How AI helps doctors and researchers • Межа

While social networks are arguing about whether artificial intelligence will replace the work of artists and copywriters, a real, but “quiet” revolution is taking place in scientific laboratories and in Earth orbit. There, AI has long ceased to be a toy for generating cat videos and has turned into a full-fledged tool capable of changing reality. Neural networks allow us to solve problems that for decades seemed insurmountable to scientists. Here are 20 examples of how algorithms have been saving lives, helping to protect the planet, and moving science forward for several years.

Solving the “protein folding problem”

For more than half a century, scientists have been trying to predict the 3D structure of proteins, because it is their shape that determines their function in the body. Previously, decoding a single protein could take years of laboratory work. Interestingly, the authors of EVE Online even involved their community in studying this issue. DeepMind’s AlphaFold system predicts the three-dimensional shape of proteins based on their amino acid sequence, achieving atomic precision with a median error of 0.96 angstroms (one ten-millionth of a millimeter), which is comparable to experimental methods.

The neural network is based on the Evoformer architectural block, which considers protein folding as a graph inference problem in 3D space, combining data from multiple sequence alignments and attention mechanisms for pairwise interactions. To generate the final structure, a separate module with the “invariant point attention” (IPA) mechanism is used, which iteratively refines the coordinates of atoms. Now biologists have an open database of almost 200 million protein models – almost all the cataloged proteins known to science today.

Discovery of the antibiotic Halicin

Researchers at MIT used a deep neural network to find new antibiotics, training it on a set of 2,335 molecules to identify structural features of antibacterial activity against E. coli. The model was used to conduct a computer simulation of a screening of 6,000 compounds from the Drug Repurposing Hub, which avoided human biases toward familiar chemical scaffolds by learning structure-activity relationships directly from the data.

The algorithm identified SU-3327 (renamed halicin), a former experimental diabetes drug, as the best candidate. Despite successful preclinical results, the drug was ineffective against Pseudomonas aeruginosa and has not yet been tested in humans due to its complex pharmacokinetic profile, highlighting the need to optimize future AI models for predicting ADME (absorption, distribution, metabolism, and excretion) parameters. Fun fact: the antibiotic itself is named after the artificial intelligence Hal from the movie 2001: A Space Odyssey.

Early diagnosis of breast cancer

In radiology, every detail in an image is critical, but the human eye is tired and has inherent defects and nuances of perception that lead to errors. AI models from Google Health learned to analyze mammograms, detecting the smallest patterns that are invisible even to experienced doctors. The system was evaluated on depersonalized datasets: over 25,000 women from the UK and over 3,000 from the US. Compared to the decisions of individual doctors, the algorithm demonstrated an absolute reduction in the number of false positives by 5.7% (US) and 1.2% (UK), and false negatives by 9.4% (US) and 2.7% (UK).

The model also demonstrated generalizability: AI trained exclusively on UK data successfully outperformed human experts on US patient data. In a simulated experiment, using AI as an independent “second reader” showed that the system could provide diagnostics of no worse quality, reducing the doctor’s workload by up to 12% of the current volume. The technology has not yet been implemented in clinical practice and requires prospective clinical trials.

Neuroimplants for speech restoration

For people who have lost the ability to speak due to a stroke, communication often becomes almost impossible. Scientists have developed a neurointerface system Intelligent Throat – a portable device that combines ultra-sensitive textile strain gauges and artificial intelligence algorithms. The sensors record microvibrations of the laryngeal muscles during silent articulation and carotid pulse signals. Instead of traditional phrase recognition, the system breaks the speech signal into tokens of 144 ms and uses a 1D convolutional neural network for continuous decoding in real time.

A separate algorithm analyzes heart rate variability to determine the patient’s emotional state (neutral, relief, frustration). Next, two agents based on a large language model (LLM via the GPT-4o-mini API) come into play: the first (TSA) collects tokens into words and corrects recognition errors, and the second (SEA) expands short phrases into full sentences at the user’s request, taking into account the decoded emotion and context, such as weather or time. During testing on 5 post-stroke patients, it achieved an error rate of only 4.2% at the word level and 2.9% at the sentence level, ensuring smooth communication without delays.

Diabetic retinopathy detection

Diabetic retinopathy is a leading cause of blindness, potentially affecting an estimated 415 million people with diabetes worldwide, but many lack access to regular screening due to a shortage of specialists. Google researchers have developed a deep learning algorithm to automatically detect signs of the disease – lesions, hemorrhages or fluid leaks – in 2D photographs of the retina.

A dataset of 128,000 images was created to train the model in collaboration with doctors from India and the US. During comparative testing, the algorithm demonstrated performance on par with a panel of board-certified ophthalmologists, achieving high sensitivity and specificity. This automated screening system is designed to help quickly identify patients with pathologies and refer them to specialized specialists, while algorithms are being developed in parallel to analyze more complex 3D retinal images.

“Eyes” for the blind

The Be My Eyes app has introduced a digital visual assistant called Be My AI to help blind and visually impaired people. The tool is fully integrated into the app and is powered by OpenAI’s GPT-4 language model. The user simply opens a tab and takes a photo, after which the system generates a detailed description of the image.

The main advantage of the algorithm is interactivity: the user can chat with the AI and ask additional clarifying questions about the captured image. In cases where the algorithm cannot provide an answer, or if the result requires human verification, the system provides a mechanism for quickly switching to a live volunteer. Access to the tool is completely free for the target audience.

Flood forecasting

Floods claim thousands of lives every year, especially in regions with poor warning infrastructure. The Flood Hub AI platform is designed to help predict potential disasters more effectively. It uses a combination of two AI models. The hydrological model analyzes public weather, precipitation, and basin data to predict river levels, while the inundation model uses these predictions along with satellite imagery to simulate water flow and determine the exact extent and height of flooding.

Trained on global meteorological products, the algorithm outperforms the standard GloFAS model and is able to predict water level dynamics 7 days before a flood. An important feature of the system is its ability to use algorithms to transfer information from regions with a large number of real river sensors to areas where such data is lacking (thanks to almost 250,000 “virtual sensors”). The system currently covers river basins in more than 150 countries, providing free warnings to 700 million people through Google Search, Maps, Android and specialized APIs for researchers.

Quick weather forecast

Google DeepMind’s GraphCast model performs medium-range weather forecasting using a graph neural network architecture optimized for processing spatially structured data. The algorithm is trained on ERA5 archive data from the European Centre for Medium-Range Weather Forecasts over the past four decades, allowing it to study the cause-and-effect relationships of atmospheric evolution without using traditional differential equations.

Given only two weather conditions (current and 6 hours ago), the model iteratively generates a 10-day forecast with a spatial resolution of 0.25 degrees (28×28 km) for surface and atmospheric variables at 37 elevation levels. The entire calculation process takes less than a minute on a single Google TPU v4 processor. In comprehensive testing, GraphCast outperformed the industry-leading numerical system HRES on over 90% of 1,380 test variables, and also demonstrated the ability to predict extreme weather events more accurately and earlier, even though it was not specifically trained to detect them.

Forest fire detection

In a changing climate, wildfires spread quickly, so response time is everything. ALERTCalifornia uses a network of high-resolution cameras with infrared night vision and spatial laser ranging (LiDAR) data to monitor natural disasters. Together with CAL FIRE and Digital Path, an artificial intelligence tool was developed that automatically analyzes video streams to find fires.

When the algorithm detects a potential fire, it sends firefighters an alert with a percentage of confidence and the approximate location of the incident. Once the event is verified by trained human operators, fire crews can respond to the threat at an early stage, reducing observer fatigue and minimizing false alarms. This AI system has been deployed in all 21 CAL FIRE 911 dispatch centers. It has proven highly effective in remote areas and at night, sometimes generating alerts faster than calls from witnesses.

Traffic light optimization

Air pollution levels at urban intersections can be 29 times higher than on open roads, with half of these emissions coming from cars accelerating after stopping at red lights. The Green Light project uses artificial intelligence and aggregated traffic trend data from Google Maps to model traffic flows without the need for expensive additional equipment or manual counting.

The algorithm analyzes existing traffic light parameters (cycle duration, waiting time, coordination between intersections) and generates timing recommendations for city engineers, which can be implemented in just 5 minutes using existing city tools. According to preliminary data, such changes can reduce the number of stops by 30% and reduce greenhouse gas emissions by 10%. Launched in 2023, the system is currently operating in 20 cities on four continents, reducing emissions for 47 million car trips every month.

Laser weed control

Modern farmers are trying to move away from herbicides to preserve soil and grow clean food. Carbon Robotics’ LaserWeeder combines computer vision, edge computing, and laser kinematics to identify weeds among crops in milliseconds and destroy them with pinpoint laser shots.

Real-time plant classification uses a proprietary Large Plant Model neural network trained on over 150 million images, allowing it to semantically separate crops from weeds even under changing lighting conditions. The thermal pulse instantly destroys the plant’s cellular structure without physically damaging the soil or neighboring crops, allowing it to kill up to 5,000 weeds per minute with submillimeter precision. Carbon Robotics is an NVIDIA investment.

An enzyme that “eats” plastic

Researchers at the University of Texas at Austin used a machine learning model based on a 3D convolutional neural network to generate mutations in the natural enzyme PETase. The algorithm accurately predicted which mutations would allow the enzyme to rapidly depolymerize polyethylene terephthalate (PET) at low temperatures. The resulting variant, called FAST-PETase, is capable of completely breaking down PET plastic into its original monomers in just 24 hours at temperatures below 50 degrees Celsius.

The technology has been practically proven to work in testing on 51 types of consumer plastic containers and five types of polyester fibers and fabrics. Unlike energy-intensive industrial recycling methods, this biological solution requires significantly less energy, and the developers have made the optimized AI model available to the academic community via the MutCompute web platform.

Assessment of damage after disasters

When an earthquake or hurricane strikes, it’s important for rescue workers to know where the damage is most severe so they can direct help there. The xView2 AI system compares before-and-after satellite imagery to automatically create a map of damaged buildings and roads. The algorithms are powered by the xBD dataset, which contains over 850,000 annotated building polygons in high-resolution satellite imagery, before and after the disaster, for six types of disasters in 15 countries. Because there was no single methodology for classifying damage, the developers created the Joint Damage Scale, combining standards from FEMA, HAZUS, and others to unify assessments across agencies.

The system is able to independently assess a segment of the destroyed area. This eliminates a key bottleneck, replacing slow manual search for damage with fast automated analysis, which is critical for the prompt dispatch of rescue teams. What used to take days of analysis is now done in minutes.

Discovering new materials

Google DeepMind’s GNoME (Graph Networks for Materials Exploration) tool uses a graph neural network architecture to predict the stability of new inorganic crystals. The model generates candidates through two parallel pipelines and is refined through iterative active learning, where predictions are tested against quantum mechanical density functional theory calculations and the results are fed back into a training dataset. As a result, the algorithm discovered 2.2 million new crystals, of which 380,000 are recognized as the most stable, equivalent to almost 800 years of experimental discoveries.

The model has been validated in practice: independent researchers have already synthesized 736 predicted structures, and the Berkeley Robotics Lab has used this data to autonomously create more than 41 new materials. The open database now contains 52,000 layered compounds for potential superconductors and 528 candidates for lithium-ion conductors for next-generation batteries.

Fusion control

DeepMind researchers, together with the Swiss Plasma Center (EPFL), have developed the first deep learning-based reinforcement learning system for autonomous plasma control in a tokamak reactor.

Instead of the traditional approach of using 19 separate algorithmic controllers for each magnetic coil, the new architecture uses a single neural network that adjusts the voltage thousands of times per second directly based on sensor data to keep the unstable plasma from contacting the tokamak walls. The system was first trained in a simulated environment, then applied to the real machine to accurately shape the plasma configuration.

By changing the target parameters, the algorithm was able to create and stabilize a “snowflake” shape for exhaust energy distribution, a configuration for the future ITER reactor, as well as an unprecedented “drop” with two separate plasma fragments inside the chamber at the same time.

To further improve AI controllers, the developers released TORAX, an open-source simulator written in the JAX framework that models the inner core of the plasma and predicts changes in its temperature, density, and electric current. This approach not only prevents damage to the reactor walls and reduces cooling costs by optimizing the shape of the plasma, but also significantly speeds up scientific experiments. In the future, this autonomous control technology will contribute to the design of next-generation tokamaks, bringing closer the creation of a clean fusion energy source.

Navigation of the Perseverance rover

Real-time guidance of the rover is impossible due to signal latency, so Perseverance was recently trained to use an AI-based routing system. The unnamed generative AI model analyzed high-quality images of the Martian surface from the Mars Reconnaissance Orbiter’s HiRISE camera, as well as a digital elevation model of the area where Perseverance is located. After identifying critical terrain features—rock outcrops, dangerous boulder fields, sandy areas, and so on—the AI model created a continuous path for the rover with designated waypoints without human intervention.

Restoration of ancient texts

Historical stone inscriptions often come to us damaged and fragmented. To address this issue, researchers developed Pythia, the first deep neural network model designed to reconstruct missing characters in damaged ancient texts. The algorithm works simultaneously at the level of individual characters and whole words, allowing it to efficiently process incomplete fragments and take into account the long-term context of the inscription, such as the repetition of names in other parts of the text.

To train the model, a PHI-ML pipeline was created, which converted the largest digital corpus of ancient Greek inscriptions dating from the 7th century BC to the 5th century AD into a machine-readable format. During experimental evaluation, Pythia demonstrated an error rate of 30.1% at the symbol level, surpassing expert historians from the University of Oxford, whose rate was 57.3%. At the same time, in 73.5% of cases, the correct answer was among the 20 most likely hypotheses of the algorithm, making it an effective auxiliary tool for epigraphists.

Search for exoplanets

The Kepler space telescope has collected data on hundreds of thousands of stars, recording microscopic changes in their brightness when planets pass in front of the star’s disk. But many of the signals are so faint that traditional algorithms either miss them or confuse them with noise. Researchers at Google and NASA used neural networks to search for exoplanets in archival data from the Kepler space telescope. The model was trained to recognize transit signals—minimal dips in a star’s brightness as a planet passes in front of it—using a set of 15,000 previously tested signals.

After achieving 96% classification accuracy on a test set, the algorithm was directed to search for weak, previously missed signals in 670 star systems that already had several known planets. Google’s neural network analyzed archival data and discovered previously unknown exoplanets, including eight planets in the Kepler-90 system. The AI proved that old astronomical data hides new worlds waiting to be discovered.

The first photo of a black hole

The Event Horizon Telescope project combined telescopes from around the world into a single “virtual telescope” the size of Earth, collecting 5 petabytes of data. Since telescopes don’t cover the entire planet, there were huge gaps in the data. Machine learning algorithms, including CHIRP, helped to “stitch” these fragments and reconstruct the image, filling in the gaps with the most likely options. This allowed humanity to see the shadow of the black hole M87* for the first time, confirming the general theory of relativity.

Saving whales from ship collisions

Humpback whale songs have long been a mystery to scientists. Long-term recordings have yielded thousands of hours of sounds, but they have been largely unanalyzed. Google AI, in partnership with the National Oceanic and Atmospheric Administration, ran these recordings through an AI that can now distinguish whale songs from ship and other noise. This allows scientists to track the animals’ whereabouts and adjust ship routes to avoid harming the whales. Anyone can join in the study of songs, potentially finding new patterns and understanding the “language of whales.”