Ming-Ming Zhou, PhD, insisted in his New York City office, as if he overlooked the cloudy Central Park.
Zhou began his teaching career in 1997 and is currently a professor of physiology and biophysics at the Icahn School of Medicine in Mount Sinai. His lab designs compounds that regulate chromatin-mediated gene transcription for therapeutic applications. Zhou's energetic research in the chemical targeting of bromodomains, a set of proteins that recognize acetylated lysines in histones, opened the pharmaceutical field of bromodomains' drug discovery to address a wide range of cancer and inflammatory disorders.
Reflecting treatment research over the past 30 years, Zhou says that structure-based drug discovery has been converted to drug discovery that supports artificial intelligence (AI). According to Zhou, this paradigm shift is something that Mount Sinai's new AI Small Molecule Drug Discovery Center is tackling head-on.
Building a house for AI innovation
The AI Initiative, Professor at Avner Schlessinger, PhD, Pharmacological Sciences and Associate Director at Mount Sinai Therapeutics Discovery, leverages a computational approach that revelled New York's latest hub in April.
To expand access to AI-driven drug discovery, the center will provide hands-on training to the next generation of scientists through seminars, internship programs and drug discovery hackathons, promoting AI-centric research collaboration with pharmacy companies, biotech companies and academic institutions.
“Drug discovery is an inefficient process. One of the number one limiting factors is inadequate communication, interaction, or thinking outside the box,” Zhou said. gen. “This centre is a way to unlock new ideas and technologies that will help people connect and address this limitation.”
In contrast to traditional drug discovery, which relies on slow, resource-intensive experimental workflows, AI models trained with a vast dataset of molecular structure and biological activity can predict the properties of new compounds prior to synthesis.
Mount Sinai's center focuses on three core areas: designing new drug-like molecules using produced AI, optimizing existing compounds to enhance efficacy and safety, and predicting drug target interactions to reuse known drugs or natural products for new indications.
“I was trained with AI and machine learning long ago before it cooled down,” laughed Shressinger as he dodged a New York taxi while on a tour of the Mount Sinai campus. “But now is especially good time to improve the model of a real solution using Mount Sinai datasets and experts.”
As a medical school embedded in the hospital system, the Sinai Mountain community emphasizes the impact on patient care. Many research projects have a highly translational focus, ranging from target identification of Alzheimer's disease to the development of machine learning algorithms, to predicting the pathogenicity of mutations based on patient data.
Dr. Malta Filizola is a professor and dean of the Graduate School of Biomedical Sciences in Mount Sinai, leading the Center's graduate education activities, highlighting the need for interdisciplinary education to create the next wave of AI innovation.
“We have created an infrastructure to increase the visibility of AI training here in Sinai and provide practical experience in research programs that are directly related to improving human health,” she said. gen.
Please show me the data
Historically, structure-based drug discovery has been primarily fueled by the Protein Data Bank (PDB), a published data set that houses more than 200,000 entries for experimentally determined protein and nucleic acid structure data collected by researchers for over 50 years.
PDBs are powerful resources for advances in AI, such as Alphafold's Nobel Prize for chemical-earning protein structure prediction algorithms, but many new drug targets are outside the PDB, which motivates many AI biotechnology to invest in their own data generation. Much of this unique industry data remains locked and keyed.
“A key issue for parties building and innovating new model architectures is the inability to benchmark their own data. The validity of industrial-grade research cannot be assessed.” gen. “Access to industry data for benchmarking is a huge value-added for everyone building the model.”
Apheris is a startup focused on enabling governance, private, and secure access to data for machine learning. In March, the Berlin-based company announced an initiative with the AI Structural Biology (AISB) Consortium. This tweaks OpenFold3, a protein structure prediction algorithm developed by Dr. Mohammed Alkuraisi, a professor of systems biology at Columbia University.
This collaboration evaluates and refines OpenFold3 to predict the 3D structure of molecular complexes focusing on the interaction of small molecule proteins and antibodies and antigens for biological discovery. As of May, the list of participating drug developers has expanded to include AstraZeneca, Bohringer Ingelheim, Sanofi and Takada.
Pushing open source code
Other scientists aim to make AI molecular models widely accessible and advance collaboration. In June, researchers at the Massachusetts Institute of Technology (MIT) Jameel Clinic for Machine Learning in Health announced the open source release of Boltz-2. It predicts molecular binding affinity with new rates and accuracy to democratize commercial drug discoveries.
The Boltz-2 is available under a highly acceptable MIT license. This allows commercial drug developers to use the model internally and apply their own data. The work was done in collaboration with Recurring, a Salt Lake City-based artificial intelligence (AI) drug discovery company that was combined with Excientia last year. The MIT research team was led by Dr. Regina Barzilay, a well-known professor at MIT's AI and Health.
The Boltz-2 is the answer to the community's protests over the limited accessibility of Alphafold3. Nature In May 2024, Google Deepmind and Isomorphic Labs do not involve open source code. Alphafold 3 expanded its protein structure prediction tool to a wide range of biomolecule interactions, including small molecules, DNA, RNA, and more, providing a powerful next step in drug discovery.
However, the omission of the code has prevented other scientists from replicating the results of the publication, preventing the use of the model in their own research efforts, leading over 1,000 scientists to sign protest letters calling for transparency in the Alphafold 3. To address the protest, the AlphaFold 3 developer released the code after six months under a non-restrictive non-commercial license Nature Published.
Dr. Anshul Kundaje, an associate professor of genetics and computer science at Stanford University, wrote in a letter sent to Nature I posted on social media platform X, and although commercial entities have no obligation to open source or share their products, “this does not mean bypassing standard standards for what constitutes peer reviews and verifiable scientific publications. Nature The articles published as peer-reviewed are actually advertisements, at best white paper. ”
Back at MIT, Corso said the biggest reward for releasing Boltz is seeing community gatherings behind the open source project.
“When it seemed inevitable that a closure model like Alphafold 3 would dominate the field, many researchers in academia and the industry decided to contribute to open source projects like Boltz to build new features and make them available to everyone,” Corso said. gen.
Lift all the boats
Alphafold 3 has made progress in accurate prediction of molecular complex structures; In Silico Binding affinity calculations were not shown (officially) by deep fine and isomorphic labs, as achieved by Boltz-2. Binding affinity is an important drug discovery metric that measures the strength of interactions between a drug and its target and can determine candidate progress through a development pipeline from hit discovery to lead optimization.
From an accuracy perspective, Boltz-2 was the leading affinity performer in two experiments assessing the latest model of structural biology in a critical assessment of the protein structural prediction 16 (CASP16) competition in December 2024. In speed, Boltz-2 is reported to calculate binding affinity values in just 20 seconds, 1,000 times faster than the current physics-based computational standard, Free Energy Perturbation (FEP) simulation.
Dr. Najat Khan, Chief R&D Officer and Chief Commercial Officer of Repursion, said the open source release of Boltz-2 will “lift all the boats” in advancing integration of technology, biology and chemistry.
“Binding affinity was the core of starting and ending treatment, and was the fundamental problem that many of us have been trying to tackle. [with]Khan said.
In May, the re-regression said it would terminate development of four of the 11 pipeline programs and pause the fifth program. The company looks forward to applying Boltz-2 to future discovery candidates.
Unique restrictions remain the reality of commercial interests, education, data partnerships, and open source modeling, moving forward to promote a culture of collaboration. Time will tell whether a new AI drug discovery paradigm will become one of the true democracies.
