AI with little data sounds like a joke to a lot of people. You read about models trained on huge piles of examples, then you look at your company spreadsheet and think: no way. It is just a few rows gathered over the years, often full of gaps. This feeling shows up in businesses of every size in Brazil. And it almost always becomes an excuse to push any artificial intelligence project to later.
A story from MIT News, published this week, shows that the math can work out differently. MIT researchers created CheMeleon, an open model that predicts how molecules will behave even when there is little information about the task at hand. The secret is not gathering more data. It is not starting from scratch.
Chemistry may feel far from your daily routine. But the logic behind the project speaks to any company that wants to use pretrained models without a mountain of records. Below, you will see what MIT did, why the pharmaceutical industry paid so much attention, and what you can take from this idea into the real world here.
What CheMeleon is, the MIT AI for molecules
Think about the work of people developing a new drug. Before getting anywhere useful, the team needs to find out if a molecule dissolves well, if it is toxic, if it is potent enough to work. Each of these answers usually takes slow, expensive lab tests. Since there is still no way to perfectly predict how a compound will behave, much of the process runs on trial and error.
CheMeleon MIT steps in right at this point. It is a machine learning model that tries to estimate these traits before any physical test. That way, the team can pick out the most promising molecules and focus time and money on them. The name is a nod to the chameleon, as the caption in the story explains: a single model that adapts, like an animal changing color, to predict different properties such as potency, toxicity and solubility.
According to MIT News, the work came from the group led by Professor William Green, in the Department of Chemical Engineering. The study was written by graduate student Akshat Shirish Zalte, recent PhD Jackson Burns, research scientist Charlles Abreu, Green himself and three researchers from BASF, a chemical company. The research appeared in the September 14 issue of the Journal of Chemical Information and Modeling, although an earlier version of the paper had been available online since last year.
There is one more detail that carries real weight. The model is open and was built on ChemProp, another open source tool created and maintained by the same group. Anyone can download CheMeleon through it and adjust it for their own use. According to the outlet, several drug companies already rely on the model on a regular basis.
How a pretrained model works (no jargon)
The idea happens in two steps.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
In the first, the model learns the basics of chemistry in a general way, without aiming at any specific problem. In the second, it gets fine-tuned for the task that matters, like predicting solubility or toxicity. Since it already knows a lot when it arrives, this tuning needs far fewer examples than training something from scratch would.
According to MIT News, pretraining used about 1 million molecules from PubChem, a large public chemical database run by the US National Institutes of Health. For each molecule, the model learned to predict around 1,600 descriptors, which are simple traits to calculate that chemists have been developing for decades. Molecular weight and the number of oxygen atoms are two examples mentioned.
Here is a curious point: the team did not have the model read scientific papers. Zalte told the outlet that other successful models lean on data published in chemistry journals, but the group took another path and worked only with these calculated descriptors. One million molecules is a tiny slice of all the possible combinations of atoms. Even so, it was enough for the system to spot patterns and apply them to molecules it had never seen.
A comparison helps picture it. Imagine two new hires joining the same company. One is an intern who has never worked in the field. The other has plenty of years in the industry. Both need to learn the internal system, but the experienced one catches on fast, because they get the context. The pretrained model is that experienced pro. You do not need to teach everything, only what is unique to your case.
This approach stands in contrast to so-called classic models, built by experts. In those, someone who knows the field manually points out which traits likely affect the result, like the size of the molecule or how branched it is. It works, but it depends heavily on human knowledge built in by hand.
The numbers that caught the market by surprise
CheMeleon became famous even before the paper went through review by other scientists. According to MIT News, this year OpenADMET held a competition to predict properties of molecules that had not even been synthesized yet. When the results came out, 9 of the top 20 finishers had used the MIT model at some point in their process.
From there, drug companies started reaching out to the Green group, which had only just submitted the paper. The story says at least one of them sent a message to its entire internal team: whenever anyone modeled a property, the starting point should be CheMeleon. It became company policy. For an open source academic project, that is not the kind of adoption you see often.
Formal tests backed up that impression. According to the outlet, the researchers compared the model with other foundation models and also with classic methods based on expert knowledge. In an evaluation across 58 test cases, Zalte says CheMeleon came out ahead overall.
Notice the choice the group made. They did not settle for comparing the model only with modern competitors. They also put it up against the traditional way of doing things. That care is worth its weight in gold for anyone evaluating a new tool, and we will come back to it later.
The lesson for anyone with little data
The core message is simple: do not start from scratch. Zalte told the outlet that most current deep learning models need more data than traditional methods. CheMeleon gets around this limit because it already carries a base of general knowledge. When it is time to solve a specific problem, a few examples may be enough for solid performance.
The team itself says it was inspired by large language models, which learn from huge amounts of text before any specialization. This principle of learning the general first and specializing later shows up across many areas of AI. It is no guarantee of success in every scenario, especially since the MIT story only covers molecules. But the logic resembles other projects trying to get useful predictions out of limited data, as in the case of artificial intelligence that predicts extreme events without historical data.
What changes, in practice, is the question the company asks. Instead of how much data do I need to collect before I begin, the question becomes which model already exists and understands a bit of my problem. It looks like a small swap. But it can save a good chunk of project time and keep the idea from dying during data collection.
How to train an AI model with little data in practice
The first step is to research before you build. Many fields already have some open or ready-to-tune model, whether for text, images, demand forecasting or classification. Setting aside time to map these options usually saves a lot of work down the road. That is what happened with those who adopted CheMeleon: they downloaded the model and skipped the heaviest part.
Picture, as an illustrative example, an online store that wants to sort customer complaints automatically. Training a language model from scratch would be out of reach. Starting from a model that already understands Portuguese and tuning it with the messages the store has saved is far more realistic.
The second step is to take care of the data you already have. When the dataset is small, every bad row weighs more. It pays to standardize names, remove duplicates, check dates and fill gaps where you can. A lean, clean dataset tends to deliver more than a big, messy one. For those who want to go further, tracking how the model behaves once it is in use also helps, a topic covered in the piece on observability in AI models.
The third step is to test for real. Compare the pretrained model with a simple solution, like a historical average or a rule your team already uses every day. MIT did exactly that by putting CheMeleon up against classic methods. If the new model does not clearly beat the basics, it may not be worth the cost. If it does, you will have a solid argument to bring to leadership.
That simple? Not always. But it is a roadmap that lowers the risk of spending months on a project that does not deliver.
Artificial intelligence for Brazilian businesses: where this fits
Many sectors in Brazil live with small datasets. Picture a farming cooperative with only a few years of yield records per field plot. Or a regional factory with a short history of machine failures. Or a local retail chain that only recently started organizing its sales data. In situations like these, a pretrained and fine-tuned model can be a workable start, as long as there is a foundation model related to the problem.
In healthcare, the picture is similar. Smaller clinics and labs rarely have the volume to train something from scratch, but they can benefit from models that already bring general knowledge and only need tuning, always with the privacy and validation care the field demands. More accessible tools also make this shift easier, as shown by the rise of artificial intelligence with low-code in companies.
For people in marketing, the conversation is much the same. A small agency, for example, can use ready-made models to analyze the tone of social media comments for a client, tuning them with a few samples labeled by its own team. You do not need a team of data scientists to take the first step.
The real gain is cutting down on trial and error. CheMeleon does not replace the lab, it helps pick what is worth testing. In your company, the logic can be the same: use AI to rule out bad paths sooner and focus energy on what has the best shot at working. Having little data is still a limitation, but it no longer has to be the reason to shelve the idea.
Pick a small problem in your business, check if there is already an open model that understands something similar, and test it against what your team does today. And if you want to keep exploring how AI is being applied beyond the obvious, it is also worth reading what military artificial intelligence teaches businesses.
