademicBlogging bastian.rieck.me/blog/2026/40/
bastian.rieck.me
40 Bits of Advice That, in Murky Light, Might Pass as Wisdom
021
Reposted by Lars Dornfeld
Daniel Probst @skepteis.bsky.social · 29/06/2026
Great new work by @florisvdf.bsky.social!
051
Reposted by Lars Dornfeld
Martin Pacesa @martinpacesa.bsky.social · 15/06/2026
My favourite sequential folding video, featured on the homepage
2242
Reposted by Lars Dornfeld
Prism Science @prismscienceorg.bsky.social · 09/06/2026
We present sampleworks, an open, modular platform for generating and evaluating biomolecular conformational ensembles from structure predictors guided by experimental data doi.org/10.82153/jkx....
195
Reposted by Lars Dornfeld
Sebastian S. Cocioba @atinygreencell.bsky.social · 05/06/2026
Why does tryptophan have only one codon, UGG?
6235
Reposted by Lars Dornfeld
Kevin K. Yang 楊凱筌 @kevinkaichuang.bsky.social · 11/05/2026
Genie3: all-atom SE3-equivariance for fast and performant protein design. @yeqinglin.bsky.social @moalquraishi.bsky.social www.biorxiv.org/content/10.6...
0102
Reposted by Lars Dornfeld
Janine George @molecularxtal.bsky.social · 10/05/2026
To lead, you need a positive vision. A vision that is updated together with all of those that you lead. It can even be a challenging and demanding vision, but it needs to be there.
172
Reposted by Lars Dornfeld
Yeqing Lin @yeqinglin.bsky.social · 08/05/2026
Introducing Genie 3, a generative protein model that substantially advances the state-of-the-art for binder design, increasing in silico success rates by up to 20x on hard multimeric targets. It also debuts a form of inference-time scaling unobserved in other design models. 🧵1/8
25716
Reposted by Lars Dornfeld
Diego del Alamo @delalamo.xyz · 04/05/2026
LinkedIn as an employer looking for candidates is like ipTM: great for filtering candidates discovered elsewhere, but absolute dogshit for discovery. Quite a few people find ways to fluff up mostly-empty CVs and post a firehose of AI-generated posts. Anyway, role still open
182
Reposted by Lars Dornfeld
Stephen Turner @stephenturner.us · 03/05/2026
Free and open-source images, icons, and tools for creating scientific illustrations doi.org/10.59350/5zt... 🧪
doi.org
Free and open-source images, icons, and tools for creating scientific illustrations
Phylopic, NIH Bioart, Bioicons, Scidraw, Open Science Art, Health Icons, Servier Medical Art, Biodiversity Heritage Library, the Noun Project, Segment Anything, Excalidraw, draw.io, Biographics
6406227
Reposted by Lars Dornfeld
Irina Bezsonova @irinabezsonova.bsky.social · 28/04/2026
Excited to share that my 2025 collection of science-inspired drawings is now live on the Protein Data Bank (PDB) website. Thank you @rcsbpdb.bsky.social! It is free to explore and download. I’d love to hear what you think! pdb101.rcsb.org/sci-art/bezs...
pdb101.rcsb.org
PDB101: Irina Bezsonova Gallery
PDB-101: Training, Outreach, and Education portal of RCSB PDB
19934
Reposted by Lars Dornfeld
Pedro Beltrao @pedrobeltrao.bsky.social · 22/04/2026
"A minimal, pedagogical PyTorch reimplementation of AlphaFold2" github.com/ChrisHayduk/...
github.com
GitHub - ChrisHayduk/minAlphaFold2: A minimal PyTorch re-implementation of AlphaFold2's model & training
A minimal PyTorch re-implementation of AlphaFold2's model & training - ChrisHayduk/minAlphaFold2
0112
Lars Dornfeld @ldornfeld.bsky.social · 22/04/2026
This is great! I often realize that a healthy portion self-doubt makes people good scientists. It can be comfortable muting doubt with endless, calming LLM discussions, so I get why people do it. But it dilutes potential great ideas with little evidence.
130
Reposted by Lars Dornfeld
Simon Hammann @simonhammann.bsky.social · 20/04/2026
Don't be shy to take on a little two-week side project. These five months will be the most precious three years of your academic journey.
191958537
Lars Dornfeld @ldornfeld.bsky.social · 16/04/2026
Great work @neuroscinikolai.bsky.social 🙌
110
Reposted by Lars Dornfeld
Russell England @russellengland.bsky.social · 12/04/2026
France is switching from Microsoft to Linux "We must become less reliant on American tools and regain control of our digital destiny. We can no longer accept that our data, our infrastructure, our strategic decisions depend on solutions whose rules, pricing, evolution, and risks we do not control."
numerique.gouv.fr
numerique.gouv.fr
279104872255
Reposted by Lars Dornfeld
Alex Tong @alextong.bsky.social · 08/04/2026
Such an incredible journey building DISCO 🪩. Love working with this team. DISCO is a co-design model with functional, de novo, new-to-nature enzymes. Huge shoutout to my co-authors for making this a reality! 🚀👇
02411
Reposted by Lars Dornfeld
Joe Greener @jgreener64.bsky.social · 01/04/2026
On this 1st April I'm happy to release a technical report from our startup MirageBio, describing a new protein structure model with state-of-the-art performance. jgreener64.github.io/posts/techni...
jgreener64.github.io
MirageBio technical report
14618
Reposted by Lars Dornfeld
Martin Steinegger 🇺🇦 @martinsteinegger.bsky.social · 17/03/2026
AlphaFold database has entered the era of complexes. Together with NVIDIA, DeepMind and EBI, we use ColabFold, OpenFold and MMseqs2-GPU to predict ~31 million complexes (homo & hetro-dimers) resulting in 1.8 million high-quality predictions 📄 research.nvidia.com/labs/dbr/ass... 🌐 alphafold.ebi.ac.uk
8263110
Lars Dornfeld @ldornfeld.bsky.social · 03/03/2026
Great opportunity to work on interesting problems in protein design :)
021
Reposted by Lars Dornfeld
Amanda Moehring @flybehaviour.bsky.social · 13/02/2026
Ooooh. Cool new paper on origins of life. A simple 45-nucleotide RNA molecule that can perfectly copy itself. www.science.org/doi/10.1126/...
science.org
A small polymerase ribozyme that can synthesize itself and its complementary strand
The emergence of a chemical system capable of self-replication and evolution is a critical event in the origin of life. RNA polymerase ribozymes can replicate RNA, but their large size and structural ...
014358
Reposted by Lars Dornfeld
Joe Greener @jgreener64.bsky.social · 04/02/2026
I'm seeing a few neural networks that predict very long molecular dynamics time steps, and I'm not convinced that this is the way to go. We often want the trajectory not for its own sake, but in order to get snapshots of the underlying ensemble. (1/4)
141
Reposted by Lars Dornfeld
Khmelinskaia Lab @akhmelinlab.bsky.social · 03/02/2026
🗓️ Save the Date! The 5th European RosettaCon – Crossing Boundaries with Protein Design will take place in Lisbon, Portugal 🇵🇹 🗓️ October 28–30, 2026 Join the protein design community for an inspiring scientific meeting at the intersection of innovation and design.
185
Reposted by Lars Dornfeld
Wolfgang Huber @wkhuber.bsky.social · 30/01/2026
Plato's allegory of the cave is so relevant to the natural sciences, yet so many colleagues seem to think they can get away with writing about the shadows.
3274
Reposted by Lars Dornfeld
Martin Pacesa @martinpacesa.bsky.social · 22/01/2026
Here are the success rates of de novo pipelines based on which designs I could actually identify the methods for.
22312
Reposted by Lars Dornfeld
Kresten Lindorff-Larsen @lindorfflarsen.bsky.social · 20/01/2026
Another day, another paper describing an ML method for generating conformational ensembles without comparison to any experimental data 😩
3405
Reposted by Lars Dornfeld
Meredith Whittaker @meredithmeredith.bsky.social · 12/01/2026
We love to see it! Community-created harm reduction infrastructure to contest the alarming integration of AI agents into the Windows operating system (which is currently the most reckless deployment environment) ♥️ github.com/zoicware/Rem...
github.com
GitHub - zoicware/RemoveWindowsAI: Force Remove Copilot, Recall and More in Windows 11
Force Remove Copilot, Recall and More in Windows 11 - zoicware/RemoveWindowsAI
9941405
Reposted by Lars Dornfeld
Felix Pultar @pultar.bsky.social · 12/01/2026
Ever wanted to run MD simulations of entire proteins in water with DFT accuracy? Meet AMPv3-BMS25, the latest iteration of our AMP multiscale neural network potential by @rinikerlab.bsky.social Read more in the preprints: doi.org/10.26434/che... doi.org/10.26434/che...
1145
Reposted by Lars Dornfeld
Diego del Alamo @delalamo.xyz · 08/01/2026
2182