Drug development is often described as searching for a single needle, but the actual work is more like first finding all possible needles in a warehouse, then measuring, categorizing, and comparing each one, and also determining which ones are worth developing further. On November 3rd, Google DeepMind demonstrated its chemical informatics capabilities for Google Antigravity: The agent searched through thousands of compounds around real lung cancer drug cases, completed analysis of matching molecular pairs, and proposed rigid macrocyclic molecules along with synthesis routes. The demonstration compressed a process that could originally take several days into about an hour, but “proposing hypotheses faster” does not equate to “getting new drugs faster.”
The agent is responsible for scheduling, and the computing tools are responsible for the calculations.
The key to this workflow is not to have the language model guess chemical conclusions from memory, but to have it call upon deterministic tools. The system connects to RDKit and integrates resources such as ChEMBL and PubChem. After researchers set a goal, the agent first searches for public compound data, filters out irrelevant entries such as buffers and solvents, records the sources, and then hands over the organized collection to subsequent analysis. The language model takes on the roles of disassembling tasks, selecting tools, and organizing reports, while molecular descriptors and structure comparisons are calculated by specialized software.
This division of labor is much more reliable than “letting chatbots directly design drugs.” Chemical structures, bonds, and molecular properties need to be represented precisely; any minor error at the character level could turn what seems like a reasonable answer into a non-existent molecule. Deterministic tools confine calculations within verifiable rules, while agents reduce the time researchers spend moving information between databases, scripts, visualizations, and documents.
Why compare when there is only one differing factor?
The demonstration uses matched molecular pair analysis to compare compounds that differ in structure by only one discrete change, in order to observe the impact of this change on activity and properties. The team paid particular attention to the results brought about by 'cyclization': large ring formation may reduce rotatable bonds and is associated with higher potency in case data. The agent also established various potency prediction models, which were validated using reserved data; when time information is available, it tends to use a training and testing method divided by era, otherwise, it can use skeleton segmentation to estimate generalization ability.
The value of this step lies in breaking down “relevance” in more detail. It is difficult to determine where the differences in efficacy come from when directly comparing two completely different sets of molecules; by matching molecule pairs and only changing one factor at a time, it is closer to the way chemists typically make judgments about structure-activity relationships. However, the common changes observed in the data still cannot automatically prove a causal relationship, as experimental conditions, sample bias, and measurement noise can all affect the conclusions.
One hour later, the result is still just a candidate answer.
DeepMind clearly describes the final design as assumptions that need to be synthesized and verified. The agent proposes rigid large rings and possible synthesis routes, but this does not mean that these molecules can already be stably prepared, let alone that they are effective in cells, animals, or the human body. After drug discovery, there is still a long chain of steps including purity, selectivity, toxicity, metabolism, dosage, and clinical trials. Converting the demonstration time directly into "how many times the R&D cycle is shortened" ignores the most expensive and uncertain experimental stages.
Data boundaries are equally important. The efficacy values in public databases may come from different laboratories, using various measurement methods under different conditions; sparse data can make a model perform well near its familiar structure, but it may fail on new datasets. Intelligent systems can run analyses more quickly, but they may also amplify dirty data more swiftly. Therefore, source records, filtering rules, model segmentation, and failed samples must be preserved along with the desirable results.
What is truly improved is the research throughput.
If such systems enter the laboratory, the most immediate change may not be to replace chemists, but rather to enable each researcher to try more directions at a lower cost. Explorations that in the past required waiting in line for data scientists to write scripts can now be initiated by intelligent agents, which generate analytical drafts that can be reviewed; chemists can then spend their time questioning hypotheses, selecting experiments, and interpreting unexpected results. DeepMind also plans to provide Google Antigravity at the highest level of usage for six months to 10,000 eligible scientists in order to expand the trial period.
Higher throughput also brings new governance issues: who approves database access, whether the automatically generated routes involve controlled substances, how to record the candidates produced by the models, and who is responsible for discovering errors. The most seasoned “AI chemists” do not present answers as definitive facts; instead, they leave each step of tool invocation, data sources, and uncertainties for review by others.
What a one-hour demonstration truly illustrates is that scientific research software is evolving from a passive tool to an active collaborator. It can link retrieval, data cleaning, modeling, visualization, and reporting into a cohesive process, yet experiments are still necessary to validate the conclusions. The significance of speed lies not in bypassing scientific methods, but in enabling scientists to more quickly approach the questions that are most worth verifying.
