Electronic Components
Stay ahead of the storm
Industrial Automation
Industrial Automation at your fingertips
Source: Deemerwha studio/stock.adobe.com
Artificial intelligence (AI) is reshaping the pace and scope of scientific discovery in medicine, environmental science, and physics, often by making stages in the research pipeline faster or by altering them with new processes that require significantly less time. Much of that shift relates to a new generation of AI architectures that changed how models learn from data and the depth of representations they can capture. As a result, the progression from hypothesis to results is much faster, making failed attempts less costly in fields where success is rare, such as drug discovery.
This blog explores how modern AI architectures are compressing research timelines across multiple scientific disciplines, where these advances are delivering measurable results today, and the challenges that must be addressed before they can be applied more broadly.
Older statistical machine learning (ML) in science depended on hand-built features. For example, chemists had to decide which molecular descriptors were important for drug design. However, deep learning models, such as transformer and graph-based architectures, instead learn directly from raw sequences, images, or structured graphs. This shift is significant because for many scientific problems, the limiting factor was knowing which features to encode manually. Learning from raw data also enables automated analysis of unstructured inputs like images, allowing experts to process volumes of data that would otherwise be impractical to handle manually. Protein folding is a clear example of a problem that challenged researchers for decades before AI models demonstrated that protein structures could be predicted directly from amino acid sequences.
AlphaFold2's performance at the 14th Critical Assessment of Structure Prediction (CASP14) contest demonstrated protein structures at an accuracy rivaling experimental methods. Its database now covers over 214 million predicted structures, up from roughly 300,000 when it launched, and has been used by more than 3 million researchers across more than 190 countries.[1] That widespread availability fed directly into drug design. The preclinical phase—from target identification to drug candidate nomination—traditionally takes three to six years. However, AI-native firms like Insilico Medicine completed theirs in under 18 months and reached Phase I clinical trials within 30 months of program initiation.[2] Even though it's not guaranteed that AI-assisted drugs will receive FDA approval, the reduction in both time to trial and costs means pharmaceutical companies have a greater chance of success than ever before.
A similar case shows up in genomics. Missense variants are mutations that alter a single amino acid in a protein and may affect its function. Before the deep learning model AlphaMissense was developed, only about 2 percent of the roughly 4 million such variants in the human genome had been clinically classified.[3] Interpreting the rest required slow, case-by-case laboratory experiments, many of which were left unresolved. AlphaMissense predicts a pathogenicity score for each variant in seconds, though clinical classification still requires confirmation. Digital pathology follows the same pattern. Large deep learning models were trained on over a billion small, standardized image patches extracted from digitized tissue slides. Now, those models can detect rare cancers directly from a scanned biopsy sample.[4] This diagnostic ability previously required hours of manual review by specialist pathologists, an expertise that many hospitals lack.
In environmental science, AI-based weather models have improved the speed of discovery and the application of results. Traditional weather forecasting divides the atmosphere into a 3D grid of boxes, then uses supercomputers to solve physics equations forward in time, step by step. Where the grid is too coarse to compute directly, simplified representations of small-scale processes like cloud formation are used instead. Models such as GraphCast, however, learn from decades of historical atmospheric data to predict how weather patterns evolve, picking up the real statistical behavior of those small-scale processes instead of relying on simplified formulas.[5]
Because GraphCast skips solving the full physics equations at runtime, it produces a forecast in under a minute on a single chip, compared to hours on a supercomputer, with accuracy that matches or beats traditional systems on most measures. This lets researchers run far more forecasts and scenarios in the same amount of time and at the same cost, accelerating the pace of follow-on research. These models inherit known precipitation biases from their training data, a limitation stemming from their reliance on historical patterns rather than physical laws. Even so, the speed and accuracy gains represent a substantial shift in what is computationally feasible for atmospheric science.
In materials science, testing a candidate crystal’s stability has meant either synthesizing it in a lab or running quantum-mechanics-based density functional theory (DFT) calculations, both of which are slow and expensive enough to make exhaustive searches of the possible materials space impractical. DeepMind’s Graph Networks for Materials Exploration (GNoME) model predicts a candidate’s stability almost instantly, sends only the most promising ones to DFT for verification, and feeds those results back into the model. This active learning loop boosted the accuracy of stability predictions from around 50 percent to 80 percent on the field’s benchmark.[6] This approach accelerates research because low-cost computational screening combined with selective DFT checks allows GNoME to cover far more of the space of possible materials than DFT alone could. The result was 2.2 million candidate crystals identified, of which 381,000 were confirmed stable, delivered as a pre-filtered shortlist rather than requiring researchers to test candidates one by one.
Beyond the data bias and absence of guaranteed results already discussed, a lack of explainability and reproducibility can undermine the reliable use of AI in scientific research. In most cases, these models do not reveal how they arrived at a prediction, which is a more significant limitation in science than in many other uses of machine learning, since scientific outputs must typically support a testable hypothesis. The field of interpretability addresses this by developing methods to examine how input data leads to specific predictions in a measurable way. Reproducing a result also requires access to the same code, parameters, training data, and often comparable computing resources. When these are not shared openly, even non-proprietary results are difficult for others to verify or build on, reducing confidence in the findings and slowing broader adoption.
A common pattern is emerging across medicine, environmental science, and physics. Tasks that once took years, or had no practical solution at all, can now be completed in days, hours, or even minutes, shortening the surrounding research pipeline in the process. Protein structures can be predicted in minutes, variant pathogenicity can be assessed in seconds rather than remaining unresolved indefinitely, and millions of candidate materials can be identified without exhaustive laboratory testing.
What these examples demonstrate is that the acceleration is already measurable, even though the end goal in each field, such as an approved drug or a synthesizable material, has not yet been fully realized. Significant challenges still remain, particularly in interpretability and reproducibility. Addressing these challenges will enable faster research to produce results that are trusted, verified, and ready for wider use cases.
Sources
[1] https://deepmind.google/blog/alphafold-five-years-of-impact/ [2] https://insilico.com/phase1 [3] https://www.nature.com/articles/s41576-023-00668-9 [4] https://www.nature.com/articles/s41591-024-03141-0 [5] https://www.science.org/doi/10.1126/science.adi2336 [6] https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/
Becks Simpson is a Machine Learning Lead at AlleyCorp Nord where developers, product designers and ML specialists work alongside clients to bring their AI product dreams to life. She has worked across the spectrum in deep learning and machine learning from investigating novel deep learning methods and applying research directly for solving real world problems to architecting pipelines and platforms to train and deploy AI models in the wild and advising startups on their AI and data strategies.
