AI for Drug Discovery: How Google Antigravity Searches for Molecules and What the Experiment Showed

In the experiment described, Google Antigravity acts as an AI system for searching for and initially evaluating drug compounds. According to the concept outlined in the review materials, it combines language models, scientific publications, patents, and specialized cheminformatics software, progressing from data retrieval to formulating hypotheses about new molecules.

This result cannot be considered a finished drug.

What Antigravity Is and Which Part of the Work AI Performs

Antigravity does not replace a chemist, pharmacologist, or laboratory. In the review author's account, the system works as a research assistant: it finds publications and patents, extracts information about compounds, compares chemical structures, organizes filtering, and generates a report. The language model also plans the sequence of actions, selects suitable tools, and explains the results obtained.

In the demonstration, Google stated that tasks that previously took several days and required substantial preliminary preparation in computational chemistry were completed in approximately one hour. This is an estimate for a specific demonstration scenario, not a universal speed standard for all drug discovery projects: the duration depends on the quality of the initial data, the number of compounds, the search settings, and the need for manual verification.

For a research group, such a system is most useful at the early stage. It reduces the time between defining a question and preparing a list of compounds worth examining in greater detail. Decisions about synthesis, biological testing, and further development remain with specialists.

How the Language Model Connects to RDKit and Scientific Databases

The main feature of this approach is the separation of roles. The language model handles query understanding, information retrieval, action-chain planning, and preparation of an explanation. Specialized chemical tools perform precise operations on molecules. Among the tools mentioned in the materials reviewed by the author is RDKit, a software toolkit for working with chemical structures and performing computations in cheminformatics.

The review author describes RDKit as a tool that represents molecules in machine-readable form, compares their fragments, calculates descriptors, searches for structural similarity, and performs filtering based on specified features. This characterization refers to RDKit's role in the architecture under consideration; the materials provided do not confirm that Google officially listed this particular toolkit or used it for each of the operations named.

The workflow looks like this:

• the researcher formulates a task, such as searching for compounds similar to a known inhibitor;
• Antigravity examines publications, patents, and available information about molecules;
• the system sends the chemical structures to specialized software;
• the algorithms calculate features and discard unsuitable options;
• the language model compares the results, generates a report, and suggests the next stage of analysis.

Such an architecture reduces the risk of a simple textual error but does not eliminate it completely. An error in the source database, an incorrect interpretation of a publication, or an improperly specified filtering criterion will affect the entire subsequent report.

The Experiment with Crizotinib and Lorlatinib

For the demonstration, Google selected crizotinib and lorlatinib. In the experiment materials, these drugs were used in the context of treating certain forms of lung cancer and were presented as first- and third-generation agents, respectively. The reference to medical indications here conveys the context of the demonstration rather than an independent conclusion by the review author. The researchers used information from articles and patents to study the molecular features of the compounds and find promising options for further analysis.

Antigravity found thousands of compounds, then filtered them and prepared an analytical report. In one example, the system retained 568 molecules and presented the results as interactive charts. All 568 compounds do not thereby acquire proven therapeutic activity: the number indicates the size of the selected set for further study.

Why did ring structures attract attention? According to the description of the demonstration, the system noticed a relationship between more rigid molecular scaffolds and high activity in the available data. Such structures often contain fewer rotatable bonds, and a change in molecular shape can affect its interaction with a biological target. Google specifically noted that observations of this kind require further validation.

Antigravity also compared pairs of similar molecules. This approach helps identify a structural change accompanied by an increase or decrease in activity within the data set. In the demonstration, the AI proposed this analysis independently before the researchers formulated it as a separate task. The observation applies to the described scenario and does not prove that the system always independently finds meaningful patterns.

How the System Predicted the Efficacy of New Compounds

After preparing the set, Antigravity created several machine-learning models to predict the efficacy of new compounds. Google reported that the models were tested on data not used during training. This test shows whether the predictive quality is maintained on previously unseen examples.

Testing on held-out data is useful, but it is insufficient to conclude that a molecule has therapeutic effects. A model may reproduce the patterns of a particular database well and perform worse on compounds that differ substantially from the training set. The result is also affected by the completeness of the experiments, the consistency of the measurements, and how accurately negative and positive examples were collected.

For a pharmaceutical team, the prediction serves as a prioritization tool. From a large number of candidates, researchers select compounds that are more rational to synthesize and test first. AI does not state that a molecule will become a drug. It estimates the probability of matching the selected features given the available data.

Synthesis Schemes: A Useful Hypothesis, Not a Laboratory Protocol

In the experiment described, the AI proposed several new chemical compounds, predicted their activity, and prepared possible laboratory synthesis schemes. The system explained the presumed reactions and electron movement, linking a textual description of the transformation to the chemical logic of the selected route.

Such a result helps a specialist evaluate an idea more quickly and find an approach for discussing the synthesis. However, the proposed scheme cannot be transferred to laboratory work without verification by a chemist and experimental optimization of the conditions. An actual reaction depends on reagent purity, solvent, temperature, pressure, equipment, side processes, and the characteristics of the specific substrate.

The proposed molecules can also be studied further using AlphaFold and molecular dynamics simulations to assess possible interactions with proteins. These methods provide computational hypotheses about structure and interaction. They do not replace biochemical analysis, cell-based tests, toxicity studies, or subsequent stages of preclinical and clinical development.

Why the Result Cannot Be Considered a Finished Drug

Google explicitly emphasizes the hypothetical nature of the results. Prediction accuracy depends on data quality, and every new molecule must be synthesized and tested experimentally. Only after such testing can it be determined whether the compound exists in the required form, retains its presumed activity, and has acceptable safety properties.

And what happens if the model correctly predicts activity in one computational database? That still will not prove efficacy in humans. Between a computational prediction and a drug lie confirmation of the mechanism of action, pharmacokinetic assessment, toxicological studies, quality control, and clinical trials conducted according to established procedures.

The weak point of this approach may lie not in the language model itself but in the data and the way the task is formulated. If publications describe predominantly successful compounds, the system will receive a biased picture. If experimental results were obtained using different methods, the comparison will be less reliable. The researcher must verify the origin of the data, the measurement conditions, and the model's applicability limits.

Access to Chem Informatics Skills

The materials on which this review is based mention the use of specialized cheminformatics tools, including RDKit. The review author has no confirmation that a toolkit called Chem Informatics Skills is available in Google Antigravity through a specific Science plugin.

For a researcher, the value of such a tool is determined not by the name of the function but by the reproducibility of the workflow. It is important to see which sources the system used, which compounds it excluded, which parameters it passed to the chemical tool, and why it selected a particular prediction model. Without this information, the report is more difficult to verify and reproduce.

A Free Program for Scientists

The available materials do not confirm the terms of the program described as providing free access to the highest level of Antigravity use for 10 000 scientists for six months. Therefore, the number of participants, the access period, and the claim that applications are openly accepted cannot be considered established facts. If such a program is indeed operating, its terms depend on the official rules, selection criteria, region, and set of available features.

The idea of providing research access can be assessed separately. Groups that want to test AI on their own sets of publications, patents, and chemical structures should define success criteria in advance: reduced search time, filtering quality, the proportion of correctly identified structures, reproducibility of calculations, and the number of candidates that reach laboratory testing.

What the Experiment Showed

According to the review author's assessment, Google's demonstration shows a combination in which the language model manages the research process while specialized chemical software performs formal computations. In the scenario described, Antigravity collected information about crizotinib and lorlatinib, selected 568 molecules in one example, built charts, trained predictive models, and proposed variants of new compounds with possible synthesis routes.

The main conclusion is more modest than the promise of automatically creating a drug. AI can accelerate the search, organization, and initial evaluation of scientific hypotheses. The reliability of each hypothesis still has to be verified experimentally, and the decision to advance a compound is made by the research team.

Frequently Asked Questions

Does Antigravity create drugs without human involvement?

No. The system proposes compounds and computational hypotheses, but synthesis, biological validation, safety assessment, and clinical development require the involvement of specialists.

What does the number 568 mean?

In the demonstration example, Antigravity selected 568 molecules from a broader set for further analysis. This is not the number of confirmed drug candidates.

Why is RDKit needed?

In the concept outlined in the review, RDKit performs specialized operations on chemical structures: calculating descriptors, comparing molecules, and filtering. The language model uses the results of these operations when planning and explaining the research. The review materials do not confirm that Google has officially stated that RDKit is used at each of these stages.

Can the proposed synthesis scheme be trusted?

The scheme serves as a starting hypothesis. Chemists test it, then determine actual reaction conditions and confirm the composition of the resulting substance using analytical methods.