From Physics to Longevity: How Gero’s ProtoBind-Diff Tool Reads Nature’s Hidden Drug Library

By Allison Proffitt 

July 9, 2025 | Peter Fedichev, the CEO and co-founder of Singapore-based longevity biotech company Gero, has an unconventional theory: somewhere in our genome lies a hidden library of every drug that could ever work against human disease. His company’s latest AI tool, published on bioRxiv last month, aims to crack that code—and share the results. 

“If a more advanced alien species could read our genome, maybe in a thousand years from now, that alien civilization would read the genome and would be able to predict not only the proteins as AlphaFold is doing right now, but also the set of all the active molecules from the sequences,” Fedichev told Bio-IT World. 

It’s a bold claim from a physicist-turned-longevity-researcher whose fascination with non-aging mammals—from naked mole rats to certain bats to whales—led him to switch fields and co-found Gero a decade ago. (The company is based in Singapore, where the government is very supportive of longevity research.) Now, his team has developed a tool they believe could be a breakthrough approach to one of drug discovery’s most expensive bottlenecks: finding small molecules that bind to protein targets. 

ProtoBind-Diff is a structure-free masked diffusion model that generates molecules conditioned on protein sequences via pre-trained language model embeddings, Fedichev explained. “Trained on over one million active protein–ligand pairs from BindingDB, ProtoBind-Diff generates chemically valid, novel, and target-specific ligands without requiring structural supervision,” he and his co-authors write in the BioRxiv preprint.  

Nature’s Trial-and-Error Record 

Traditional drug discovery relies heavily on screening vast libraries of compounds, a process that can take up to 18 months and cost millions of dollars—particularly risky when targeting completely novel pathways. Computational approaches promise to speed this up, but modeling the quantum mechanical dance of molecules interacting with moving, breathing proteins remains “hell on earth,” as Fedichev puts it. 

Gero’s approach sidesteps this complexity entirely. Instead of trying to predict molecular interactions from first principles, their AI system assumes that evolution has already done the hard work. 

“Evolution cannot plan. We know that evolution has no plan. But evolution has a record of all trials and errors,” Fedichev explained. “What if nature actually experiments with small molecules for the proteins nature operates with? And what if nature optimizes proteins in such a way that they recognize certain chemical classes and those chemical classes are used in order to control biology?” 

Similar to how image generation models create pictures from text descriptions, Gero’s system generates drug-like molecules from protein sequences. The key premise: nature’s successful experiments—the molecules that actually work—are somehow encoded in our DNA. 

Emergent Understanding 

The results surprised even the researchers. When they trained their network on known examples of biologically active molecules, then asked it to generate new compounds based only on protein sequences, something remarkable emerged. The AI developed what Fedichev calls “an emergent understanding of 3D geometry.” 

“The attention layers of our networks actually read the active sites in the sequences of the proteins,” he said. “And they see also the parts of the molecules that interact with them on the graph representation of the molecule.” 

They benchmarked the results against structure-based models and found that ProtoBind-Diff performs competitively in docking and Boltz-1 evaluations and generalizes well to challenging targets, including those with limited training data. 

“It turns out that the model that has never seen a structure generates molecules that appear to be active, according to the state-of-the-art computational methods,” Fedichev noted. 

In the preprint paper, Fedichev and his co-authors predict: “This sequence-conditioned generation framework may unlock ligand discovery across the full proteome, including orphan, flexible, or rapidly emerging targets for which structural data are unavailable or unreliable.” 

Open Science Approach 

In an era where many AI companies closely guard their models, Gero is taking an open approach. The company’s preprint details ProtoBind-Diff and Gero plans to make both the code and trained model weights freely available. 

“We are believers in reproducibility in science,” Fedichev emphasized. “It’s very important because in many cases when you have all these AI algorithms, you never know what was in the training set, what was not, how new are the molecules, do they just memorize or they actually predict something new.” 

The rollout is planned in stages. The team has already released scripts for generating molecular structures, with the full model code coming within weeks. Soon after, they plan to launch a user-friendly interface allowing biologists without machine learning expertise to generate molecules for their targets. 

The ultimate test, of course, will be experimental validation. Fedichev confirmed that wet lab testing of the AI-generated compounds is currently underway, with results to be added to the current preprint and included in a future journal submission. 

Beyond the Algorithm 

For Gero, this tool represents just one piece of a larger puzzle. The company’s primary focus remains identifying new biological targets for aging and age-related diseases, rather than becoming a drug design software vendor. 

“The true leverage of AI in medicine and biomedicine and biotech is target discovery, not actually the drug discovery,” Fedichev argued. “The competition here is what to do, what are the targets that modify human disease and aging.” 

This philosophy stems from Fedichev’s study of “negligible senescence”—species that don’t show traditional signs of aging. His realization that numerous mammals, from bats to certain whales, maintain constant disease risk throughout their lifespans convinced him that aging is not inevitable. 

“Humans believe that humans are the best in nature,” he said. “But [they are] not.” If evolution found ways for other mammals to avoid aging, Fedichev believes humans can too—with a little help from AI.