In recent years, artificial intelligence in science has moved well beyond pure experimentation. Neural networks have become a standing working tool. They help find relevant papers, systematize data, and model experiments.
The result is most effective when you use neural networks for science together, rather than relying on a single tool. For example:
- Perplexity helps find relevant links, while Gemini structures and visualizes information;
- GPT-5.1 provides solid support with analysis.
Let's look in practice at how this neural network "symbiosis" works inside the GPTunneL hub.
The best neural networks for science: strengths of each tool
Scientific R&D almost always follows the same pattern: finding sources, analyzing them, working with the information, and drafting the paper. Artificial intelligence in science and education helps speed up these processes — provided the right neural network tools are chosen for the job.
- GPT-5.1: the central analytical link. It has enough power for long reasoning chains, can compile statistics, and turn conclusions into a clear, logical article;
- Claude 4.5 Sonnet: plays the role of an impartial reviewer. It checks whether conclusions are logical, evaluates methodology, spots contradictory statements, and bluntly points out weak arguments;
- Gemini 3.0 Pro: a powerful neural network capable of processing large datasets, writing or fixing scripts, and examining images or video from experiments;
- Perplexity Sonar Pro: quickly finds information and provides plenty of clickable links;
- LLaMA 4 Scout / Mistral Medium 3: supporting neural networks that can shorten text without losing meaning and structure a draft that can later be fed into other tools.
Case 1. Automated literature review on CRISPR
Task: a review of CRISPR genome-editing efficiency in mammalian cells: search → structure → text → visualization.
Perplexity – literature search
Prompt:
"Find 10–15 recent publications (2020–2025) on the topic: 'Efficiency of CRISPR genome editing in mammalian cells.' For each article, provide:
- DOI;
- a brief summary;
- method;
- key findings;
- limitations.
Present the answer as a table."
Result
Perplexity suggested 12 up-to-date publications. For each source, the neural network provided a DOI/link, a concise description, key findings, and limitations. The data was compiled into a convenient table that can be uploaded into other models.
Among the downsides is the model's careless approach: selecting truly relevant sources, at least in this case, had to be done manually. They're all there, but the list needs to be cleared of informational "noise." You can evaluate how Perplexity handled the prompt here.
Claude – structuring the framework
Prompt:
"Structure the data from the CRISPR articles in the following format:
- method;
- cell type;
- results;
- limitations;
- repetitions;
- contradictions;
- research gaps."
Result
Claude grouped the articles by method, cell model, results, and limitations. The AI also flagged sources with repeated information and research gaps. For example, Claude noted that it is still unclear what the long-term effects of genome editing in cells might be.
It's worth remembering that the model didn't verify the reliability of the sources — it simply "trusted" what Perplexity shared with it. See an example of the model's work in this dialogue.
GPT-5.1 – analytical review
Prompt:
- "Based on the structured data, write a scientific review (800–1,200 words) on the topic: 'Efficiency of CRISPR genome editing in mammalian cells.'
- Style — Nature Reviews.
- Include an analysis of trends, limitations, and future directions."
Result
GPT-5.1 wrote a coherent review covering methods, results, limitations, and prospects. The model sticks closely to the requested structure and doesn't distort the provided facts. Stylistically, the text is close to a short article in a scientific journal.
That said, GPT-5.1 overdid it on length — the text needs to be shortened manually, or you can simply ask the AI to make it more concise. You can read the AI-generated review here.
Gemini – visual comparison of methods
Prompt:
"Build a heat map:
- rows — CRISPR methods;
- columns — efficiency, precision, safety, ease of use, and clinical potential.
Rate each parameter with a qualitative label."
Result
Gemini did exactly what was asked. It rated the methods by precision, efficiency, safety, and other criteria, accompanying each rating with an explanation.
The result is a handy illustration for a presentation slide or an article. There's a nuance, though: all the ratings are more like comments than quantifiable characteristics, so not every study can rely on them. You can evaluate the model's work here.
Case 2. Hypotheses on the microbiome and metformin
Task: turn a request into a testable hypothesis, a study design, criticism, and a simple model.
GPT-5.1 – hypotheses and design
Prompt:
- "Generate 5 testable scientific hypotheses about how microbiome composition affects metformin absorption in adults.
- Provide a mechanism, experimental design, data, metrics, and criteria for statistical significance."
Result
GPT-5.1 proposes several hypotheses, for example a link between Akkermansia bacteria and response to metformin. The model also describes the cohort, measurement points, biomarkers, and key statistical criteria.
The result is a workable draft protocol. However, it's important to know that the design turned out somewhat overloaded and lacks an actual power calculation — which could be a problem in a real clinical study. You can review the hypotheses here.
Claude – critique and strengthening the protocol
Prompt:
"Check the hypotheses for scientific correctness.
Point out:
- strengths;
- weaknesses;
- biological limitations;
- possible design errors."
Result
Claude analyzed each block, pointed out confounders (diet, antibiotics, PPIs, genetics), the compositional nature of microbiome data, and the need for a power calculation and phased design. It proposes a more rigorous clinical study protocol.
The answer turned out to be very detailed, but the information still needs to be structured and prioritized afterward. See the result here.
Mistral – publications supporting the hypothesis
Prompt:
"Find studies (with DOI) from 2020–2025 that confirm or refute the proposed hypotheses about how the microbiome affects metformin absorption. Provide brief quotes."
Result
Mistral:
- found several sources;
- provided DOIs;
- included quotes from the sources.
That said, the sample is fairly limited, so it needs to be supplemented manually or with the help of other neural networks. See the result in this chat.
Gemini – a simple statistical model
Prompt:
"Build a simple statistical model (linear regression) linking microbiome composition to blood metformin concentration. Describe the variables and limitations of the model."
Result
Gemini described a multiple regression with "metformin concentration" as the dependent variable and predictors including dose, microbiome composition, SCFA, eGFR, and diet. The model also notes its limitations.
It's worth noting, though, that the neural network offers this at the concept level — without code and without accounting for the compositional nature of microbiome data. If you'd like to see the AI's work, take a look!
Case 3. Statistical analysis of a climate time series
Task: process a global temperature series, build models, and carefully validate them.
Gemini – cleaning and initial trend
Prompt:
"Here's the Global Temperature Time Series dataset.
- Clean the data.
- Remove gaps.
- Standardize the values.
- Build an initial trend."
Result
Gemini indeed removes gaps, outputs the mean value by year, performs standardization (Z-score), and builds a linear trend with a 10-year moving average.
The result is clean code that can be transferred to a local system. However, Gemini doesn't check for autocorrelation or stationarity, and the model doesn't justify the chosen smoothing window. You can evaluate the AI's work here.
LLaMA – regression, ARIMA, Prophet
Prompt:
"Perform a statistical analysis:
- regression;
- ARIMA;
- Prophet;
- outliers;
- trend significance testing.
Draw scientific conclusions."
Result
LLaMA built a linear regression (significant trend), fitted an ARIMA(1,1,1) model, ran Prophet, and identified outliers using the Z-score.
However, not all the data turned out to be well justified. See an example here.
Claude – methodological review
Prompt:
"Check the correctness of the analysis:
- are the models correct;
- are there errors in the conclusions;
- what alternative methods are needed."
Result
Claude pointed to the need for an ADF test, ACF/PACF analysis, checking for structural breaks, using HAC errors, and Mann-Kendall tests. The AI also suggested using more flexible models: polynomial, GAM.
The neural network also wrote code for further analysis. Claude generated a detailed answer, but extracting the useful information from it takes extra work. Here's an example.
Gemini – visualizations
Prompt:
"Build charts:
- trend by decade;
- ARIMA forecast;
- Prophet forecast;
- outliers.
Result
Gemini built a box plot by decade, ARIMA/Prophet forecast charts with confidence intervals, and an outlier chart based on residuals. The visualizations are easy to read and suitable for an article or report. You can see the AI's work in this chat.
Case 4. Analysis of HeLa microscopy images
Task: show how models can make mistakes and how another model corrects the biological interpretation.
GPT-5.1 – initial image analysis
Prompt:
"Analyze the uploaded image of HeLa cells. Identify the cells and nuclei, measure their size, density, and morphology."
Result
GPT-5.1:
- described the "nuclei," "cytoplasm," and culture density in detail;
- proposed a morphometry plan (ImageJ, thresholds, measurements);
- made a fundamental error — it mistook a SEM/3D render for fluorescence microscopy and misinterpreted the surface relief as nuclei and organelles;
- as a result, produced an incorrect report.
Claude – biological validation
Prompt:
"Check the biological accuracy of GPT's analysis."
Result
Claude correctly identified that the image was a SEM/3D render rather than fluorescence microscopy, and reinterpreted the structure: subconfluent culture, pronounced protrusions, migration, moderate heterogeneity. The neural network noted that apoptosis could neither be confirmed nor ruled out from a single frame. See an example here.
Case 5. Automating a chemistry protocol (aspirin)
Task: check how neural networks for research help automate a typical synthesis — from protocol to yield calculation.
Mistral – generating the protocol
Prompt:
"Draft a synthesis protocol for acetylsalicylic acid from salicylic acid and acetic anhydride for a 5 g product yield. Specify:
- masses;
- solvents;
- temperature;
- time;
- mechanism;
- safety measures."
Result
Mistral produced a complete textbook-style protocol with mass calculations, reaction conditions, precipitation, recrystallization, a description of the mechanism, and basic safety measures. The stoichiometry is set so that acetic anhydride is potentially introduced in a reasonable excess. However, the safety section is too generic. You can review the protocol here.
LLaMA – checking the chemistry and safety
Prompt:
"Check the chemical correctness of the protocol, and point out inaccuracies, risks, and safety errors."
Result
LLaMA confirmed the stoichiometry was correct, showed that the limiting reagent is salicylic acid while the anhydride is in excess, and highlighted the risks of working with acetic anhydride and concentrated H₂SO₄. It also noted that personal protective equipment and a fume hood are absolutely essential. At the same time, the safety section says nothing about waste disposal or handling hot ethanol. The detailed protocol is in this dialogue.
GPT-5.1 – yield calculation
Prompt:
"Calculate the theoretical, practical, and expected yield of the reaction."
Result
GPT-5:
- recalculated the molar amounts based on stoichiometry;
- obtained a theoretical yield of ≈5.02 g;
- gave an expected yield range of 80–90% (4.0–4.5 g), with formulas for recalculating based on the actual product mass.
That said, the calculation doesn't account for actual process losses and only works in the context of "pure" stoichiometry. Here's a clear example.
Case 6. Automatically writing an IMRaD paper
Task: automatically generate a paper and assess its reliability using different neural networks.
GPT-5.1 – generating an IMRaD paper
Prompt:
"Write an IMRaD paper 'Comparison of LoRA, QLoRA, and Full Fine-Tuning on Small Medical Imaging Datasets,' 500 words, Nature Methods style, including limitations."
Result
GPT-5.1 delivers the correct structure:
- introduction;
- methods;
- results;
- discussion;
- a limitations section.
The text looks realistic, but there are no specific metrics, architecture, hyperparameters, or dataset descriptions. You can read the simulated scientific paper here.
Claude – rigorous text review
Prompt:
"Check the article for scientific correctness, logical errors, overly general statements, and claims that need citations."
Result
Claude acted as a full-fledged reviewer. It flagged the absence of:
- architecture;
- loss function;
- hyperparameters;
- ROC curves;
- statistical tests;
- dataset characteristics.
The model also identified contradictions in the conclusions. The AI additionally proposed specific revisions. You can read the analysis by opening this chat.
Perplexity – DOIs for the key claims
Prompt:
"Find publications from 2020–2025 (with DOI) that support the article's key claims."
Result
Perplexity found numerous articles on the topic, provided DOIs, and included concise excerpts. The selection is available at this link
Case 7. Multimodal analysis of physics experiment data (damped oscillator)
Context: a CSV with measurements of t(s), x(mm) for a mass on a spring; the task is to extract oscillation parameters and assess how well the model fits.
Gemini – initial CSV analysis
Prompt:
"Here's a CSV file with measurements of a damped harmonic oscillator: t(s), x(mm).
Analyze the data:
- plot x(t);
- estimate the frequency, period, and damping coefficient γ;
- estimate the amplitude at the start of the experiment;
- determine whether the exponential decay is linear on a logarithmic scale;
- assess the quality of the experiment (noise, outliers).
Show all intermediate calculations."
Result
Gemini built a damped cosine curve, found the peaks and zero crossings, estimated a period of ≈0.75 s, a frequency of ≈1.3 Hz, a damping coefficient γ≈1.55 s⁻¹, and showed that ln(A(t)) is nearly linear at low noise levels.
At the same time, the model ignores the fact that the period changes over time, meaning that describing the system as linear is incorrect. You can check out the AI physicist's work here.
GPT-5.1 – oscillator model and parameters
Prompt:
"Calculate the parameters of the damped harmonic oscillator from the experimental CSV.
- Average angular frequency ω.
- Frequency f.
- Damping coefficient γ (via a linear fit of ln(A(t))).
- Quality factor Q.
- Build an analytical model x(t) and compare it with the experiment.
- Assess the discrepancy between the model and the data (RMSE). Provide formulas, substitutions, final calculations, and a brief conclusion on how well the data fit the theory."
Result
GPT-5.1 performed the requested calculations. However, the model failed to account for the fact that the parameters can't be constant — the period and damping depend on amplitude. You can check the calculations here.
Claude – checking the math and methodology
Prompt:
"Check the correctness of the oscillator parameter calculations: ω, f, γ, Q, and RMSE.
Check:
- correctness of the formulas;
- correctness of the amplitude envelope extraction;
- correctness of the logarithmic fit;
- modeling errors;
- statistical errors.
Suggest corrections to the model."
Result
Claude demonstrated that the period noticeably decreases, γ changes with amplitude, and ω increases as amplitude drops, and it re-estimated the parameters (ω≈8.5 rad/s, γ≈1.7 s⁻¹, Q≈2.4) accounting for all the extrema.
The AI additionally proposed a nonlinear model with amplitude-dependent ω(A) and γ(A), and assessed the statistical errors. To dig deeper into the calculations and conclusions, take a look at the chat.
Conclusion
Using artificial intelligence in science is an effective way to quickly handle a wide range of tasks: finding literature, analyzing it, and drafting articles and visualizations.
By combining neural networks, you get exactly the result you're after — a wealth of relevant sources, logical reasoning, no contradictions, and a properly structured scientific text. And the GPTunneL hub gives you fast access to GPT, Perplexity, Gemini, and other AI assistants suited to different scientific tasks.