Recent advancements in artificial intelligence are transforming drug discovery processes, enhancing predictions of small molecule binding affinity and cellular reprogramming. This evolution in AI applications is marked by the integration of extensive biological data and innovative model architectures. According to Arman Zaribafiyan, PhD, the head of product, AI simulation and platforms at SandboxAQ, the drug discovery field has yet to experience its “ChatGPT moment.” He emphasizes that solely depending on machine learning models trained on text data is insufficient for addressing the most complex medical challenges, stating, “The cure for cancer is not written in Wikipedia.”
SandboxAQ, a spinout from technology giant Alphabet, aims to harness large quantitative models (LQMs) for societal benefits, including drug discovery. Unlike large language models (LLMs) that focus on text data, LQMs are based on physical principles to simulate real-world systems.
The release of Google DeepMind“s AlphaFold31 and RoseTTAFold All-Atom in 2024 marked a pivotal moment, significantly advancing protein structure prediction abilities beyond simple peptide chains to encompass interactions with small molecules and nucleic acids. These models have accurately predicted the “pose” of ligands binding to target proteins, though the prediction of binding affinity—which measures the strength of these interactions—remains a critical goal. Achieving this capability could provide a less resource-intensive alternative to traditional experimental methods, thereby shortening discovery timelines and reducing costs.
In June 2024, a collaborative team from the Massachusetts Institute of Technology (MIT) and Recursion unveiled Boltz-2, an open-source model that delivers rapid and accurate binding affinity predictions, democratizing AI-driven small molecule drug discovery. Boltz-2 excelled in the December 2024 Critical Assessment of Protein Structure Prediction 16 (CASP16) competition, demonstrating a binding affinity calculation speed of just 20 seconds—significantly faster than the current physics-based standard, free-energy perturbation (FEP) simulations. Importantly, Boltz-2 operates under an MIT license that permits commercial developers to use it with proprietary data.
Zaribafiyan points out a significant challenge in AI-based drug discovery: the scarcity of data connecting protein-ligand structural complexes with pharmacokinetics (PK) and pharmacodynamics (PD). One proposed solution is to generate synthetic training data through computationally predicted structures. Following Boltz-2″s introduction, SandboxAQ and Nvidia announced the Structurally-Augmented IC50 Repository (SAIR), an open-access database that uses Boltz models to create protein–ligand structures linked to experimental drug affinity values. This repository includes over one million unique protein-ligand pairs and 5.2 million curated 3D structures derived from binding affinity databases such as ChEMBL and BindingDB.
While Boltz-2″s training set includes data from these databases, other research teams have begun integrating SAIR into their drug discovery initiatives. For instance, Technetium Therapeutics is employing SAIR to develop AI models aimed at identifying new drug candidates for oncological and immunological disorders. Similarly, researchers from Texas A&M University are utilizing SAIR to create foundational models that design novel ligands targeting specific protein pockets.
“Understanding the pose is crucial if you wish to modify the small molecule once binding is confirmed,” stated Ian Quigley, PhD, CEO of Leash Bio. He acknowledges the growing attention within the community toward predicting binding affinity.
Leash Bio focuses on addressing data gaps in small molecule drug discovery. Quigley notes that life science datasets often suffer from batch effects and technical noise, complicating the accuracy of biological predictions. He illustrated this with an analogy to a famous data collection of horse images that included a watermark, which misled AI models into incorrectly identifying horses based on the watermark rather than the actual content of the images.
By generating large, high-quality datasets that assess millions of small molecules against numerous protein targets, Leash aims to achieve robust predictive performance despite using relatively simple model architectures. In July, the company announced Hermes, a binding prediction model developed solely from internal data. Hermes, which does not rely on structural modeling, predicts binding likelihood based on amino acid sequences and SMILES representations of small molecules. Leash claims this model operates 200 to 500 times faster than Boltz-2 while providing improved predictive accuracy compared to other AI models. Following this, Leash introduced Artemis, a hit expansion tool that utilizes Hermes to explore chemical space around specific targets.
As the focus on small-molecule drugs persists, several experts are turning their attention to the complexities inherent in proteins to tackle therapeutic issues effectively. Simon Kohl, PhD, CEO of Latent Labs and a former researcher at Google DeepMind, is committed to constructing advanced models for biology, starting with the design of proteins from scratch. His team develops generative AI models to create new antibodies, refine existing enzymes, and advance genetic engineering.
In July, Latent Labs launched Latent-X, their inaugural frontier model for de novo protein design, achieving impressive binding affinities in the picomolar range by testing merely 30 to 100 candidates per target in laboratory experiments. This approach is a significant advancement over traditional drug discovery methods, which often require screening millions of random molecules for hit rates below one percent. Latent-X”s designs focus on therapeutically relevant mini-binders and macrocycles, demonstrating competitive binding affinities compared to leading protein design models such as RFdiffusion and AlphaProteo.
Kohl highlights the unique model architecture of Latent-X, which simultaneously models sequence and structure at an all-atom level. “We”ve released videos showcasing how the model creates specific hydrogen bonds and aromatic ring stacking. This end-to-end biochemistry generation allows us to develop superior molecules from the outset,” Kohl shared.
Latent-X is accessible via a web interface, enabling researchers without computational expertise to utilize the platform. Kohl underscores the importance of making this technology available without requiring extensive AI knowledge or infrastructure, calling it a true democratization of science.
In addition to protein design, Retro Biosciences is leveraging AI models for cellular reprogramming in aging research, pursuing diverse therapeutic strategies from cell therapies to small molecules. CEO Joe Betts-LaCroix, PhD, explained that the company sees itself as a portfolio investing in various approaches to achieve its mission. “Different modalities offer unique advantages and disadvantages, providing robustness to Retro as a single entity,” he stated.
In collaboration with OpenAI in August, Retro announced enhanced versions of the Yamanaka factors—four transcription factors that can reprogram adult somatic cells into induced pluripotent stem cells (iPSCs). Utilizing the GPT-4b micro model, tailored for protein engineering, Retro achieved over a 50-fold increase in the expression of stem cell reprogramming markers compared to wild-type controls in laboratory settings.
The model incorporates protein information through textual descriptions and protein interaction networks, allowing it to generate new sequences with specified properties. This adaptability is especially beneficial for both structured proteins and intrinsically disordered proteins, such as the Yamanaka factors, whose activity hinges on transient interactions with various binding partners.
From predicting small molecule binding affinities to advancing cellular reprogramming, the integration of large-scale biological data with innovative model architectures is driving the AI revolution in drug discovery. The future will reveal whether increasing computational power will further enhance therapeutic potential.
