Idris (@Edrees_Adl) header
Idris (@Edrees_Adl) profile picture

A young man interested in biology and bioinformatics

Joined January 2022
Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)Photo by Idris (@Edrees_Adl)
Thanks to the thousands of you that used the RFD4-Proteina online GUI and gave feedback! Thanks to the one and only @sokrypton it is now even easier to use, both on HF and on your own laptop/tablet/even phone! Make your own pesticide degrader like legend @SethWoodbury1 in our paper, including refolding! Links etc in 🧵
2
17
1
100
4,525
LLMs are powerful, so why do we still make them call tools to handle 3D molecules? Introducing Nahual: a decoder-only model for any sequence of text and 3D molecules. #NeurIPS2026
2
20
2
105
5,132
In roughly a day's worth of work with a few dollars of GPU compute, I could computationally design miniprotein binders against IL-36 in Codex. It involved a bunch of steps, including detecting and mitigating against reward-hacking of the design function and optimizing for developability. A slew of tools was involved: RFD4-Proteina, Boltz2, ESMFold, Protein-Sol. Runpod GPUs. Sounds impressive? Read through to the end for the point I am making. 1. RFD4-Proteina generation. Method and output: Original versus V2 reward, eight matched seeds, 16 successful searches. Each search used 60 steps, beam width 2 and branching 2, with checkpoints at steps 30 and 60. Saved 128 rollouts, including intermediate checkpoints; 74 distinct coordinate structures after exact deduplication. 2. Selection and deduplication. Method and output: Two final candidates per seed per arm: 16 original + 16 V2 entries. Cross-arm sequence deduplication yields 22 molecules: 6 V2-only, 6 original-only and 10 shared. 3. Standalone folding. Method and output: ESMFold predictions for all 22 unique sequences; assessed consistency with the designed binder structures. Added cost. 4. Complex prediction. Method and output: Boltz-2 predictions for all 22 sequences with IL-36γ, three poses each: 66 complexes. Fixed settings, three recycles, 200 sampling steps, shared target MSA and empty binder MSA; no coordinate templates or affinity model. 5. Developability prioritization. Method and output: All 22 scored with Protein-Sol and NetSolP-D, plus isolated-monomer exposed hydrophobic surface. Four candidates on the first Pareto front. 6. Detailed structural review. Method and output: U04 and U10: six existing complex poses examined for pose consistency, close contacts and hydrophobic patches. 7. Figures. Method and output: PyMOL ribbon pictures, interface close-ups and editable sessions from existing model-0 coordinates; no coordinate optimization. At the end I have a few candidates that are predicted to have good binding affinity and developability against IL-36. So have I designed a new protein binder or drug? Of course not. Of course I need to experimentally test these predictions. But I did this exercise to make a point: we are constantly hearing about especially the frontier labs designing all kinds of proteins against different targets. This work is legitimate and valuable and a great advance relative to where we were even two years ago. But the point is that someone like me who is not an expert can now do this with a few dollars of compute in a day or two; if I had a couple of thousand dollars to blow on tokens I could have done much more. It's not as impressive as it looks - and that's a good thing. Only large-scale experimental validation can validate the predictions that are coming out of AI-designed protein and small-molecule pipelines. This is no different from twenty years of computational chemistry and biology where we saw the best algorithms and docking and other models die on the rocks of experimental validation. AI is a great thing and a fantastic accelerant for work like this, but it's really important to put it in perspective and not get carried away. Retrosynthesis, protein design, small molecule design - all of these are now possible with AI, but none of them turn into proven science until there's large-scale validation of their predictions. We need to stop being unduly impressed with such exercises and treating every rapid computational molecular design exercise as a big deal, and in a way that's a great thing because it really shows us how far AI has come so quickly. A few computational design pipelines don't equal a step change in scientific discovery, especially new therapeutic discovery. If we unduly raise expectations from AI-enabled biology or drug discovery and related areas, we risk facing the same disillusionment that proponents of computational techniques in these fields have faced for fifty years. The proof is in the pudding, and those who eat it know the difference between how it tastes and how it looks like in a recipe book.
10
19
110
6,233
Idris retweeted
AlphaGenome 2 is live! A step change in performance and new capabilities over AG1! @adamgayoso's thread covers the key results. Manuscript: storage.googleapis.com/deepm… Blog: deepmind.google/blog/alphage… API: github.com/google-deepmind/a… A few notes under the hood & what's next:
Single-cell atlases continue to map the diversity of cell states, but how can we use them to interpret non-coding genetic variants? Excited to share AlphaGenome 2, our model for predicting cell-type-specific regulatory activity and variant effects from DNA sequence. deepmind.google/blog/alphage…
15
75
11
309
21,787
Creation of light emitting proteins by AI. AI designed a completely new enzyme capable of producing light through a chemical reaction. The artificial luciferase was only 13.9 kDa, remained stable above 95°C and showed catalytic efficiency comparable to natural luciferases. doi.org/10.1038/s41586-023-0…
2
37
1
167
9,781
Protein function often depends on protein dynamics. To design proteins that function like natural ones, how do we predict their dynamics? @HWaymentSteele and I are thrilled to share the first big, experimental datasets on protein dynamics and our new model: Dyna-1! 🧵
8
134
9
695
105,369
Idris retweeted
In collaboration with the @francesarnold and @pengliu_group research groups, we @GoogleDeepMind are proud to release AlphaProtein Novo, a pipeline for small-molecule enzyme design. We developed AlphaProtein Novo on a few standard model reactions, and validated designs in our own wet lab. We then wanted to challenge this system against nascent chemistries, so we asked leading scientists in biocatalysis whether we could partner on the reactions they cared about.
12
40
5
198
23,270
Idris retweeted
Introducing AlphaProtein Novo, a pipeline for de novo enzyme design! In collaboration with @francesarnold's group, we developed AP Novo and used it to get SoTA "novel-scaffold" activity on 2 benchmark reactions and create enzymes to synthesize the pharmaceutical motif piperidine and degrade the environmental toxin DEHP. Work led by @ZvxyWu @ZiqiLi0513 with teammates @alexechu_ @kelly_ruijie @schulluc @jacobjinkelly @rosaliags @davidla @solserpens @j_reisenbauer, @pengliu_group, Josh, Thomas, Amy, Tristan, Harshnira, and others. It was a joy and privilege to work with you all! Preprints and code: biorxiv.org/content/10.64898… biorxiv.org/content/10.64898… github.com/google-deepmind/a…
20
179
24
1,024
139,771
Today we’re releasing the fastest? protein folding models ever! @biohub actually trained ESMFold2 variants from the 300M and 600M parameter ESMC backbones, however, they did not release the models with confidence heads. Synthyra fixed that, now, the FastPLM versions of ESMFold2-300 and 600 can return pLDDT, pTM, and iPTM alongside structure. These models are actually very strong. On FoldBench, moving from 300M to 600M only gains about 0.008 IDDT and 1.8 percentage points in PPI pass rates. Moving from 600M to 6B only gains another 0.028 IDDT and 6 points. Our new confidence heads showcase high correlation to truth on held out evaluation sets. Roughly 0.8-0.9 spearman correlation across LDDT, TM, and DockQ. They can even tell which residues are intrinsically disordered from pLDDT alone with ROC AUC ~0.86. Importantly, correlation to the 6B model is ~0.7 on pLDDT, ~0.8 for pTM, and ~0.7 for iPTM. There are many gradient-based protein design methods at this point, (Bindcraft, Mosaic, etc.) so this monotonic relationship between these tiny model’s outputs and the production ESMFold2 are super important for cheap gradient based design, where critic stages can still come from the larger model. We have also integrated @AnthropicAI's new optimizations for protein language models, including our ESMFold2 implementations, for even more speed! Of course, like any FastPLM model, these are super easy to use – even on consumer hardware. I predicted the barnase-barstar complex structure in 18 seconds on my modest 3070 GPU, including imports and model download. Check out FastPLMs and try ESMFold2-300 and 600 today! github.com/Synthyra/FastPLMs huggingface.co/Synthyra/ESMF… huggingface.co/Synthyra/ESMF…
1
2
1
9
571
Avengers assemble: 1 year BakerLab + NVIDIA GenAIR = 🔥 rfd4-proteina.org Presenting RFD4-Proteina, our latest protein design model, significantly better and faster than everything we compared against, both in silico and in the lab! 1/30
27
207
45
675
73,228
📦 RFD4-Proteina just open-sourced its full training-structure store. Not classic RFdiffusion — this is the Baker Lab IPD × NVIDIA GenAIR release stack. Yesterday’s Hub commit (kierandidi) moved it from a partial PDB+AFDB cut to every source used by the release flow + AE recipes. Hard facts: → huggingface.co/datasets/nvid… → 2.13 TB · 19 configs · 33,527 objects → AFDB monomers, PDB interfaces, PLINDER ligands, glycans/Re-Glyco (~1.57 TB alone), RNA/DNA, Ab/TCR, NCAA/PTM, TED dimers, OMol, cyclic peptides… → Packaging/indices: CC BY 4.0 (structures keep upstream terms) → Model weights already BSD-3; you don’t need this dump to run inference Why it matters: rare for a frontier biomolecular design model to ship weights + code + the actual training shards/provenance. Reproducibility and fine-tune fuel, not just a demo. Caveat: distill ≠ experimental; Hub mix ≠ every checkpoint’s exact history; preprint, not peer-reviewed. 📄 Dataset: huggingface.co/datasets/nvid… 🧠 Model: huggingface.co/nvidia/RFD4-P… 💻 Code: github.com/RosettaCommons/RF… 📑 Paper: rfd4-proteina.org/assets/pap… #ProteinDesign #AIforScience #OpenSource #RFD4 #Proteina #NVIDIA
25
1
124
4,851
1/4 One DNA sequence, many chromatin states. Excited to share SMFNet, a diffusion model predicting single-molecule chromatin accessibility. Explainer video below! With co-first author Till (his thread: bsky.app/profile/tschwammle.…), Yuliya, Roze, @anshulkundaje & Arnaud Krebs.
2
43
3
168
21,934
We can now design all kinds of fancy de novo proteins. But can we produce them at scale, at a reasonable cost, and without extensive optimization? That’s what got this project started. Our preprint is now out! Much of the work was led by Selena @sesameandrea  and Kaori. 1/8
7
84
3
315
19,750
In parallel to the remarkable feat of predicting protein structure from amino acid sequence, we are seeing steady progress for doing the same—sequence to structure— for RNA via foundation models. The latest one today nature.com/articles/s41586-0…
7
32
1
158
21,816
1/18 🧵A thread about the evolution of four generative AI models for biomolecular design (Proteína, La-Proteína, Proteína-Complexa and RFD4-Proteina), which makes me think we're moving toward fully programmable protein design 🔥
1
25
129
5,651
RNA sequencing and a new computational tool called Species-OT help identify genes with similar functions across species, offering insights into gene regulation, stem cell development, and human biology. - @Ashbi_KyotoU - @KyotoU_News - @univkyoto rna-seqblog.com/rna-sequenci…
1
9
18
989
Contextualized gene representations work surprisingly well for predicting multigene perturbation effects. I’m excited to introduce EMBER, a pretrained gene-centric foundation model designed specifically to learn context-specific gene representations. A joint work with @jmuiuc and @ericxing. With innovations in single-cell pretraining and dataset curation, EMBER gene representations can be combined with a lightweight probe for single gene knockouts, genetic interactions (double knockouts), and even drug perturbations. Both a 🧵 and a 🎥, got in front of a camera to share some highlights! (1/n)
10
36
4
155
15,381
Can we help discover new, unknown small molecules using modern generative AI and scientific agents? Excited to share our work, led by superstar @zyxiong123, on ChefNMR2🧑‍🍳⚛️, a new approach for structure elucidation from NMR spectra! Paper: chefnmr.cs.princeton.edu/ 🧵 1/
8
47
8
257
25,854
#HighlightoftheWeek Speak to a Protein: An Interactive Multimodal Co-Scientist for Protein Analysis. #compchem pubs.acs.org/jcisd8/article/…
14
30
1,987