AI tools pushing biologics faster
Companies and labs are rolling out AI updates that accelerate biologics discovery and engineering, from Insilico’s Pharma.ai refresh (including PandaClaw and single‑cell tools) to sequence‑based viral‑feature models that predict glycosylation and cleavage sites. Observers point to real-world impacts — faster candidate generation, compressed antibody optimisation timelines, and improved trial readiness — that make these models among the few AI bets with commercial traction today. Those platform moves are already being paired with partnerships using real‑world data to sharpen biomarker and trial designs. (x.com) (x.com) (x.com)
Biologic drugs are made from living cells, so the hard part is not just finding a molecule that works. The hard part is finding a protein that still folds correctly, survives manufacturing, and behaves the same way every time a patient gets a dose. (nature.com) That is why artificial intelligence in this corner of biotech looks different from chatbot hype. These models are being used to rank drug targets, design candidate molecules, and predict whether a protein will have the chemical decorations and cut points that change how it acts in the body. (insilico.com) (link.springer.com) One of those decorations is glycosylation, which is a sugar chain attached to a protein after the protein is built. Those sugar chains can change folding, stability, immune visibility, and even whether a viral or therapeutic protein works at all. (nature.com 1) (nature.com 2) One of those cut points is a cleavage site, which is a short stretch of amino acids where an enzyme snips a protein like scissors cutting a ribbon. In SARS‑CoV‑2, a furin cleavage site on the spike protein helps control viral infection, and nearby glycosylation can increase or inhibit that cutting. (nature.com) That matters because sequence-only models can read a protein’s amino-acid string the way a language model reads text. Researchers are now using that approach to predict glycosylation patterns from sequence and structure signals, including new “glycoimpact” work and newer protein language model systems that improve site prediction. (biorxiv.org 1) (biorxiv.org 2) For drug companies, that turns wet-lab guesswork into a shorter shortlist. A model can flag which antibody variants are more likely to keep strong binding while also improving stability, viscosity, specificity, and manufacturability before scientists spend weeks making and testing each one. (link.springer.com) Insilico Medicine’s March 2026 PandaClaw launch is one example of that shift from single prediction tools to full research copilots. The company says PandaClaw sits inside PandaOmics, turns natural-language prompts into multi-step biology workflows, and can pull from more than 140 scientific skills and over 1,000 bioinformatics tools. (insilico.com) Insilico is pairing that software push with commercial deals, which is the part investors care about. On March 29, 2026, the company announced a collaboration with Eli Lilly worth $115 million upfront and up to about $2.75 billion in milestones, with Lilly getting an exclusive worldwide license for a portfolio of preclinical programs. (insilico.com) The reason these deals keep showing up is that biologics development has a lot of expensive bottlenecks but also a lot of measurable data. Insilico says more than 40 pharmaceutical companies use its technology, and its media materials say its internal pipeline includes 31 programs for 29 targets, with 7 already in clinical stages. (insilico.com) The same pattern is moving downstream into trials. Recent clinical research reviews describe artificial intelligence and real-world data being used to improve biomarker selection, patient matching, and trial design, which means the model is not only helping invent the drug but also helping decide who should get it first. (esmorwd.org) (pharmiweb.com) This is why biologics has become one of the few areas where artificial intelligence already has a believable business case. If a model cuts months off antibody optimization, avoids dead-end candidates, and makes trials more targeted, it does not need to replace scientists to earn its keep. (link.springer.com) (insilico.com)
