AI & Drug Discovery - True Disruption

Medical science currently provides a pitiful return on both time and money:  the UK’s Welcome Trust conducted a study with the RAND Corporation(https://wellcome.org/sites/default/files/wtp056595_0.pdf), and in this study, “Of the interventions considered between 1991 and 2010, smoking reduction accounted for 65% of the net monetary benefit to the UK, followed by cervical screening (24%) and breast cancer treatment (10%).”  Furthermore, the study illustrated a 15 year lag between funding for medical research and its impact in cancer treatment.

However, the results are even worse in the drug industry: 90% of clinical drug programs fail (https://pmc.ncbi.nlm.nih.gov/articles/PMC9293739/):

In the past few decades, each step of the above drug development has been rigorously optimized and validated, while many successful strategies have been rightfully implemented in the drug development process to select the best drug candidates for clinical studies. Despite this validated effort, the overall success rate of clinical drug development remains low at 10%–15%... Such persistent high failure rate raises several questions: Why 90% of clinical drug development fails despite implementation of many successful strategies in the past several decades? Did we overlook certain aspects of drug development process leading to high failure? How can we improve the success rate of clinical drug development? 

Of this 90% failure, roughly half of it is attributed to “lack of clinical efficacy”:  translation:  the drug fails to deliver a statistically-significant clinical outcome (i.e. more efficacious than placebo).  Toxicity is then the second leading factor (30%) of failure, where the cure indeed proves worse than the disease.  The remaining 20% is split between poor drug-like properties (e.g. fragility of the drug, itself) and commercial issues.

Thus, with the capital required to develop drugs only increasing over time and the yield of this investment proving so poor, it leaves little to our imagination why Americans are “expected to pay so much” for their drugs:  precisely because of the one drug that goes to market, roughly nine others fail in the process!

LLM’s Alone Are Not Enough:  The Power of Composition & Multi-Modal Models

Because AI is automated, it can quickly and cheaply help scientists converge on viable solutions.  This is an iterative process, and a great example of this is the ongoing evolution of retrosynthetic planning, where chemists determine the best commercially-viable recipe for making a given drug or polymer.  At the moment, strict LLM models in retrosynthetic planning struggle with bias, because of the dearth of unpublished failed studies and other factors in published papers: (https://www.chemistryworld.com/opinion/slow-march-of-the-retrosynthesis-robots/4018263.article)  The current viable approach for using strict LLM’s on the retrosynthetic planning application often entail the inclusion of an “expert” to help guide or correct bias within the LLM (https://gwern.net/doc/reinforcement-learning/model/alphago/2024-striethkalthoff.pdf).

Despite the bias issue when applying a strict LLM to the application of retrosynthetic planning, multimodal models  present an opportunity for composing a LLM trained on published chemistry data with, say, graph models that capture deeper physical properties for molecules and chemical reactions.  The power of composing LLM’s with other models unleashes vastly improved results, and the research has only just begun on how to best compose models within a LLM to produce meaningful predictions that significantly accelerate the process of retrosynthetic planning.  In the case of this paper’s composition of a LLM with a graph model for molecular structure and reactions, its algorithm, Llamole, “…increases the success ratio from 5% to 35% for drugs and to 17.9% for polymers.” (https://arxiv.org/pdf/2410.04223)

Pharmacokinetics (PK):  Conquering Sparse(or slow & expensive) Data

Pharmacokinetics strongly impact both drug efficacy and toxicity.  As mentioned above, efficacy and toxicity combined account for 70-80% of clinical drug failure.  Currently, the state-of-the-art is to model Pharmacokinetics with differential equations fit using non-linear mixed effect models (NLME).  

A disruption in PK modeling would come from accurately estimating the coefficients in the existing PK models without physical experimentation.  Very often, PK datasets are sparse, thus requiring either costly, slow experimentation and study to infer key constant values to solve the PK differential equations, or else estimations that may prove statistically intractable. AI Multimodal models show great promise in a new path to overcoming this challenge.

In a study from 2022, a multimodal model achieved estimates for PK equation parameters(clearance (CL) and volume of distribution (Vd), specifically) for humans comparable to those from animal experiments:

… the prediction accuracy of this method is comparable with the many animal scale-up methods. Different from the conventional methods, since these models do not need new experimental data, it seems to be appropriate for predicting the human parameters in not only the clinical stage but also the early drug discovery stage. (https://pubs.acs.org/doi/10.1021/acs.jcim.2c00318)

Another successful paper demonstrates how a sequence of deep learning models successful estimated a key coefficient, the tissue-to-plasma coefficient, in modeling drug distribution.  The tissue-to-plasma coefficient represents the steady state concentration ratio between plasma and the surrounding tissue, e.g. the concentration of a compound within fatty tissue vs. the fatty tissue's plasma compound concentration. Furthermore, the challenges in estimating specific tissue-to-plasma coefficients is not uniform: the liver, for example, proves extremely complicated to characterize and thus estimate its tissue-to-plasma coefficient.

The key contribution of this study was inferring values in the absence of data:  by using deep learning to identify structure and relationship the data, the missing  (https://pubmed.ncbi.nlm.nih.gov/38639496/) data no longer required extensive experimentation to determine the tissue-to-plasma("Kp") coefficient, in particular, the costly and difficult-to-estimate Kp for the liver proved un-needed for estimating other tissue Kp values:

Significant improvements were observed in the Kp values of adipose tissue, brain, kidney, liver, and skin. These improvements indicated that the Kp information of other tissues can be used to predict the same for a specific tissue. Additionally, we found a novel relationship between each tissue by evaluating all combinations of explanatory variables. For example, Kp values in no other tissues were needed for the prediction of adipose Kp, and liver Kp was not needed for the prediction of Kp in other tissues.

Conclusion: Clinical Drug Success Rate Will Only Increase

While it is cliche to claim AI will "change the world", it will indeed change the ROI for clinical drug development: this article has already revealed, in a tiny sample of recent papers, the massive progress well-underway in applying AI to catalyze far, far faster and more successful clinical drug development.

Further Reading - Drug Competitors: LLM's Advancing Gene & Protein Therapy

This article did not touch on two other major topics, those of LLM's for both genomics and protein design. Gene and protein-based therapies are completely different from conventional drug design and offer novel approaches where drug-based solutions have failed.