Artificial intelligence (AI) is rapidly transforming the way science is conducted across industries, from aerospace and energy to pharmaceuticals. In drug discovery, AI is enabling researchers to move beyond trial-and-error experimentation and toward more predictive, structure-aware approaches. At the same time, regulators such as the FDA are signaling openness to AI, modeling, and simulation as tools to modernize how new therapies are designed, evaluated, and eventually approved.

DDN spoke with Arman Zaribafiyan, Head of Strategic Alliances at SandboxAQ, an AI and quantum technology company that has recently launched Structurally Augmented IC50 Repository (SAIR), an open dataset of over 5 million protein-ligand structures paired with experimental binding affinities. This dataset provides a major resource for researchers to advance AI models in drug discovery, significantly enhancing the speed and accuracy of binding affinity predictions. Arman shared his perspective on the opportunities, challenges, and regulatory considerations shaping the future of AI-enabled drug development.

The FDA and other regulatory agencies are increasingly exploring AI, modeling, and simulation to modernize drug discovery and evaluation, and potentially reduce animal testing. From your perspective, what are the biggest opportunities and challenges in this shift?

The explosive growth of real-world and in silico data, such as single-cell RNA sequencing and physics-based simulations at scale, combined with advances in AI infrastructure and model architecture, is finally bringing AI-informed clinical outcomes within reach.

Aside from the direct effects of cost and time reduction, the biggest opportunity is to bring AI-informed clinical predictions earlier in the drug design process. This enables triaging targets, prioritizing chemistries, and focusing wet-lab work where it matters, or skipping it altogether once AI models are validated with regulatory bodies in a particular context of use.

The main challenges are trust and generalizability. AI predictions in this field must be trustworthy, reproducible, explainable, and capable of generalizing well beyond what they are trained on. Both challenges can be addressed with larger volumes of high-quality and diverse data, as well as rigorous benchmarking frameworks — two areas we are beginning to tackle head-on with SAIR as our first publicly available dataset.

How is AI currently reshaping drug discovery and which areas do you think stand to benefit most from computational approaches?

AI is everywhere in drug discovery. You see its relevance across the entire life sciences value chain. It is shifting discovery from intuition-driven heuristics to structure-aware, physics-informed design. The biggest near-term gains are in hit triage, lead optimization, and early safety, where fast AI surrogates can prune vast search spaces and anticipate off-target risks before synthesis.

The most exciting near-term advance is structure-aware AI. With the rise of protein-folding and co-folding models, physics-informed systems can predict on- and off-target interactions at inference speed — milliseconds to seconds — moving computer-aided design beyond static snapshots and toward accurate predictions of interaction dynamics.

We released the SAIR dataset precisely to enable every scientist to build, validate, and benchmark these structure-aware models for drug discovery.

What lessons from other sectors — like aerospace, defense, or energy — can the drug discovery industry apply when integrating AI and predictive modeling?

The key lesson is to avoid AI-washing the value chain with LLMs and instead take a first-principles approach to the areas where machine learning can truly accelerate the design of better, more valuable products for critical parts of the supply chain .
- Arman Zaribafiyan

The cure for cancer won’t be found in Wikipedia, nor are the solutions to many problems across the energy, defense, and aviation industries. The common thread across these industries is that large language models (LLMs) alone aren’t enough to unlock AI’s full power for scientific discovery. Despite their exciting applications, LLMs are unable to address the core challenges these industries face.

A different AI, trained on quantitative data and grounded in physics-based equations, is needed to design sustainable aviation fuels, build better EV batteries, lightweight armored vehicles, decarbonize processes, produce green hydrogen, and turn oil-and-gas waste streams into high-value products. The key lesson is to avoid AI-washing the value chain with LLMs and instead take a first-principles approach to the areas where machine learning can truly accelerate the design of better, more valuable products for critical parts of the supply chain.

SandboxAQ recently released SAIR, a 5 million+ protein-ligand dataset linking molecular structure to drug potency. How do you envision researchers and companies using this dataset with AI to accelerate drug discovery?

SAIR provides a structural backbone for discovery, with more than five million co-folded protein-ligand complexes paired with experimental half-maximal inhibitory concentration (IC₅₀) labels, released under a permissive Creative Commons Attribution (CC BY 4.0) license for immediate use in both commercial and academic settings. Using this foundation, researchers can train and fine-tune structure-aware potency and affinity predictors, build ultra-fast docking and scoring surrogates, and extend predictions to proteins that lack experimental structures. Approximately 40 percent of the proteins in SAIR didn’t have a Protein Data Bank entry but still achieved a 97 percent pass score on PoseBusters checks. PoseBusters is a Python-based tool that evaluates the physical plausibility and chemical consistency of predicted protein-ligand structures. This means SAIR also serves as a validated, common test bed for rigorous, head-to-head benchmarking, enabling fair comparisons and faster iteration.

Can you share any case studies or early examples of how companies or researchers are already using SAIR, or how you hope they might use it in the near future?

Only a few days after its release, SAIR became the number one trending dataset on Hugging Face, with hundreds of downloads in the first days and strong interest from leading academic centers, hospitals, clinical research institutes, and industry — from pharma and biotech to AI leaders like NVIDIA.

We built SAIR to enable training and benchmarking of structure-aware frontier models, especially affinity predictors that estimate drug potency directly from 3D structure and close a long-standing gap in rational drug design.

Now we see that early adopters are already taking creative paths. Technetium Therapeutics, for example, is using SAIR to retrain its agentic platform to better prioritize high-potency chemistries. At Texas A&M’s School of Medicine, researchers are converting SAIR protein and ligand crystallographic information files into language representations. These representations are used to pretrain a foundation model that can generate molecules directly within a target’s binding site.

This is the power of open source and exactly the trajectory we hope to see: teams using SAIR to tighten the design-make-test loop and move more of discovery from the wet lab to the workstation.

Looking ahead, how might datasets like SAIR influence regulatory practices, particularly around modeling and simulation as alternatives to animal testing?

For years, in silico evidence stalled at triage because models lacked sufficient high-quality training data that directly linked 3D molecular structure to real drug potency, efficacy, and more. SAIR closes that gap by pairing millions of co-folded protein-ligand complexes with experimental IC₅₀ labels. With datasets like SAIR, structure-aware models can deliver faster, more accurate predictions of potency, selectivity, and polypharmacology. Combined with focused in vitro calibration and explicit uncertainty estimates, these models can move from simple ranking tools to credible inputs for regulatory decision-making.

The impact on animal testing is two-fold: reduction and refinement. Fewer compounds advance to in vivo, and those that do enter with tighter dose ranges, clearer off-target risks, and stronger mechanistic hypotheses — shortening studies and reducing cohorts. For certain low-information screens, a transparent, benchmarked modeling package combined with in vitro confirmation can replace animal testing, reserving animal work for specific residual uncertainties.

Open sourcing a validated, auditable dataset like SAIR makes resulting evidence easier for regulators to trace and compare, accelerating the shift from animal-heavy discovery to reproducible, structure-grounded modeling and simulation.

From your perspective, what regulatory updates or frameworks would help AI and modeling approaches become more widely adopted in drug development?

The industry needs a simple, risk-tiered standard for model trust, something as intuitive as Good Laboratory Practices (GLP), but for AI. We need to define how sponsors demonstrate provenance from data to model to decision, require calibrated uncertainty, and set clear expectations for low-risk versus decision-critical uses. Give the field one playbook and a set of clinically relevant benchmarks, and you will get speed without sacrificing rigor.

Pair that with shared yardsticks so everyone is judged on the same field. Open, auditable benchmarks, such as SAIR, allow reviewers to compare methods and trace claims. With those two pieces in place, regulators can enable in silico bridging, where validated models plus targeted in vitro evidence replace some early animal work where the context of use is validated. That is how AI becomes decision-grade evidence, and not just a promising demo.

This interview has been condensed and edited for clarity.