BigHat Biosciences is betting that the real bottleneck in antibody development is asking whether a candidate can be made, and that a cell free system paired with machine learning can answer that question before scale up.
The bottleneck in antibody development has shifted. Finding a candidate that binds its target is no longer the hard part. The hard part is knowing, before committing to manufacturing scale, whether that candidate can actually be produced at the volumes and qualities a clinical program demands.
A team at BigHat Biosciences argues the constraint sits in the wet lab, specifically in the CHO cell line that anchors most mammalian biologics production. Chinese Hamster Ovary cells remain the industry workhorse for a reason: they generate antibodies with the human-like glycosylation and folding that regulators expect. Running an antibody variant through a CHO expression screen, though, is slow and expensive, which limits how many sequences a team can test for manufacturability before locking in a candidate.
At the PEGS Boston conference, Hunter Elliott, PhD, vice president of machine learning at BigHat, described the company's workaround: build machine learning models of antibody yield on top of a cell-free expression system, then use those models to screen thousands of sequence variants in silico for the small subset that looks most likely to be both manufacturable and biophysically well-behaved. Only those candidates move into the physical lab.
The distinction matters because it reframes what in silico tools are for. They are not a replacement for wet-lab biology or for CHO-based process development. They are a way to compress the learning loop between sequence design and manufacturability prediction, surfacing problems earlier than a traditional pipeline can.
Elliott framed the gain in operational terms. Combining in silico screening with physical experiments "derisks" antibody development by catching manufacturability problems before they become scale-up crises, according to the PEGS Boston write-up. The model learns from cell-free expression data, which is faster and cheaper to generate than CHO titer runs, and the predictions guide which sequences are worth expressing in mammalian cells at all.
That approach also pushes back on a common objection to computational screening: that it might discard the best antibody simply because the starting sequence looks "suboptimal" or "imperfect." Elliott's counter, reported at the panel, is that a weak-looking starting sequence can be optimized through several rounds of in silico mutagenesis and engineered into manufacturability rather than killed early. The screen functions as a design tool, not a filter.
The larger gap the panel surfaced is organizational. A second speaker, described in the GEN write-up of the PEGS Boston session as a preclinical and manufacturing communication advocate, argued that harder-to-manufacture drugs require earlier and more sustained dialogue between the teams that discover candidates and the teams that will eventually make them. In silico tools do not solve that handoff. They make the inputs to the conversation more legible by attaching a manufacturability score to each candidate, but the conversation itself still has to happen.
The honest framing, then, is that cell-free expression paired with machine learning is a constraint-reliever, not a silver bullet. It expands the mutation space a small team can explore before committing to scale. It does not replace CHO cells, regulatory comparability work, or the cross-functional alignment that biologics development has always required.
The public data limitation is real. Elliott acknowledged at the panel that there is a shortage of publicly available manufacturability and developability data, especially for candidates that failed before reaching the clinic. That shortage shapes what any in silico model can learn and what claims its users can credibly make, and it is part of why the BigHat approach has to be read as one company's operating bet rather than a settled industry method.
What to watch next is whether other groups publish yield-prediction models trained on cell-free data that hold up across different antibody scaffolds, and whether any of those models get used inside a CMC filing or a process technology transfer rather than only in early discovery. That is where the approach moves from a credible research tool to a manufacturing-relevant one.