{
  "id": "biomaterials-discovery-needs-the-failures",
  "type": "note",
  "title": "Biomaterials discovery needs the experiments that failed",
  "date": "2025-11-12",
  "description": "Machine learning on biomaterials data keeps running into the same problem: the published record is a survivorship-biased sample of what worked. High-throughput experimentation is one of the few ways to fix that at the source.",
  "tags": [
    "Machine learning",
    "Biomaterials",
    "Research practice"
  ],
  "readingTimeMinutes": "6",
  "aboutPublication": "mapping-biomaterial-complexity-machine-learning",
  "url": "https://emanahmed.org/notes/biomaterials-discovery-needs-the-failures.html",
  "apiUrl": "https://emanahmed.org/api/notes/biomaterials-discovery-needs-the-failures.json",
  "author": {
    "name": "Eman Ahmed",
    "orcid": "https://orcid.org/0009-0008-7822-9530",
    "affiliation": "Rutgers, The State University of New Jersey",
    "url": "https://emanahmed.org/"
  },
  "body": "Biomaterials are hard to design because their performance usually comes from several subtle properties interacting at\nonce. Surface chemistry, molecular weight distribution, charge density, hydrophobic balance and architecture all\ncontribute, and rarely independently. The structure–function relationship is real, but it is not something you can\nusually write down.\n\nThe traditional response is to hold everything constant and vary one thing. It is rigorous, and in a space with this\nmuch interaction between variables it is close to hopeless. You sample a line through a space that has structure\nin every direction.\n\nThe obvious fix has a less obvious problem\n\nMachine learning is the natural tool for a high-dimensional structure–function map, and the barrier to using it\nhas dropped sharply: an experimentalist can now train a competent model without a background in statistics. That is a\ngenuine shift, and it is the shift our review\nin Tissue Engineering Part A was written around.\n\nBut models are only as good as what they learn from, and the biomaterials literature is a filtered sample. Papers\nreport the formulations that worked. The ones that aggregated, or failed to release, or provoked an immune response, are\nmostly absent, not through dishonesty, but because a null result is hard to publish and the material was\nabandoned.\n\nA model trained on that record learns which successful materials resemble other successful materials. It has very\nlittle to say about where the boundary of the useful region lies, because it has never been shown the other side of it.\n\nWhy high-throughput data is different in kind\n\nThis is the part I think is underappreciated. When you run a 96-well plate, you keep every well. The formulations that\nprecipitated are recorded with the same rigour as the ones that performed, because they were measured by the same\ninstrument in the same run. You are not making a publication decision about each data point.\n\nThe resulting dataset is balanced in a way that a literature-derived one structurally cannot be. That is a large part\nof why high-throughput experimentation and machine learning belong together, not because the throughput is\nimpressive, but because of what it does to the shape of the data.\n\nWhere data mining still helps\n\nNone of which means the published record is useless. Text and data mining across existing literature can map regions\nof a design space cheaply enough to tell you where not to spend benchtime, and for well-studied material classes the\naggregate signal is strong. Our review covers these approaches alongside direct experimentation, across tissue\nengineering, gene delivery, drug delivery, protein stabilization and antifouling materials.\n\nThe practical position is that mined data is good for narrowing and high-throughput data is good for deciding. Used\nthe other way round, you get a model that is confident about a region nobody has actually measured.\n\nWhat follows for how we run experiments\n\nIf you accept that the failures carry information, some things follow. Record them with the same metadata as the\nsuccesses. Do not discard a plate because most of it did not work. Report the full screened range, not the subset that\nsupports the conclusion. And design assays so that failure produces a number rather than an absence: a solubility\nof zero is data; a well you stopped measuring is not.",
  "describedBy": "https://emanahmed.org/openapi.json",
  "license": {
    "content": "CC-BY-4.0",
    "contentUrl": "https://creativecommons.org/licenses/by/4.0/",
    "note": "Site prose and this API's metadata may be reused with attribution to Eman Ahmed (emanahmed.org). Published abstracts are reproduced from the publisher of record named in each record's `doi`; their reuse is governed by that publisher, not by this licence.",
    "attribution": "Eman Ahmed, https://emanahmed.org"
  }
}