The Chan Zuckerberg Biohub has unveiled an artificial intelligence system capable of designing therapeutic molecules in days rather than years, a breakthrough that could reshape drug discovery timelines for research institutions worldwide. Portuguese labs now have access to this globally available platform at no cost.
Why This Matters
• Speed: Protein-based drug candidates now designed in days instead of months or years, cutting preclinical timelines by orders of magnitude.
• Open Access: The entire platform—ESM Atlas, ESMFold2, and ESMC language model—is free and open-source, allowing smaller biotech startups globally to compete with pharma giants.
• Scale: The atlas maps 6.8 billion protein sequences and predicts 1.1 billion 3D structures, dwarfing Google DeepMind's AlphaFold database by 800 million entries.
• Proven Results: Early lab tests show 36-88% success rates for mini-binders and 15-29% for antibody-format molecules targeting cancer and autoimmune pathways.
What the Technology Actually Does
The Chan Zuckerberg Biohub—a nonprofit research organization founded by Facebook co-founder Mark Zuckerberg and physician Priscilla Chan—released what it calls a "world model of protein biology" this month. The system functions as a scientific search engine that can predict protein structures, design new therapeutic binders, and map evolutionary relationships across the entire tree of life.
At its core, the model comprises three components. ESMC is a language model trained on 2.8 billion genetic sequences, representing proteins as virtual entities the AI can manipulate. ESMFold2 predicts three-dimensional protein structures and their interactions with unprecedented accuracy, while the ESM Atlas organizes 6.8 billion sequences into a relational map that reveals connections missed by conventional databases.
The practical upshot: researchers can now input a disease target—say, a receptor implicated in tumor growth—and receive dozens of candidate binder molecules within hours, complete with structural predictions and binding-affinity estimates. In traditional drug discovery, generating even a handful of candidates typically requires 6-12 months of iterative lab work.
Head-to-Head with AlphaFold
Google DeepMind's AlphaFold famously solved the protein-folding problem in 2020, achieving near-90% accuracy in predicting how amino acid chains collapse into functional 3D shapes. The third iteration, released in May 2024, expanded to predict interactions with DNA, RNA, and small-molecule ligands.
The Biohub's ESMFold2 now claims to surpass AlphaFold 3 on several key benchmarks, particularly for protein-protein interactions and the notoriously difficult task of predicting antibody-antigen binding poses. When given the same evolutionary input data (multiple sequence alignments), ESMFold2 outperforms AlphaFold's latest version in head-to-head tests, according to the Nature report cited by AFP.
More importantly for the research community worldwide, the Biohub system is entirely open-source with no commercial restrictions, whereas AlphaFold remains a proprietary Google product with a freely accessible database but closed underlying code. For biotech ventures globally, including those in Portugal, this distinction could level the playing field against established pharmaceutical R&D divisions.
Early Lab Victories: Cancer and Immune Targets
Biohub researchers put the model through its paces by designing binders for five high-value targets in oncology and immunology: EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45. The first two are growth-factor receptors that drive tumor proliferation; the next two are immune checkpoints that cancer cells hijack to evade detection; the last is a signaling regulator in immune cells.
The AI-generated binders achieved validation rates of 36-88% for compact "mini-binder" formats and 15-29% for full antibody architectures when tested in lab assays. Crucially, molecules targeting PD-L1 successfully restored T-cell signaling in dish experiments, blocking the same pathway exploited by approved checkpoint inhibitors like pembrolizumab.
These are not yet approved therapies—clinical trials remain years away—but the preclinical validation rates far exceed industry norms. Conventional high-throughput screening campaigns often yield hit rates in the low single digits.
In a separate demonstration, the model generated stable enzyme variants capable of degrading microplastics, with 78% of designs validated in follow-up experiments—a result that has impressed independent scientists quoted in Nature.
The 500M USD Virtual Biology Initiative
The atlas release is the centerpiece of Biohub's Virtual Biology Initiative, a five-year, 500 million USD (approximately €465 million) program to build open-access, AI-accelerated databases for human cell biology. The effort brings together a coalition that includes the Allen Institute, Arc Institute, Broad Institute (MIT/Harvard), Wellcome Sanger Institute (UK), the Human Cell Atlas consortium, Sweden's Human Protein Atlas (SciLifeLab), and chipmaker NVIDIA.
The goal is to create predictive models of the human cell—digital twins that can simulate drug responses, immune reactions, and disease progression in silico before a single lab animal is dosed. Priscilla Chan, writing in Time magazine, framed the initiative as a way to "finally offer the scientific community tools to answer the most difficult and urgent questions about human health," including full simulation of the immune system.
For Portugal's research sector, this represents an opportunity. Research institutions now have free access to computational infrastructure that would otherwise require tens of millions in licensing fees and hardware investment.
What This Means for Portuguese Researchers
Portuguese institutions active in oncology, immunology, and rare-disease research can immediately integrate the ESM Atlas and ESMFold2 into early-stage discovery pipelines. The open-source nature means no institutional subscriptions, no per-query fees, and no export restrictions on commercial applications.
Researchers and biotech teams in Portugal working on antibody therapeutics stand to benefit particularly, as the model excels at antibody-antigen predictions—a task where even AlphaFold struggled until its most recent version.
Practically speaking, a lab in Portugal can now run the same protein design experiments as a Novartis or Roche R&D team, provided they have access to standard computational infrastructure. The democratization of advanced AI tools offers potential acceleration for research institutions engaged in life sciences innovation.
The CRISPR Connection and Evolutionary Surprises
One of the atlas's most intriguing findings emerged not from drug design but from pure evolutionary biology. The model identified structural homology between bacterial CRISPR-Cas gene-editing proteins and a previously uncharacterized gene-editing protein found in fungi, yeast, and algae. The two families diverged hundreds of millions of years ago, yet the AI detected shared architectural motifs invisible to conventional sequence-alignment tools.
This discovery hints at a broader application: using the atlas to uncover "dark matter" in the proteome—the roughly 60% of catalogued protein sequences whose biological function remains unknown. For agricultural research, microbiology, and industrial biotech, this could unlock enzymes and pathways with applications far beyond human medicine.
Skepticism and Validation Gaps
Not all reactions have been uncritical. Independent structural biologists interviewed by Nature acknowledged the impressive validation rates but cautioned that in vitro binding does not guarantee in vivo efficacy. A molecule that binds PD-L1 in a dish may fail to reach tumors, trigger immune reactions, or degrade too quickly in bloodstream conditions.
Moreover, the model's training data—2.8 billion sequences—includes vast swaths of microbial and environmental DNA from metagenomic surveys, where annotation quality varies wildly. Some researchers worry that noise in the training set could propagate errors in edge-case predictions, particularly for poorly characterized protein families.
Still, the sheer scale of the resource and the early lab results suggest the platform will become a standard tool in computational biology, much as AlphaFold did after 2020.
Accessing the Tools
The full ESM Atlas, ESMFold2 model weights, and ESMC language model are now available via the Biohub's open-science portal. Researchers can run predictions on their own hardware or use cloud instances; the organization has not imposed compute quotas or usage tiers.
For labs globally, including those in Portugal, the immediate action item is straightforward: download the models, test them against existing drug targets, and begin integrating AI-predicted binders into hit-to-lead pipelines. The technology is freely available to the scientific community worldwide, and early data suggests it works effectively.
