What Stays Useful After the Next Model
A pharma company that builds its own model to predict compound toxicity owns something valuable, but it'll be out of date within a few years.
Academic labs and startups keep releasing open models that do much of the same work for free. Stanford researchers released ADMET-AI in 2024, which predicts toxicity and other drug properties. MIT and Recursion released Boltz-2, which predicts protein structure and binding affinity, as open source in June 2025, seven months after Boltz-1. AstraZeneca published its molecule-generation software, REINVENT, for anyone to use. With each release, the in-house team loses some of the lead it had at launch.
But the same company likely also owns several things it probably hasn't written down anywhere: the assay results it generated as training data, including thousands of compounds that failed; its agreements with the contract research organizations (CROs) that ran those assays; the pipeline its team built to clean and label the results; and the evaluation suite the team used to check whether the model worked. The company will keep using all of these after it retires the model, and it would struggle to replace any of them. Most companies don't count them until they need them: when a team tries to replace the model, a regulator asks for validation records, or a partner asks what a data deal would include.
Pharma labs should treat the model as the perishable part of an AI project and manage everything else as the asset. Scientists can run a model a million times without wearing it out; it goes stale because other labs keep releasing better ones.
The model ages fastest
Researchers worked on protein structure prediction for about fifty years. DeepMind released the code for AlphaFold 2 in 2021, and from then on any academic lab could download a model that solved much of the problem. Researchers have repeated this across the field: they publish new architectures, other labs adopt the training methods, and chipmakers and cloud providers keep cutting compute costs. A company that trains a model in-house on public data can't keep any advantage past the next open release.
A company that trains on proprietary data does keep an advantage, but it comes from the data. Train a newer architecture on the same data, and you get a better model, and you have no reason to keep the old one. You carry the data forward and retire the model.
Companies get the most lasting value from negative results. They publish their successes and rarely their failures, so no competitor can rebuild from the literature a record of thousands of compounds that didn't bind, didn't absorb, or proved toxic in a given assay. Public datasets mostly hold compounds that worked, and anyone who trains a model only on them gets a skewed view of chemistry.
Three things that last
To find out what counts as an asset, ask one question:
If your model disappeared tomorrow, what would you need to rebuild it?
Anything on that list you couldn't buy back is an asset. You can sort most of the answers into three groups.
Data. Proprietary assay results, negative results, longitudinal patient data from trials, and the tools that make the data usable: the pipelines that clean and label it, and the software your scientists use to inspect it. A company that stops adding to a dataset falls behind the field. A company that adds to it every quarter keeps up. Teams that keep the tools can add new data in the same form as the old. Teams that lose them start the cleanup over each time.
Rights. Agreements with hospitals, CROs, and trial sites, and the scope of patient consent. You need these to keep adding to the data and to reuse what you already hold. If patients consented only to the original study, you can't use their data to train the next model, however good that model is. Companies often get more from a supply agreement that renews every year than from any single dataset it delivered.
Evidence. Benchmark sets held out from training, validation protocols, the wet-lab loop where scientists check predictions against real experiments, and the tools that show where a model fails. Your team will measure every future model against these. With them, a team can swap in a new model in weeks and know whether it's better. Without them, the team has to take the vendor's word or rerun the whole validation program.
Regulators ask for the same evidence. In draft guidance on AI in drug development, issued in January 2025, the FDA asks sponsors to establish a model's credibility for a specific context of use and to keep monitoring it over its life. Sponsors do that work with benchmark sets and validation records, and they can carry most of it from one model to the next.
A worked example
Take a screening program. A CRO runs assays on a company's compounds across several screening rounds. An internal team writes a pipeline that normalizes the readouts across plates and batches, flags bad runs, and labels each compound. The team then trains a model on the labeled assay results.
Which part is the asset: the assay results, the CRO contract, or the pipeline?
Ask the rebuild question. The assay results are fixed. The team can use them as training data for this model and maybe the next one, but the company stops adding to them the day the contract ends. Under the contract, the company can order more assay results, in the same format, from the same lab. With the pipeline, the team can use those new results without another six months of cleanup, and the pipeline includes rules for what counts as a bad run, rules that took months to agree on. If the company lost the assay results, it could regenerate many of them with the contract and the pipeline. If it lost the contract and the pipeline, it would hold a fixed set of results and have no easy way to add to them.
So the contract and the pipeline are the assets, and the assay results are what the company has produced with them so far.
Calling the contract an asset may sound odd. Most people think of an asset as something no one else has, like a patent or a unique dataset, and CROs sell a commodity: the company could hire another one to run the same assays. But when you ask the rebuild question, you count what you couldn't buy back, not only what no one else has. A new CRO would run the assays with its own protocols, plate layouts, and instruments. The team would have to retune the pipeline's rules and check that the new results line up with the old ones before mixing them. The company could replace the CRO or rewrite the pipeline, but its team took months to get the two working together, and it would spend those months again.
What Netflix gave away
In 2006, Netflix released about 100 million customer movie ratings and offered $1 million to anyone who could improve its recommendation algorithm by 10 percent. A team won in 2009¹.
Netflix could hand over 100 million ratings, rare proprietary data, because competitors couldn't do much with them. Competitors couldn't copy Netflix's subscriber base, or the system Netflix used to turn recommendations into people who kept paying. Netflix let outsiders build the algorithm, the part it cared least about owning.
Researchers then showed they could identify some Netflix users by matching anonymized ratings against public IMDb reviews. A subscriber sued, and Netflix canceled a planned second competition. Netflix believed it could release the data safely, but it still owed obligations to the people whose data it came from. Drug companies owe heavier ones to patients, which is why rights belong on the asset list next to the data.
From the Netflix case, you can take a second question to ask alongside the first: what could a competitor do with each item if you published it tomorrow? Most executives' first answer is to publish nothing, so ask it item by item. For most companies, a competitor could do little with the model architecture or with training data drawn from public sources. A competitor who got your negative results, your consent terms, or your evaluation suite could save years; those items are your assets.
Teams build these assets by accident
Nobody plans a labeling pipeline as a deliverable, and nobody writes the evaluation suite into the project charter. Teams build these things because they need them to finish the work. Because nobody planned them, nobody owns them. One engineer keeps them in a personal repository, or a postdoc keeps them in notebooks with no documentation, and when that person leaves, no one else knows how to run them.
So name them at the start. When a team scopes an AI project, it should list the parts that will outlive the model: the datasets, the contracts, the pipelines, the evaluation suite, and the validation records. Treat those as products, with owners and a maintenance budget, and treat the model as one thing they produce. Budget for upkeep, because a team that hasn't updated its evaluation suite in three years is testing models against conditions that no longer hold. Write down the decisions built into each pipeline so the next team doesn't have to argue them out again.
Companies collect on this work when they need to replace the model, lose a vendor, or adopt a better architecture. A company that has kept up its data, rights, and evidence can retrain and validate a new model in a fraction of the time. A company that focused mainly on the model starts over.
What to ask your vendors
Many pharma AI projects run through vendors, and under the standard contract the buyer gets a trained model; but the vendor usually keeps the evaluation tools, validation code, and records of how it transformed the data.
Most buyers don't ask for them.
Before signing, get answers on four things:
Whether you receive the evaluation suite and validation code along with the model
Who owns the annotations and derived data the vendor creates from your data
Whether you get the data lineage: what the vendor filtered, transformed or excluded, and why
What happens to all of the above when the contract ends
If the vendor keeps the evaluation suite, you can't compare its next model, or a competitor's, against the one you have without rebuilding that suite yourself. The model you're paying for will fall behind the next open release. You'll still use the evaluation suite years from now, if you own it.
Separate costs from investments
Once a company knows what each project produced, it can see what its teams have been building all along: a factory for models. The team will retire each model the factory produces, but it keeps the factory: the data coming in, the rights that keep it coming, the pipelines that clean it, and the evidence that shows whether a new model beats the old one. Few of those parts are rare on their own. The team builds value by getting them to work together, so it can take a new architecture and turn it into a validated model in weeks.
With that in view, the company can split its AI spending into two buckets.
The model is a cost. Buy or build the cheapest option that does the job, swap it when someone releases a better one, and drive costs down over time.
The factory (data, rights, and evidence, and the work that ties them together) is an investment. Spend on it deliberately, maintain it, and judge it by how much more it lets you do next year.
Companies already make this split with property. A company that rents an office looks for flexibility and a fair price. A company that buys one looks at the location, the building's condition, and what it will be worth in ten years. Most pharma companies budget an AI project as one cost, so when the project ends, they cut the pipelines and the evaluation suite along with the model.
Once you split the two, you can ask better questions: which projects left you with lasting assets, which vendor terms gave them away, and where the next budget should go. Biotechs that plan to license or sell programs get one more benefit.
A partner won't pay a premium for a model that will age out before the drug reaches the clinic; but they will pay for the factory that built it.
¹ Netflix never put the full winning solution into production. Its engineers later wrote that they couldn't justify the engineering work for the extra accuracy, and by then Netflix had moved to streaming and needed different things from its recommendations.