PTM-Mamba is out in @naturemethods!! 🥳 PTM-Mamba is the first/only pLM to encode both WT and post-translationally modified residues as tokens! And we do this without losing any base protein information from ESM-2! 🤝 📜: nature.com/articles/s41592-0… 💻: github.com/programmablebio/p… We all know PTMs like phosphorylation, acetylation, glycosylation, and ubiquitination massively expand the functional space of proteins (and can often cause disease if dysregulated!) — but for some reason, protein-specific models like AlphaFold, ESM, RFDiffusion never explicitly include them.🤷‍♂️ And the all-atom models? Well, they rely on sparse structural data of PTM'ed proteins, precluding PTM modeling of conformationally diverse proteins. 🤦‍♂️ PTM-Mamba fixes this: We EXPLICITLY allow you to input PTMs in your sequence.🪙We combine novel bidirectional Mamba blocks with ESM-2 embeddings and fuse them using a learnable gating mechanism. This enables PTM-Mamba to model >25 types of PTMs as actual sequence tokens and to retain wild-type signal, all while efficiently and accurately capturing new PTM information. 🚀 But PTM-Mamba doesn’t just include PTMs — it understands them. 🧠 Embeddings cluster modified and WT sequences distinctly yet coherently, while token embeddings show class-aware organization (i.e., phospho and acetyl groups form neighborhoods). 🏘️ We benchmark across 4 PTM-centric tasks that require explicit PTM information: disease association, druggability, PTM effects on PPIs (new!), and zero-shot PTM discovery. On the first two tasks, PTM-Mamba perform strongly vs. ESM-2 and PTM-Transformer. 💯 On the PPI task, we even outperform a newly-trained PTM-SaProt, indicating that sequence information may be enough to model PTM effects on PPIs. 😋 And finally, just by analyzing the per-residue output logits for masked positions in WT sequences, we can accurately discover which PTMs are likely at each position in an input protein! A super useful tool for biologists! Just put in your protein sequence, and get diverse PTM predictions. 🧑‍🔬 What's next? Like ESMFold, we can take these new embeddings and predict the structures of PTM-ed proteins via a folding trunk! But you know us: we will focus on developing peptide binders to modified proteins for therapeutic application. 💊 I'm super proud as always of my amazing PhD student, @pengzhangzhi1! 🌟 In just over one year in my lab, he's produced many beautiful works -- from PTM-Mamba, to path planning (P2) for discrete diffusion models, and flash attention variants of pLMs for efficient use. I can tell you that he really cares about making pLMs useful to the community. 🥹 And this all came together with great teamwork from my students @TongChen888, @SophieVincoff, and our collaborator @WChentong at @Westlake_Uni! 🤝 Try out PTM-Mamba for yourself and happy PTM-aware protein modeling!! 😊