HIV evades immune detection through rapid mutation of its surface proteins, yet essential steps in viral entry, such as CD4 and co-receptor engagement, remain highly conserved. While therapies like Lenacapavir represent major advances, the emergence of resistant strains highlights the urgent need for adaptable, rapid-response antivirals. This challenge extends beyond HIV, demanding scalable design strategies for diverse viral threats. Here, we demonstrate that AI-driven design can address this need by generating cyclic peptide binders targeting a previously unexploited interface on the HIV-1 fusion protein gp41. Using only sequence information, without prior structural or binding site data, we designed and experimentally validated a single candidate. This inhibitor potently blocked infection by two HIV-1 strains in cell-based assays with no detectable cytotoxicity. Affinity analysis with SPR confirms the interaction with gp41 as designed. Our findings illustrate how AI-guided peptide design, coupled with rapid in-vitro validation, can accelerate early-stage therapeutic discovery and enable timely intervention against emerging viral threats.

The risk of pandemics is increasing as global population growth and interconnectedness accelerate. Understanding the structural basis of protein-protein interactions between pathogens and hosts is critical for elucidating pathogenic mechanisms and guiding treatment or vaccine development. Despite 21,064 experimentally supported human-pathogen interactions in the HPIDB, only 52 have resolved structures in the PDB, representing just 0.2%. Advances in protein complex structure prediction, such as AlphaFold, now enable highly accurate modelling of heterodimeric complexes, though their application to host-pathogen interactions, which have distinct evolutionary dynamics, remains underexplored. Here, we investigate the structural protein-protein interaction network between humans and ten pathogens, predicting structures for 9,452 interactions, only 10 of which have known structures. We identify 30 interactions with an expected TM-score ≥0.9, tripling the structural coverage in these networks. A detailed analysis of the Francisella tularensis dihydroprolyl dehydrogenase (IPD) complex with human immunoglobulin kappa constant (IGKC) using homology modelling and native mass spectrometry confirms a predicted 1:2:1 heterotetramer, suggesting potential roles in immune evasion. These findings highlight the transformative potential of structure prediction for rapidly advancing vaccine and drug development against novel pathogenic targets.

Structure prediction technology has transformed protein design, yet key challenges remain, particularly in designing novel functions. Many proteins function through interactions with other proteins, making the rational design of these interactions a central problem. While most efforts focus on large, stable proteins, shorter peptides offer advantages such as lower manufacturing costs, reduced steric hindrance, and improved cell permeability when cyclised. However, their flexibility and limited structural data make them difficult to design. Here, we introduce EvoBind2, a method for designing novel linear and cyclic peptide binders of varying lengths using only the sequence of a target protein. Unlike existing approaches, EvoBind2 does not require prior knowledge of binding sites or predefined binder lengths, making it a fully blind design process. For one target protein, we demonstrate that linear and cyclic peptide binders of different lengths can be designed in a single shot, and adversarial designs can be avoided through orthogonal in silico evaluation.

Structure prediction of protein complexes has improved significantly with AlphaFold2 and AlphaFold-multimer (AFM), but only 60% of dimers are accurately predicted. Here, we learn a bias to the MSA representation that improves the predictions by performing gradient descent through the AFM network. We demonstrate the performance on seven difficult targets from CASP15 and increase the average MMscore to 0.76 compared to 0.63 with AFM. We evaluate the procedure on 487 protein complexes where AFM fails and obtain an increased success rate (MMscore>0.75) of 33% on these difficult targets. Our protocol, AFProfile, provides a way to direct predictions towards a defined target function guided by the MSA. We expect gradient descent over the MSA to be useful for different tasks.

Proteins are dynamic molecules whose movements result in different conformations with different functions. Neural networks such as AlphaFold2 can predict the structure of single-chain proteins with conformations most likely to exist in the PDB. However, almost all protein structures with multiple conformations represented in the PDB have been used while training these models. Therefore, it is unclear whether alternative protein conformations can be genuinely predicted using these networks, or if they are simply reproduced from memory. Here, we train a structure prediction network, Cfold, on a conformational split of the PDB to generate alternative conformations. Cfold enables efficient exploration of the conformational landscape of monomeric protein structures. Over 50% of experimentally known nonredundant alternative protein conformations evaluated here are predicted with high accuracy (TM-score > 0.8).