Modern biotech has the tools to edit genes and design drugs, but thousands of rare diseases are left untreated. According to executives from Insilico Medicine and GenEditBio, the missing ingredient over the years has found enough smart people to continue the work. AI, they say, has become a force multiplier that allows scientists to solve problems that have long been untouched by industry.

Speaking this week at Web Summit Qatar, Insilico’s president, Alex Aliper, outlined his company’s goal to develop “pharmaceutical superintelligence.” Insilico recently launched it “MMAI Gym” which aims to train generalist large language models, such as ChatGPT and Gemini, to perform as well as specialist models.

The goal is to create a multimodal, multitask model that, Aliper said, can solve many different drug discovery tasks simultaneously with superhuman accuracy.

“We really need this technology to increase the productivity of our pharmaceutical industry and solve the lack of labor and talent in that space, because there are still thousands of diseases without a cure, without any treatment options, and there are thousands of rare diseases that are neglected,” said Aliper in an interview with TechCrunch. “So we need more intelligent systems to solve that problem.”

The Insilico platform takes biological, chemical, and clinical data to generate hypotheses about disease targets and candidate molecules. By automating steps that once required legions of chemists and biologists, Insilico says it can sift through multiple design schemes, nominate high-quality therapeutic candidates, and even repurpose existing drugs — all at a reduced cost and time.

For example, the company recently used its AI models to determine whether existing drugs could be repurposed to treat ALS, a rare neurological disease.

But the labor bottleneck did not end with the discovery of the drug. Even if AI can identify promising targets or therapies, many diseases require interventions at a more fundamental biological level.

GenEditBio is part of the “second wave” of CRISPR gene editing, where the process moves away from editing cells outside the body (ex vivo) and towards precise delivery inside the body (in vivo). The company’s goal is to make gene editing a one-off and can be injected directly into the affected tissue.

“We’ve developed a proprietary ePDV, or engineered protein delivery vehicle, and it’s a virus-like particle,” GenEditBio’s co-founder and CEO, Tian Zhu, told TechCrunch. “We learn from nature and use AI machine learning methods to mine natural resources and find which types of viruses are associated with certain types of tissues.”

The “natural resource” Zhu refers to is GenEditBio’s vast library of thousands of unique, nonviral, nonlipid polymer nanoparticles — delivery vehicles designed to safely carry gene-editing tools into specific cells.

The company says the NanoGalaxy platform uses AI to analyze data and determine how chemical structures are linked to specific tissue targets (such as the eye, liver, or nervous system). AI then predicts which tweaks to a delivery vehicle’s chemistry will help it carry a cargo without triggering an immune response.

GenEditBio tests its ePDVs in vivo in wet labs, and the results are fed back to the AI ​​to refine its prediction accuracy for the next round.

Efficient, tissue-specific delivery is a prerequisite for gene editing in vivo, Zhu said. He argues that his company’s approach reduces the cost of goods and standardizes a process that has historically been difficult to scale.

“It’s like getting an off-the-shelf drug (available) for more patients, making drugs more affordable and available to patients around the world,” Zhu said.

His company is new received FDA approval to begin trials of CRISPR therapy for corneal dystrophy.

Combat persistent data problems

Like many AI-driven systems, biotech development will eventually run up against a data problem. Modeling the edge cases of human biology requires much higher quality data than researchers can currently obtain.

“We still need more ground truth data coming from patients,” Aliper said. “The data corpus is very biased in the Western world, where it was created. I think we need to have more efforts locally, to have a more balanced set of original data, or ground truth data, so that our models can also be better able to deal with it.”

Aliper said Insilico’s automated labs generate multi-layered biological data from disease samples at scale, without human intervention, which it then feeds into its AI-driven discovery platform.

Zhu said that the data needed by AI already exists in the human body, shaped by thousands of years of evolution. Only a small part of DNA directly “codes” for proteins, while the rest acts more like an instruction manual for how the genes work. That information has historically been difficult for humans to interpret but is increasingly available to AI models, including recent efforts like Google DeepMind’s AlphaGenome.

GenEditBio applies a similar approach in the lab, testing thousands of nanoparticles for delivery in parallel rather than one at a time. The resulting datasets, which Zhu calls “gold for AI systems,” are used to train its models and, in addition, to support collaborations with external partners.

One of the next big efforts, according to Aliper, is to create digital twins of people to run virtual clinical trials, a process he says is “still in its infancy.”

“We are at a plateau of about 50 drugs FDA approved every year every year, and we should see progress,” said Aliper. “There is an increase in chronic disorders because we are aging as a global population… My hope is that in 10 to 20 years, we will have more treatment options for the personal treatment of patients.”