Specificity-driven protein binder design with Odin-Multi 1 Odin-Multi is a multi-context de novo binder design framework that optimizes one shared binder sequence across multiple complexes simultaneously, using attractive losses for on-targets and repulsive losses for defined off-targets—so specificity/cross-reactivity becomes a design objective rather than a downstream screening outcome. 2 The core mechanism: in each optimization iteration, frozen AlphaFold2 predicts the same evolving binder sequence in parallel with each target/off-target; per-context objectives (e.g., interface confidence, contacts, compactness) generate gradients that are weighted and integrated to update a single shared sequence state via a three-stage continuous-to-discrete optimization schedule. 3 In silico cross-reactivity benchmark on class B1 GPCR pairs (GLP-1R, GCGR, GIPR): jointly optimized peptides were far more likely to exceed an interaction-confidence threshold for both receptors than single-target campaigns, reaching 83.5–96.8% dual-pass rates vs 6.8–36.3% in controls (using iPTM > 0.5 for both). 4 In silico cross-reactivity benchmark on short-chain three-finger toxins (Erabutoxin A and NK-shNTx): joint optimization increased dual-target success from 0.8% (single-target) to 9.2% (joint) at iPTM > 0.5, indicating population-level enrichment of designs predicted to bind both homologous toxins. 5 In silico specificity benchmark on pMHC where target and off-target differ by only one peptide residue (NY-ESO-1 SLLMWITQC vs M4A SLLAWITQC on HLA-A*02:01): counter-selection improved the fraction of designs meeting both target-confidence and separation criteria; at an illustrative target/off-target iPTM ratio threshold of 2.5, yield rose from 6.0% to 14.2%. 6 Experimental validation (cross-reactivity mode): from 30 AF3-filtered toxin minibinders, one lead (Poly 5) showed toxin-associated binding in DELFIA and then apparent nanomolar affinity by BLI to Erabutoxin A and to a candidate NK-shNTx-containing Naja kaouthia venom fraction (KD1 ~11.95 nM and ~34.43 nM, respectively, under a heterogeneous-ligand model). 7 Structural interpretation for Poly 5: predicted complexes suggest a conserved binding mode engaging the exposed β-sheet face of both toxins with a largely backbone-dominated core, while accommodating sequence differences via peripheral contacts—consistent with designing “family-covering” recognition rather than two unrelated interfaces. 8 Experimental validation (specificity mode) used mammalian surface display and pooled FACS screening of 500 pMHC minibinders designed under three regimes (no off-target; CMV+empty-HLA counter-selection; M4A+empty-HLA counter-selection). Increasing counter-selection reduced yield but produced a notable discriminator. 9 A lead from the stringent M4A counter-selection condition (S2) showed higher target tetramer staining than a previously reported NY-ESO-1 minibinder (NY1-B04) while also achieving better NY-ESO-1 vs M4A discrimination; AF3 modeling suggests Met4 of the target peptide sits in a minibinder pocket, offering a plausible structural basis for sensitivity to the Met→Ala change. 10 Practical takeaways: multi-context optimization increases compute roughly with the number of contexts (each needs its own structure prediction per iteration) and can reduce design yield, but it can shift the entire generated population toward desired interaction profiles—expanding controllable behaviors beyond “binds target” to “binds these targets” and/or “avoids these off-targets”. 📜Paper: biorxiv.org/content/10.64898… #ProteinDesign #ComputationalBiology #AlphaFold #BinderDesign #Specificity #MultistateDesign #Antivenom #Immunotherapy #pMHC #GPCR

🧪🐍 AI vs venom! Our Nature paper with David Baker shows AI-designed proteins can block deadly snake toxins. Now featured by Reuters! 🎥 ✅ 80–100% survival in mice 💸 Cheap, fast, scalable 🌍 A step toward better antivenoms 📄 nature.com/articles/s41586-0… 🎬 piped.video/gnvEDMEr2mc

Another Huge, insane cancer Breakthrough A new AI based method can produce specially designed proteins in just a few weeks that can arm the T cells in the body's immune system to attack and kill cancer cells Researchers from the Technical University of Denmark and the Scripps Research Institute have developed an advanced AI platform that rapidly designs protein based therapies to arm a patient's immune system against cancer, a major leap in precision medicine. Published in Science, the study shows for the first time that it’s possible to design proteins entirely in a computer to redirect T cells toward cancer cells via pMHC molecules, reducing the discovery timeline from YEARS! to just 4–6 weeks. Using this method, the team created “IMPAC-T” cells, engineered T cells that successfully killed cancer cells in lab tests. One key example was a protein targeting NY-ESO-1, a marker found in many cancers. The platform also proved effective in designing custom treatments for a melanoma patient, demonstrating its potential for personalized therapies. The innovation includes a virtual safety screening, allowing scientists to predict and eliminate dangerous side effects before lab testing.

De Novo-Designed pMHC Binders Facilitate T Cell–Mediated Cytotoxicity Toward Cancer Cells @ScienceMagazine 1. A groundbreaking study presents a rapid de novo design platform for minibinders (miBds) targeting cancer-associated peptide-bound major histocompatibility complex (pMHC) using advanced generative models, achieving high-affinity binding and T cell-mediated cytotoxicity against cancer cells. 2. The platform leverages in silico cross-panning and molecular dynamics simulations to enhance specificity and predictability of miBds, significantly improving the success rate of identifying binders compared to traditional methods. 3. A high-affinity binder targeting the NY-ESO-1 antigen was identified and validated through cryo-electron microscopy, demonstrating its potential for precision immunotherapy when incorporated into chimeric antigen receptors (CARs). 4. The study also successfully designed and validated binders for a neoantigen pMHC complex with an unknown structure, showcasing the platform's applicability to personalized cancer therapy. 5. The computational pipeline, including RFdiffusion, ProteinMPNN, and AlphaFold2, enables efficient and accurate design of miBds, with a success rate of 10% for ultrahigh target affinity binders, highlighting the potential for rapid therapeutic discovery. 6. The findings suggest that this approach could be transformative for targeting viral antigens and autoimmune diseases, in addition to cancer, by providing a versatile and precise method for pMHC targeting. @TimothyPJenkins 📜Paper: science.org/doi/10.1126/scie… #ProteinDesign #CancerTherapy #Immunotherapy #ComputationalBiology #PersonalizedMedicine

Thrilled to share our new paper in @ScienceMagazine! 💥 We developed a computational platform to design tiny 'minibinders' from scratch that can guide T cells to kill cancer cells. This is proof-of-concept for targeting intracellular cancer antigens with a new modality! 🎯

While I’m bouncing between emails and desperately waiting for my holiday, it’s good to know at least something in the lab is holding it together Thermal stability might not be glamorous, but it’s underrated Designerbodies: even when things get heated, still doing their job ;)

🧬 In March, we published our research for InstaNovo and InstaNovo+ in @NatMachIntell, our diffusion-powered ‘de novo’ peptide sequencing models built to uncover the secrets of the human proteome. bit.ly/3XBeJFh

📆 Six months, four publications, one cover. We’re halfway through the year and InstaDeep research is powering ahead with multiple boundary-pushing papers published in the @NaturePortfolio! Catch up on the highlights below 🔽

At @affinityaibio we’re moving from “finders keepers” to “designers binders”, one generative model at a time. If you could have a binder to anything, what would it be? (Proteins, obscure signalling pathways, that one reviewer who never gets back to you…;)

At AffinityAI, we combine Danish design sensibilities with precision molecular engineering. Elegant. Efficient. Engineered to bind. 🧬 Discover how we’re reshaping molecular function: affinityai.bio

🎧 Had the pleasure of joining Anders Høeg Nissen on the AI Denmark Podcast to talk about how we’re using AI to crack the protein code 🧬 Catch my segment from 9:30 (and yes — it’s in English 😉) 🎙️ open.spotify.com/episode/4g3…

Great to see AffinityAI represented by @esperdet and Oliver Morell at the @DTUtweet Startup Day 💪 Lots of interest in what we’re building, especially the fact that we deliver #real #reagents rather than just #AI predictions. Excited for what's to come! @DTUSkylab