Next Phase Newsletter

Juan Carlos Lopez

Andy Marshall

While most parts of biotech early-stage financing have been in the doldrums in the past two or three years, so-called tech-bio startups have been thriving. Since the posterchild $1.0 billion mega series A round last April of Xaira Therapeutics, which was founded by scientists out of Nobel prize winner David Baker’s group at the University of Washington, several startups seeking to develop machine learning models for designing miniproteins or peptide binders of challenging or ‘undruggable’ targets have emerged, including Enlaza Therapeutics, Vilya, and UbiquiTx. All of these have been developing their own proprietary models based on Alphafold 3, Boltz-1 or Chai-1 for structure prediction and tools based off RFdiffusion, Bindcraft and ProteinMPNN for peptide design. Predicting CDR loops for de novo antibody design is a considerably more challenging task than for simple peptides, but Nabla Bio, founded last year by scientists out of George Church’s lab at Harvard, claims it is doing just that for GPCRs and ion channels. Earlier this month, Chai Discovery also launched with a $100 million series A from Menlo Ventures to optimize multimodal generative models such as Chai-2, which, according to the company, already “achieves a 16% hit rate in de novo antibody design.”

Designing peptides that can selectively bind to a protein target and show therapeutic activity remains a challenge, however, as it often depends on the availability of high-quality structural information about the target molecule, which is seldom available for many disease-relevant proteins that are unstructured or conformationally disordered. Similarly modeling protein-protein interactions like antibody-antigen interactions that are extremely dynamic and floppy also poses problems. All of which raises the question as to whether binders could be predicted simply using amino acid sequence information instead of structural data.

Now, a team led by Pranam Chatterjee from Duke University has addressed this question. In a recent paper in Nature Biotechnology, Chatterjee and his collaborators report the creation of PepMLM, a peptide binder design algorithm based on masked language modeling. A key feature of the algorithm is that it depends exclusively on protein sequence, not structure. Built upon the ESM-2 (Evolutionary Scale Modeling 2) protein language model, PepMLM masks and reconstructs entire peptide regions appended to target protein sequences. This design compels the model to generate context-specific binders. To train PepMLM, the team used high-quality curated datasets from PepNN and Propedia comprising ~10k putative peptide-protein sequence pairs. PepMLM output was consistently found to outperform RFDiffusion on held-out/structured targets, with a higher hit rate (38% to 29%) and low perplexities that closely matched real binders, with generated sequences showing target specificity, even in stringent permutation tests.

The model generated binders predicted to have higher binding scores than native and structure-based binders designed through other methods. Indeed, in vitro validation experiments confirmed the high affinity and specificity of PepMLM-generated binders.

Chatterjee and his colleagues went on to turn their binders into degraders by fusing them to E3 ubiquitin ligase domains, such as CHIP/STUB1. When tested in vitro, over 60% of these degraders knocked down their target proteins. PepMLM peptides achieved nanomolar binding affinity on the drug targets neural cell adhesion molecule 1 (NCAM1), a key marker of acute myeloid leukemia, and anti-Müllerian hormone type 2 receptor (AMHR2), a critical regulator of polycystic ovarian syndrome (where RFDiffusion-predicted peptides failed to bind). The authors also demonstrated that PepMLM-predicted peptides fused to E3 ubiquitin ligases not only degraded MSH3 but completely eliminated mutant huntingtin protein exon 1 containing 43 CAG repeats in Huntington disease patient-derived fibroblast cells. Similar results were obtained for a PepMLM-predicted peptide binder of MESH1, a protein controlling ferroptosis, in collaboration with Ashley Chi Jen-Tsan’s group at Duke University (RFDiffusion again gave no hits). And with Madelaine Dumas and Hector Aguilar-Carreno’s group, in collaboration with Matt Delisa’s group at Cornell University, PepMLM-derived peptides bound and reduced levels of viral phosphoproteins from Nipah, Hendra, and human metapneumovirus (HMPV); indeed, in live HMPV infection models, the PepMLM peptide mediated high levels of P protein clearance.

The ability of PepMLM to design binders purely on the basis of target-protein sequence is an important advance towards designing therapeutic peptides against hitherto inaccessible targets that lack structural data. Future work should explore how to incorporate chemical modifications such as cyclization or stapling to enhance stability of the binders, as well as the evaluation of the strongest candidates in vivo. Another challenge will be to ameliorate the immunogenicity of these foreign de novo proteins. The use of protein engineering approaches, such as incorporation of mirror amino acids that can cloak foreign peptides from the immune system, may offer solutions. But it is likely that candidates discovered using sequence or structure prediction tools will still require lengthy development programs to be turned into safe and effective drugs, despite the hype.

Tags: AI, alzheimers, art, artificial-intelligence, beauty, biology, books, cancer, fitness, food, genetics, health, immune-system, immunotherapy, nutrition, painting, poetry, science, technology, writing