AI Continues to Make Strides on Therapeutics and Research Progress

By Irene Yeh 

April 16, 2025 | During the final plenary keynote of this year’s Bio-IT World Conference & Expo, we heard from Tris Dyson, Challenge Works, and Jeffrey Rothstein, Johns Hopkins University; Robert Green, Harvard Medical School; and Justin Scheer, Johnson & Johnson Innovative Medicine. They all gave presentations that highlighted artificial intelligence (AI) and machine learning’s (ML) ever-growing roles in research and healthcare, as well as the impact they have on the future of patient treatment, drug development, and more.  

The Longitude Prize on ALS 

Tris Dyson and Jeffrey Rothstein kicked off with a call for applicants for the Longitude Prize on ALS, an international initiative that brings together computational biologists, neurodegenerative researchers, and AI biotech organizations to develop novel therapeutic targets in amyotrophic lateral sclerosis (ALS), also known as Lou Gehrig’s disease.   

“There’s a real opportunity to develop new therapeutics if we can integrate all of that science properly,” said Rothstein. “AI is clearly a tool that can advance our understanding of the biology and then help us develop new drugs.” 

Officially launching in June 2025, the Longitude Prize on ALS’ goal is to identify and validate drug targets and pathways relevant to ALS’ progression, as well as raise awareness on the disease and the solutions needed to treat it. ALS has large sets of open source data that come in from multiple sources around the world (mostly academic), but it has not been compiled and organized yet. “This prize is oriented toward aggregating all of that data, making it available,” Rothstein continued. 

The Longitude Prize encourages AI experts and ALS researchers to partner and identify targets with high potential or AI capabilities within the institution supported by the prize. Dyson explained that applicants who are accepted will receive help from the Johns Hopkins network and other organizations to connect AI biotech companies and the ALS research community. 

Genomic Sequencing to Prevent Risks 

Robert Green shared the impact of the BabySeq Project (BabySeq), which pioneered integration of genomic sequencing into newborn and childhood screening. Since 2015, BabySeq has enrolled over a thousand families and sequenced their children to check for genetic risk variants. 

Despite early concerns about psychological distress, medical harm, and ethical standards, BabySeq has proven to be an effective way to help parents know early about medical risks in order to intervene sooner than later. “Of course, they’re not happy if they find their child has a risk variant, but they did not have catastrophic psychological distress,” said Green. “They were grateful for the information.” 

Green shared a video in which several participants mentioned how they learned their child was the carrier for heart conditions, cancer, and other diseases, which allowed them to take early steps to minimize the risks and chances of the conditions developing. The video also featured parents who learned that they were carriers of certain genetic risks, which allowed them to take precautions to prevent the diseases from manifesting.  

However, financial support for projects like BabySeq has become uncertain for Green and his team. “I come to you this morning with a heavy heart for so many reasons because my work has been supported by NIH through the years, Green said. He noted that, in particular, the Secretary’s Advisory Committee on Hereditary Diseases was dissolved in early April. 

Despite these difficulties, Green mentioned two projects he co-founded and is currently working on that support genomic sequencing in newborns. One of them is the International Consortium on Newborn Sequencing (ICoNS), which focuses on AI to inform clinical and public health researchers and implement genomic screening in newborns by aggregating and harmonizing scientific evidence and resources. With over 500 members (50 from 50 different countries), the drive to make genomic screening in newborns available and accessible is still active.  

“One of the areas we’re really missing is the computational and informatics input,” Green mentioned. “We could use a committee on AI.” 

The second project is Nurture Genomics, a startup that focuses on creating an AI-powered, physician-ordered clinical platform that is integrated into pediatric practice. The purpose of the platform would be to identify children with treatable genetic conditions and then equip those families and pediatricians with a knowledge base that could help them. 

“This is part of a vision for lifelong genomic medicine, whether it’s through some sort of revolutionary change in our healthcare system or through platforms,” said Green. 

Delivering the Best Therapeutic Candidates with Machine Learning 

While AI and ML receive criticism for being all hype, they have provided valuable support for drug development and research. “We truly believe and have seen with our own experience that AI/ML combined with data at scale can have a huge impact on the progress of our therapeutic discovery programs,” said Justin Scheer. 

Technology has rapidly evolved over the years. In the early 2000s, structure-based drug design, informatics, bioinformatics, and cheminformatics were considered niche sciences that were connected to only certain parts of drug discovery. In the 2010s, questions about the broader potential for predictive models emerged, with a few pilot programs beginning. Today, we have new, more sophisticated approaches and multimodal methods being considered. 

“Within the next year or two, our ambition is to take what we’re doing from discovery— driving the design-make-test cycle that’s largely in the hands of wet lab scientists—and bring this into the automation space, connecting both AI and automation and having intelligent lab automation drive a lot of the process,” said Scheer. Though it took about two decades, the industry is currently in a phase of fast expansion and change. 

Scheer brought up a few of the obstacles encountered over the years. First, making the AI tool able to bring in large volumes of different data and merge them, as well as actually use said data. The second challenge is determining whether different types of data can be used. The architecture of the AI model and the data must both be reliable. The final challenge is keeping up with rapidly changing, dynamic portfolios of different projects. 

However, it seems like these challenges have already been addressed by some in the industry. Scheer mentioned a few examples, starting with Sentinel, an early warning system that provides indication of problematic outcomes and runs in the background by profiling all compounds and giving a risk score. Sentinel allows researchers a chance to reassess their approach and how to navigate their projects, giving them space to figure out how to properly use data, what their objectives are, and more. 

Scheer also brought up SMARCA2, a protein that is a potential therapeutic target for cancer treatment, and how the project team learned how to identify the right compounds in it. The team understood they had to use an effective in silico discovery strategy and determine where they wanted to go with the discovery process. They realized that potency was essential and had to improve solubility. After they established their strategy, the team used ML tools to add predictive filers to identify potent SMARCA2 compounds that met all the criteria. Next, they applied selectivity filtering, which helped identify a structure-based signal that could be used with compound triaging. Finally, they were able to identify the compounds they were looking for in this process. 

Finally, Scheer mentioned a different project where the team developed technology that could predict from chemical structure to human predicted dose. The team distilled and organized large volumes of data and combined them with PK data from in vivo preclinical studies and human studies to create an AI/ML model that triages compounds. By doing so, the project team could predict what level of dose that compound will need to be effective in humans. The generated PK data on the compounds could then be used to refine the AI/ML model for better accuracy. Scheer also used Targeting Mask 3, a protease target, as an example. The team behind that project was able to notice a correlation between the predicted PK and actual PK, allowing them to use those models to go straight from compound synthesis to animal experiments without running the in vitro assays in between. This tool saved the researchers 2-3 weeks, which made them move through the cycle faster and make decisions quicker. 

The above-mentioned examples showcase the efficient utilization of data and the effectiveness of having reliable AI and ML tools. Furthermore, they also highlight how to use data and ensure the reliability of AI and ML tools: apply an effective strategy, understand and determine what to look for, and take a step back instead of rushing in. 

But with these challenges addressed, a new question has emerged. Where do we go from here? 

According to Scheer, there needs to be focus on driving design. “We want to combine different types of data to have much better models to be more predictive and drive design at a much higher level.” He also mentioned testing and its role in generating much better models.  

The issue isn’t the availability of motivation, resources, and people. According to Scheer, it seems like the concern centers around the data, models, and technology “increasing at the same time.” In other words, the industry is now dealing with “a multi-pronged challenge of dealing with an explosion on all fronts at the same time.” 

“We need to think hard about how we’re going to maximize the impact of AI and realize its promise in drug discovery and development—and ensure that we have the technical infrastructure and culture in place to allow us to capture the opportunity,” he elaborated. 

Scheer suggested a few strategies to help with managing and maintaining the rapid expansions. The first is to completely abandon the service model and ticket model, improve transparency between cross-functional teams, and break away from silos. This would help improve efficiency, as well as keep up with platform investments and create platforms that can rapidly and seamlessly bring in new types of data or models.  

The second is to rethink budgeting for technology. “Technology investment has to be thought of in the same way we think about synthesizing compounds… the way we think about investments in day-to-day operations of drug discovery.” This way, teams can get to the right level of resources to address challenges and make the most of the data generated. 

Finally, there needs to be a return to teams. A reshaping of team modeling infrastructure can help with keeping things organized and improve transparency. 

“We’re just at the cusp of what is possible,” concluded Scheer. “We will continue through this approach to have impacts on our outcomes, including class-leading molecules, faster development, discovery timelines, better results in the clinic, and better results for patients.”