03-18-Daily AI News Daily

USER MESSAGE BEGIN —

I need you to translate and edit this Chinese technical content into English following the specific role instructions provided.

— USER MESSAGE END —

Daily Summary

DeepMind's new London headquarters Platform 37 opens, with an AI exchange center on the first floor where AlphaFold 4 may be born.
AlphaFold3 designs directed IL-2 for safer CAR-T therapy, AI protein design achieves breakthrough in miniature inhibitors.
Life science AI tools are becoming "democratized," blooming across drug design to disease prediction—worth continuous attention.

⚡ Quick Navigation

• 📰 Today’s AI News - Latest updates at a glance

💡 Tip: Want to experience the latest AI models mentioned in this article (Claude 4.5, GPT, Gemini 3 Pro) first? No account? Head to Aivora to grab one—one minute setup, worry-free support.

Today’s AI Life Science News

👀 One-Liner

DeepMind built an “AI temple” in London dedicated to life science breakthroughs.

🔑 3 Key Hashtags

#AIProteinDesign #MedicalAI #BioComputing

🔥 Top 10 Headlines

DeepMind’s New Headquarters Platform 37 Unveiled, Focused on AI Life Science Breakthroughs

AlphaGo’s Move 37 changed the game of Go—now DeepMind names its new headquarters Platform 37 to honor that legendary moment. This building isn’t just an office; the first floor hosts an “AI Exchange Center” where regular people can visit exhibitions and attend events to learn how AI is transforming life science. Demis Hassabis calls it “a space dedicated to science and AI,” suggesting AlphaFold 4 and next-generation protein design tools may be born here.

AlphaFold3-Driven IL-2 Immunotherapy Design: Making CAR-T Cells More Precise

CAR-T therapy has a persistent challenge: activating immune cells with IL-2 also damages normal cells, causing severe side effects. Now a research team used AlphaFold3 to design a “directed” IL-2 that activates only CAR-T cells without touching other immune cells. Using a physics-constrained sequence generator to screen mutants, they achieved an ipTM (predicted structure quality metric) of 0.724 with nearly zero off-target binding. This means future CAR-T therapy could be safer and more effective—good news for leukemia patients.

Complement C9 Miniprotein Inhibitor: Using AI to Design a “Molecular Shield”

Sometimes the human immune system “friendly-fires,” with the complement system attacking the body’s own cells and causing autoimmune disease. This Nature sub-journal paper used AI to design miniproteins that specifically block complement C9 assembly, like adding a “molecular shield” to cells. These miniproteins are much smaller than traditional antibodies, penetrating tissues more easily—potentially treating rheumatoid arthritis, lupus, and other conditions. AI protein design scores another win.

LongHap: Improving Genome Haplotype Reconstruction Using Methylation Signals

Sequencing tech keeps advancing, but accurately separating the two strands of chromosomes (haplotypes) remains challenging. LongHap, a new tool, doesn’t just look at DNA sequences—it leverages methylation signals (an epigenetic marker) in long-read sequencing data. The result: lower error rates, better haplotype continuity, and especially strong performance on medically relevant genes. This is crucial for precision medicine—the same mutation on different haplotypes can have completely different pathogenic effects.

DEX: Amino Acid Substitution Matrix Based on Deep Mutational Scanning

Evolutionary biology has a classic question: which amino acid substitutions occur more easily? A research team tested 30 amino acid distance matrices and found experimental data (especially deep mutational scanning) outperforms theoretical models. Their DEX matrix combines top methods and performs best at predicting codon substitution patterns. This benefits protein engineering, evolutionary analysis, and even AI protein design—knowing which mutations are more “natural” helps design more stable proteins.

Multi-Scale Cross-Modal Fusion Framework for Drug-Target Binding Prediction

A core challenge in AI drug discovery: can this drug molecule bind to the target protein? The MSCMF-DTB framework analyzes both the chemical structure of drugs and protein sequences, using multi-scale fusion to translate both data types into a unified language. This is far more accurate than examining molecular structure or protein sequence alone, enabling earlier elimination of ineffective drug candidates and saving substantial experimental costs.

Deep Learning + Swin Transformer for Breast Cancer Diagnosis

Medical imaging AI continues evolving. This framework combines Swin Transformer (a vision AI architecture) with dual attention mechanisms to more accurately identify cancer from mammograms. The key is it captures not just local features but global patterns, reducing missed diagnoses. For early breast cancer screening, this AI-assisted diagnostic tool could become doctors’ “second pair of eyes.”

Sequential Data Mining Predicts Early Dementia Diagnosis

Early dementia symptoms are subtle—by the time they’re obvious, it’s often too late. This research uses sequential data analysis (patient medical records and test results over time) to train AI models predicting dementia risk ahead of time. This approach beats single-point testing because it captures “symptom evolution trajectories.” If scaled up, it could enable earlier intervention for more patients.

Novel Benzimidazole Compounds Inhibit Cholinesterase, Potential Alzheimer’s Treatment

One Alzheimer’s treatment strategy is cholinesterase inhibition to boost acetylcholine levels in the brain. A research team synthesized novel benzimidazole-alkanesulfonate compounds that in vitro experiments and computational simulations show effectively inhibit cholinesterase. Though early-stage, these small-molecule drugs penetrate the brain more easily than antibodies—worth watching.

Cyclic Peptide Space: Optimizing Sequence Selection with ESM-2 Language Model

Cyclic peptides are hot candidates for next-generation drugs, but designing them is tough—the chemical space is enormous. A research team used the protein language model ESM-2 to construct “cyclic peptide space,” mapping each peptide to high-dimensional vectors, then uniformly sampling this space to select initial sequences. Results beat random selection by far—when designing β2-microglobulin-binding peptides, finding excellent candidates was noticeably faster. This method could become standard in AI drug design.

📌 Worth Watching

[Research] DeepMind’s New Headquarters First Floor Opens “AI Exchange Center” - Open to the public for AI exhibitions later this year

📊 More Updates

🔮 AI Life Science Trends Forecast

AlphaFold 4 Official Release

• Predicted Timeline: Q2 2026
• Prediction Confidence: 70%
• Reasoning: Today’s news on DeepMind’s New Headquarters Platform 37 Unveiled + DeepMind’s historical pattern of major releases in spring/summer, with the new headquarters launch signaling momentum