Researchers cite ten-company drug-discovery model to argue biomanufacturers should pool data

The comparison behind the result fits the question a company asks before joining. Each member's own model is the control. According to Hart, Lee and colleagues, the pooled model significantly outperformed every one of them ^[1]. Heyndrickx's wording is narrower. He reported benefits in "most of the classification or regression tasks," and wrote that the model generally enhanced predictivity, sometimes substantially ^[2].

The 12% median is a relative measure, as its name says ^[3]. It does not mean twelve percentage points of accuracy. Turning it into hits found, or assays a chemist could skip, would take the baseline scores in the consortium's own papers.

The pooled data are also sparse. Treat each data point as one molecule-assay measurement. More than 21 million molecules crossed with more than 40,000 assays gives at least 840 billion possible pairs ^[12]. The 2.6 billion points fill at most about 0.3% of them ^[13]. That is an average of no more than about 124 measurements per molecule ^[14].

The thing this doesn't tell you is whether the gain carries over from molecules to a production plant. The MELLODDY model used "the chemical structure of potential drug compounds to predict how they will function" ^[7]. The groundwork the authors say manufacturing needs is a different job: common ontologies and schemas, and sensors from multiple vendors that communicate in real or near-real time ^[8]. Hart leads the Big Data Project at the National Institute for Innovation in Manufacturing Biopharmaceuticals ^[15]. He says redesigning legacy data infrastructure is expensive and time-consuming, and that equipment developers give multivendor operability low priority ^[9]. Regulators have not yet decided how such models will be evaluated in data submissions. People trained in both biomanufacturing and data science are scarce ^[10]. A drug-activity benchmark scores predictions on data that are already clean and pooled, and those barriers all come before that step.

Lee accepts that in-house work has its uses. "Many organizations are already pursuing in-house solutions," he told GEN. "The benefits are the ability to tailor the solution to one's specific situation." ^[11] His case for pooling is that firms have problems in common: "Many organizations face similar challenges and opportunities related to development and manufacturing," Lee said ^[16].

I think MELLODDY answers a narrower question well. Ten competitors put decades of proprietary data into one model trained on a secure platform ^[5][6], and the pooled model beat each firm's own ^[1]. To show the same for bioprocess data, someone would have to run that member-against-pool comparison on plant records. Until then, the team's warning that "Companies that do not adopt these capabilities are competitively disadvantaged from realizing the full benefits of their big data in the biopharmaceutical manufacturing marketplace" is an argument from analogy ^[17].