NAMS: In silico and in vitro approaches for Drug Development

The FDA has set forth new regulatory guidances around NAMS.

These NAMS are mainly centered around two arms: (1) in silico and (2) in vitro replacements to reduce studies in animals and promote sensible paths for approving new medicines. In particularly, the new guidelines feature cases where animals studies are less relevant to predicting human biology (eg. Immunology).

Part I. In Silico Approaches

In silico approaches are the lower hanging fruit of the two — lower cost and faster to implement. Additionally, there are good foundational layers for building in silico approaches which had been built upon years of research in thermodynamic modeling and simulation to predict physical properties. For example, modeling the conformational stability and folding equilibrium of macromolecules to predict protein folding or predicting solubility, boiling point, and free energies of small molecules. These models are build on the fundamental elemental properties of atoms and conform to physical laws of nature. The neatness of these predictive models is very admirable.

Biology is messier. Commonly used are models that correlate these chemical structures with biological activities, like QSAR and QSPR are less reliable. It’s worth repeating, that these models are correlated. Importantly, the models are guided by a set of assumptions and these assumptions require a fair amount of scrutiny. There is complexity in the biological and physical realities — how the molecule fits into the pocket — that may be outside the scope of modeling parameters, and therefore, are less reliable in predicting behavior of new molecules. Biology is more like a web where pulling on one point moves others in some correlated but sometimes unpredictable way. So while designing small molecules or amino acids can neatly fit within the realm of modeling where parameters are well defined, and model predictions are close to what can be manufactured, biology is a restless. Homeostatic mechanism, futile cycles of biosynthesis and degradation, signaling cascades — cells want to keep correcting, maintaining, adapting, growing and so perturbations require actual experimentation.

Like reading a guidebook on the various insect species in New England. We assume we will encounter these common insect species on our hike and when we see those species, we can pat ourselves on the back when we positively identify the beetle species from page 6. However, when a new invasive species is brought to the region on wooden shipping crates, it can only identified as something unusual. That is, until we have new information, use Image Search, or consult an ornithologist who is familiar with the Asian Longhorned Beetle and can confidently identify this strange insect.

So while in silico approaches and the power of ML are great at some tasks, it is not yet powerful to describe the biology of what is not already known (frontier science). Thus, I would caution its use for predicting biological phenomena, much of which is yet to bet discovered. Thus, much of the in silico developments in Biopharma have been in the small molecule and more recently with improved performance of protein folding prediction — in biologics.

Biopharma’s In silico “AI” integration

LLMs

Significant effort has also been made in data curation or organizing information. The first flavor is the curation of literature databases using LLM. For every 5 companies with a shiny “AI” label , only 1 may actually be doing something with longterm vision. Yes, the work of querying multiple databases has been facilitated by new AI platforms like how internet searches and digitization of text has facilitated access to information. While these platforms have reduced some of the manual work of querying multiple databases, as a scientist, I need to understand what and how it’s doing and much of what is beneath the hood is proprietary. I’m less likely to trust the output of the search if I cannot scrutinize the assumptions made for the model inputs or am relying on assumptions made by computer scientist without experience in understanding nuances in biology or pharmaceutical development.

Similarly, services providers are increasing offering the use of LLMs for filing regulatory documents. Theres a gaggle of companies offering solutions to ease the pain of paperwork. Many are backed by experienced regulatory professions that have actually walked the walk. So the use of LLMs will facilitate the needed documentation and administration.

I am excited to see what comes out of this new wave of democratizing information using LLMs. While access to information across disciplines has potential to facilitate new insights, it’s unclear if in practice, we are actually seeing these Eureka moments that can happen when a scientist trained in one discipline moves to another field and has a major breakthrough (eg. Francis Crick, Cecilia Payne-Gaposchkin, Ivan Pavlov (Krauss 2024)).

Machine-Learning Approaches

Another flavor that has seen significant in silico development has been in the curation and use of publicly available databases, with many companies promising new target identification. Within this context, it will be important to identify differences between anomolies and noise in the data. While large publicly available databases can be powerful engines, again, there’s a lack of evidence that these can deliver bespoke solutions for specific cases and I’d be curious to find good examples where companies have delivered on promises of “new target ID”.

The most common use cases in biopharma for these new tools have been to filter the number of potential small molecule candidates that should be tested or optimizing immunological properties (Fc region) of biologics. Narrowing the pipeline of candidates early.

The true unlock is when these in silico and in vitro approaches are meaningfully integrated. The in vitro approaches which can reliably model the disease and are well supported by literature reviews and readouts can recapitulate real biology.

Stay tuned for Part II. In Vitro