Interaction‑first protein models
DeepSeqAI described training models on protein interactions rather than structure to map biologics across viruses and the proteome, noting evaluation by Gilead and funding from DARPA/NSF. The work was presented at SynBioBeta 2026 and framed as useful for designing variant‑resilient therapies. (x.com)
Proteins are tiny machines, and drugmakers need to know which ones stick to which targets before a biologic becomes a medicine. DeepSeq.AI said it is training artificial intelligence on protein interactions, not just shapes, to predict that matching at much larger scale. (syntheticbiologysummit.com) The company described the approach in a five-minute spotlight talk scheduled for May 5, 2026, at SynBioBeta in San Jose, California. The session page says DeepSeq trains protein language models on “billions to trillions” of experimental protein interactions from a single experiment. (syntheticbiologysummit.com) Most artificial intelligence protein systems start with sequence or three-dimensional structure, which is the folded shape of a protein. DeepSeq’s pitch is “function-first”: learn from measured binding behavior between proteins, then use that data to map biologics against viruses, immune receptors, and the human proteome. (syntheticbiologysummit.com) That distinction targets a practical problem in drug discovery. The SynBioBeta session page says predicting biologic affinity and immunogenicity remains a major challenge, and says broader interaction maps could help design therapies that stay effective as viral variants change. (syntheticbiologysummit.com) Biologics are large, engineered medicines such as antibodies that work by latching onto specific proteins. If a virus mutates or a drug also binds unintended human proteins, a candidate can lose potency or create safety problems, which is why companies screen for both target binding and off-target effects. (deepseq.ai) (nature.com) DeepSeq said Gilead Sciences signed a technology evaluation agreement on November 20, 2024, to assess its platform and wet-lab capabilities for designing and selecting biologics. The company also said on January 9, 2025, that it received a Defense Advanced Research Projects Agency award through the agency’s Biological Technologies Office pitch-day initiative. (deepseq.ai 1) (deepseq.ai 2) The SynBioBeta page says the platform has been “validated by Genentech” and funded by the Defense Advanced Research Projects Agency and the National Science Foundation. DeepSeq separately announced a National Science Foundation Small Business Innovation Research Phase I grant in 2024 for integrated protein design and screening. (syntheticbiologysummit.com) (deepseq.ai) The broader idea is not unique to one startup. A 2024 Nature Methods paper from researchers led by Marinka Zitnik described PINNACLE, a model trained on protein interaction networks across 156 cell types and 24 tissues, showing that interaction context can improve target prediction over context-free models. (nature.com) DeepSeq has since tied that interaction-first pitch to commercial expansion. On November 11, 2025, the company said Illumina Ventures invested in DeepSeq and quoted the firm as saying DeepSeq maps protein sequence to function with protein-protein interaction data at “unprecedented scales.” (deepseq.ai) The next test is whether those large interaction maps hold up outside conference stages and pilot deals. For now, the company is staking its case on a simple claim: in protein drugs, what binds may matter as much as what folds. (syntheticbiologysummit.com)
