Evolutionary Virologist@VUB (form. Ntl. Ref. Gov. Lab.) - Ai-based Struct. Biology Bio-Defense/Modulome/OneHealth Comput’al Proteomics and Founder@MetaVIRIA🦠
Belgium
Joined November 2018
- Tweets556
- Following776
- Followers516
- Likes975
Rachid Tahzima, PhD retweeted
localfold.org updated to include:
- bioemu2 @FrankNoeBerlin
- mol1 @MolenceBio
- fastplm (confidence modules for esmfold fast models) @Logan_Hallee
- af3x crosslinks @jankosinski
- faster diffusion (exponential update and geometric noise schedule) @vinay_matt
Rachid Tahzima, PhD retweeted
Public preview of BioEmu2! @MSFTResearch AI for Science’s new foundation model for biomolecular structure, dynamics and function. Open weights, MIT license.
github.com/microsoft/bioemu2
Movie: BioEmu2-simulated binding of intrinsically disordered peptide PUMA to MCL-1 receptor. 🧵⬇️
Rachid Tahzima, PhD retweeted
🚨 NVIDIA × Nobel laureate David Baker’s team just unveiled RFD4-Proteina.
Protein design is entering the reasoning era.
The model unifies proteins, noncanonical amino acids, sugars, DNA, and RNA in a shared latent space, enabling iterative refinement and inference-time search.
🧬 10.6% computational success rate on protein binder design benchmarks, with 42× the throughput of BindCraft.
Even more impressive: wet-lab validation includes a de novo designed enzyme for a key step in the synthesis of the HIV drug lenacapavir.
Code and model weights are now available.
🔗 rfd4-proteina.org/
Rachid Tahzima, PhD retweeted
From childhood prodigy to AI pioneer to Nobel Prize winner, Demis Hassabis has always had a simple plan. Step one, solve intelligence. Step two, solve everything else. 🔗 on.wsj.com/4i0Axoo
Whatever
thelancet.com/action/showPdf…
30
Rachid Tahzima, PhD retweeted
The cool thing about the Fold'EM paper, which directly uses images of EM particles as inputs for diffusion guidance, is that in principle it can be ported over to a codesign model like BoltzGen for modeling display/immunization-derived antibodies that have never been sequenced
Rachid Tahzima, PhD retweeted
Computational design of protein conformational landscapes for metamorphic behavior
1. The study designs de novo metamorphic proteins that can populate two distinct, well-folded conformations without any external trigger, aiming to move protein design from single static structures to controllable conformational landscapes.
2. The workflow starts from 93 small natural domains (60–150 aa) and uses RFdiffusion to partially remodel termini while keeping cores fixed, generating 200 alternative conformations per domain and selecting large-change pairs (RMSD > 4 Å and ≥40% secondary-structure change in remodeled residues).
3. A key observation: substantial structural changes can be achieved with relatively modest sequence changes. Many alternative-state designs retained >50% sequence identity to the native protein even when the native vs alternative structures differed by up to ~16 Å RMSD.
4. To ensure transitions are structurally plausible (not just two endpoints), the authors built a coarse-grained, sequence-independent transition modeling step inside RFdiffusion using auxiliary potentials: an “unfolding” potential to disengage the moving region and a “transformation” potential to bias toward the target state, producing dense ensembles of intermediate backbones.
5. From thousands of native/alternative pairs, they focused on three transition types with clear structural logic: helix reorientation (H2H), beta-to-helix (B2H), and helix-to-beta (H2B). Alternative folds were experimentally validated; for example, crystal structures for stabilized alternative states matched designs at ~1–2 Å RMSD.
6. Two complementary sequence-design strategies were tested for metamorphism: (i) chimeric “trajectory” design that restricts each position to amino acids from either endpoint sequence (searching the sequence space between states), and (ii) fully de novo iterative ProteinMPNN redesign that optimizes one sequence against both backbones by exchanging per-residue logits biases until convergence.
7. Biophysics and NMR show the central challenge of dynamic design: many candidates are folded (CD, SEC-MALS) yet display exchange broadening, missing peaks, and heterogeneous intensities in HSQC spectra, consistent with micro-to-millisecond conformational exchange concentrated in the metamorphic regions.
8. Unbiased MD (hundreds of microseconds aggregate) supports intrinsic multi-state behavior. For H2H-chimeric, multiple replicas underwent reversible helix reorientation approaching the alternative-like state; for B2H-full, the native beta state was kinetically more stable while the alternative helix showed partial unfolding/refolding and transient beta-strand propensity.
9. A major contribution is a practical strategy to steer state populations with minimal mutational load. The authors combine AF3 (for filtering state-specific predictions) with BioEmu (for ensemble sampling) to identify “state-defining” mutations; 2–3 mutations were sufficient to shift equilibria toward either state.
10. Structural confirmation closes the loop: for B2H-full, a native-stabilized mutant yielded a crystal structure matching the beta state (~2.6 Å RMSD), while an alternative-stabilized mutant enabled an NMR structure confirming the designed helix (~2.2 Å RMSD). For H2H-chimeric, an alternative-stabilized mutant was confirmed by crystal structure (~1.8 Å RMSD), demonstrating that designed sequences can be tuned across distant conformations.
📜Paper: biorxiv.org/content/10.64898…
#ProteinDesign #MetamorphicProteins #ConformationalDynamics #RFdiffusion #ProteinMPNN #AlphaFold3 #NMR #MolecularDynamics #ComputationalBiology #Bioinformatics
Rachid Tahzima, PhD retweeted
Banger paper from Google DeepMind on co-evolving an agent's harness and its weights.
(bookmark it)
If you already log your harness changes and results, those logs can become training data for the model running inside the harness.
This is a nice paper from Google showing how to do this co-evolution.
Specifically, they propose Evo2, a framework for weight-harness co-evolution.
How it works:
One policy plays two roles. As a Proposer, it writes edits to its own harness, and as a Reflector, it diagnoses what each edit changed. Those edits and their outcomes then become training data for both roles, using SFT followed by interleaved GRPO.
The result:
Evo2 reaches 48.3% on SWE-bench Verified and 82.1% on Polyglot, against 41.0% and 62.6% for the strongest baselines. The harness part alone moves a GPT-5-mini agent from 24.4% to 48.3% on SWE-bench Verified.
The evolved harnesses also transfer to other benchmarks without further evolution.
Paper: arxiv.org/abs/2610.13025
Chat with Paper: academy.dair.ai/papers/evo-2…
Rachid Tahzima, PhD retweeted
In October 2024, an American man returned to Iowa after traveling to Liberia, unknowingly carrying Lassa fever. His illness showed how quickly a health threat abroad can become a threat at home. Drawing on 16 years of CDC partnership in Liberia, CDC-trained disease detectives and other global health teams partnered with CDC experts to quickly provide U.S. health officials with critical guidance to help prevent its spread—protecting Americans here at home. #CrisisToControl
Rachid Tahzima, PhD retweeted
E. coli Outbreak: 13 people sick in 9 states with 8 hospitalized. Over half of the illnesses are in children age 5 or younger. Do not eat recalled raw milk cheeses sold by Sierra Neveda Cheese Company. All Graizers raw milk cheeses were recalled. See full details:
cdc.gov/ecoli/outbreaks/raw-…
Rachid Tahzima, PhD retweeted
During the UN General Assembly in New York, CDC leadership met with DRC Prime Minister Judith Suminwa to discuss continued Ebola response support, research, and coordination.
#RDCongo #UNGA81 #Ebola
En marge de son séjour à New York, la Première Ministre, Cheffe du Gouvernement, Judith Suminwa Tuluka, s’est entretenue avec la Directrice générale du Centre américain de contrôle et de prévention des maladies (CDC), Dr Erica Schwartz, accompagnée de ses équipes d’Atlanta et du Directeur pays du CDC en RDC, Dr Delayo Zomahoun.
Les échanges ont notamment porté sur :
• la poursuite de l’accompagnement du CDC dans la riposte contre Ebola, tout en maintenant l’attention sur d’autres priorités de santé publique
• la valorisation de l’expertise congolaise acquise dans la gestion des épidémies d’Ebola, notamment contre la souche Bundibugyo
• la nécessité de préserver la recherche scientifique et de renforcer la transparence financière dans la gestion des crises sanitaires
• l’amélioration de la traçabilité des ressources mobilisées par les partenaires internationaux
La Cheffe du Gouvernement a réaffirmé la nécessité d’un appui international mieux coordonné, transparent et pleinement articulé avec les capacités et l’expertise nationales.
Rachid Tahzima, PhD retweeted
A new protein design optimization framework from the Baker Lab: RFOptimization.
Instead of generating more candidates, it rescues near-miss designs through RF3-guided mutation, Boltz/MPNN cycling, and AF3 filtering—boosting PPI success from 6.7% to 22.3%.
Preprint: doi.org/10.64898/2026.09.04.…
🤖 Made with AI
Rachid Tahzima, PhD retweeted
Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work.
We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets.
We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.
Singularity is …. Past
Signing the End of a Toxic Class of “Clerics and Lunatics” …. Et bon débarras
theatlantic.com/science/2026…
48
Rachid Tahzima, PhD retweeted
isn't that crazy! 2000 binders with Average_dG(rosetta) Lower than -100. Some uniprot ids; uniprot.org/uniprotkb/Q6UWZ7…
uniprot.org/uniprotkb/Q86XE0…
uniprot.org/uniprotkb/Q969H6…
The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates
1. The paper introduces the Human Bindome: 306,146 designed, genetically encodable protein binder candidates (fixed length 90 aa) spanning 8,296 human proteins, each with an explicit sequence, a predicted binder–target complex structure, and in silico confidence metrics.
2. Core motivation: conventional antibody reagents are often expensive, variably characterized, and frequently lack disclosed sequences, contributing to reproducibility issues; the Bindome aims to provide openly specified alternatives at proteome scale.
3. Methodologically, they scale the experimentally benchmarked BindCraft pipeline (reported prior success rates 10–100% depending on target) by adding automated domain-level target selection and an accelerated parallel implementation suitable for proteome-scale execution.
4. Target selection is driven by AlphaFold DB full-length models plus PAE matrices: proteins are segmented into domains via PAE-based clustering, and “targetable domains” are filtered for confident folding, compactness, contact density, acceptable length, and exclusion of membrane-embedded regions.
5. A key engineering detail: designing against cropped domains makes computation cheaper and enables multiple distinct sites per protein, but can create clashes with other parts of the full-length protein; they address this by superimposing each complex onto the full-length model and rejecting sterically clashing designs unless the clash is in low-confidence regions (high PAE / low pLDDT).
6. All candidates pass unchanged BindCraft stringency thresholds, including binder monomer fold confidence (pLDDT > 0.8) and interaction placement confidence (normalized iPAE < 0.35); distributions are not piled up at cutoffs (medians pLDDT 0.92, iPAE 0.23), enabling within-target prioritization.
7. Coverage numbers: the “targetable proteome” is 10,664 of 20,279 proteins; Bindome covers 77.8% of that targetable set and 40.9% of the full human proteome, with ~37 candidates per target protein on average.
8. Structural breadth: mapping domains to CATH superfamilies suggests 78.2% of superfamilies represented in the human target-domain set have at least one binder candidate; uncovered regions are largely outside the current scope (notably integral membrane proteins).
9. Epitope biology: binder epitopes show convergence within domain families (e.g., grooves/pockets in WD40 and bromodomains; broad kinase-domain surfaces), and epitope residues are significantly more evolutionarily conserved than matched surface controls—more consensus across binders corresponds to higher conservation.
10. Functional relevance and perturbation potential: epitopes often overlap annotated functional sites (PTMs, protein–protein interfaces, ligand-binding sites, DNA-binding regions), and 26% of ASD-annotated allosteric targets have at least one binder overlapping an allosteric site; examples include candidates positioned to occlude TP53 DNA-binding or disrupt SOD1 interaction interfaces.
11. Interface properties vs native PPIs: binder interfaces are native-like in mean area and slightly more polar, but show a strong bias toward helix-mediated contacts on the designed binder side; binder folds are often not globally novel, yet interfaces appear largely novel by interface similarity comparisons.
12. Accessibility is treated as a first-class output: a free web interface supports ranking/inspection and bulk downloads; a standardized 3D-Beacons API enables programmatic retrieval; and an MCP server allows natural-language “agentic” querying via LLM tools.
13. ML utility: the dataset is released with leakage-controlled train/val/test splits based on interface clustering (205,553 train; 21,205 val; 21,955 test), designed to support development/evaluation of next-generation binder design models.
💻Code: github.com/wejulius/Bindome
📜Paper: biorxiv.org/content/10.64898…
#ProteinDesign #ComputationalBiology #StructuralBioinformatics #AlphaFold #ProteinBinders #Reproducibility #FunctionalGenomics #DrugDiscovery #MachineLearning
Rachid Tahzima, PhD retweeted
Humanized Anti-PD-1 Antibodies Generated Using the Conditional Kernel-Elastic Autoencoder
1. The work proposes an AI-guided “interpolation” strategy to generate new humanized anti–PD-1 antibodies between two FDA-approved endpoints—pembrolizumab and nivolumab—that bind the same PD-1 FG-loop hotspot but in nearly opposite orientations (~174° apart).
2. Core method: a conditional kernel-elastic autoencoder (CKEA) trained on ~130,000 antibody sequences (Feldhaus library) learns a chemically informed latent space; generation is conditioned on a high “antibody medicine-likeness” target score (set to 10) to bias designs toward developable sequence properties rather than only binding.
3. Six candidates (P2N_1 to P2N_6) are generated by linear interpolation in latent space; P2N_1 is closest to pembrolizumab, P2N_6 closest to nivolumab. Sequence divergence is quantified using the Gonnet substitution matrix (scaled at 250 PAM).
4. A key design rationale is structural: pembrolizumab and nivolumab overlap on PD-1 and both target the FG loop, yet use distinct auxiliary contacts (pembrolizumab engages the C’D loop; nivolumab engages the N-terminus), creating two very different binding geometries that define informative endpoints for latent-space sampling.
5. Instead of relying on static crystallography alone, the study uses long-timescale molecular dynamics (MD) to assess whether generated antibodies can form stable complexes and to refine equilibrium interaction patterns; each designed antibody is simulated for 400 ns, starting from both pembrolizumab-like and nivolumab-like docking poses to address kinetic trapping between the two orientations.
6. The paper emphasizes “MD-derived electron density / electrostatic potential maps” built from trajectories as a stability diagnostic: well-defined interfacial density features indicate limited fluctuations and a more stable bound state, while poorly defined maps suggest lack of convergence within the simulated window (not necessarily non-binding).
7. Mechanistic motifs highlighted by MD: stable binding often involves Arg/Lys “finger” insertions from PD-1 into antibody pockets plus clustered hydrogen-bond networks (often to backbone atoms). These motifs appear in several designed complexes (e.g., PD-1 R86 insertion in P2N_1/P2N_2; dense HB cluster across PD-1 E84–Q91 in P2N_3).
8. An important negative/nuanced result: proximity in latent space does not reliably predict binding orientation or mode. Example: P2N_1 is closest to pembrolizumab in latent space but shows a sizable domain rotation (~43°) relative to the pembrolizumab pose, suggesting future generative models may need explicit antigen/epitope conditioning.
9. Experimental validation: antibodies were expressed and tested by BLI against immobilized PD-1-Fc. Five of six designs were purified (P2N_1 showed favorable MD complex formation but failed purification). Three of six designed antibodies showed detectable PD-1 binding; signals were unexpectedly weak, limiting confident kinetic extraction and raising possible assay/immobilization and temperature (MD at 37°C vs BLI at 25°C) considerations.
10. Overall contribution: a combined pipeline—sequence-generative modeling conditioned on developability, physics-based MD evaluation from multiple starting poses, and wet-lab binding tests—demonstrates a practical route to propose novel humanized anti–PD-1 binders that can be further improved via computational saturation mutagenesis or laboratory evolution.
📜Paper: biorxiv.org/content/10.64898…
#ComputationalBiology #AntibodyEngineering #ProteinDesign #MachineLearning #GenerativeModels #Immunotherapy #PD1 #MolecularDynamics #Biophysics #Bioinformatics
