Page 42 - AIH-2-2

Artificial Intelligence in Health New drug discovery in the AI era

3.4. Deep learning (DL)

DL, a subbranch of ML, learns from algorithms and
their outcome data to further improve using neural
networks. This disruptive technology has been effectively
44
applied in various complex scenarios. For instance, NCEs
may have weak interaction strengths but be highly target-
selective, meaning they exhibit strong target selectivity but
relatively low absolute potency. In these cases, the goal is
to balance NCE potency and selectivity, finding the most
selective molecule with the minimum desirable potency.
Tools such as Affinity2Vec, DeepDTA, and DeepGS can
predict drug–target binding affinity, assign binding affinity
scores, and rank compounds. 45-47

One approach to improving a ligand’s druggability
Figure 8. Role of machine learning in drug toxicity prediction during the lead optimization phase is to expand it by
adding a single chemical group (fragment). Fragment-
and BenevolentAI’s dermatitis drug BEN-2293) was based drug design, on the other hand, involves adding
a setback. However, BenevolentAI’s repurposed drug multiple fragments. Geometric DL helps expand the ligand
Olumiant (baricitinib, a rheumatoid arthritis therapy) was by identifying the site(s) on the ligand to add fragments,
approved by the US FDA in May 2022 for the treatment of suggesting the most suitable fragments, and predicting the
COVID-19. geometry of the added fragments. 48

3.3. ML Furthermore, DL tools predict chemical toxicity
by comparing millions of known substances based on
ML, a subfield of AI, uses defined datasets to create biological mechanisms or physicochemical features. DL
algorithms for developing predictive and descriptive models, algorithms trained on datasets of well-known medications
which are useful for analyzing data, drawing insights, and can accurately forecast the activity of NCEs. 49
making informed decisions. There are two types of ML
models based on the data used to generate an algorithm: 3.5. GenAI
Supervised and unsupervised. Supervised ML models are GenAI, a subset of DL, creates new content based on
trained with labeled input data and a defined outcome, learned information. As one of the most advanced
whereas unsupervised ML models use raw input data to find forms of AI, GenAI can generate new molecules from
relevant patterns and relationships without prior knowledge. training data. To develop novel molecules for specific
In recent years, ML has become a powerful tool in applications or predict their behavior in biological
drug discovery, transforming the way we investigate environments (e.g., receptor binding), algorithms are
and understand large, complex information about drug trained on the chemical–physical features and 3D forms
behavior in biological systems. ML predicts novel drug– of molecules.
target interactions with reasonable accuracy. Clinical trial GenAI, coupled with data analytics, can design
43
datasets have been used to build ML models that forecast structures with optimal druggability attributes and predict
trial success early, thereby reducing unrecoverable costs their physicochemical properties, drug–target interactions,
and saving time. potency, efficacy, and toxicity with reasonable accuracy. It
Large language models (LLMs), a component of AI’s also aids in designing the most suitable drug formulation
natural language processing that overlaps with ML, are and delivery system, which improves stability and oral
gaining importance and popularity. By drawing scientifically bioavailability. In addition, GenAI can reduce the time
valid conclusions from large datasets, such as genomic, required for regulatory dossier submissions. During clinical
proteomic, and metabolomic data, and existing literature, evaluation, GenAI assists in drafting the best-fit clinical
LLMs help researchers generate hypotheses and make sense trial design and patient recruitment strategies. AI tools
of voluminous experimental data. LLMs effectively analyze help identify novel biomarkers and surrogate endpoints
large biological datasets to predict new druggable targets to predict patient responses to treatment. Furthermore,
that conventional approaches may miss. Furthermore, GenAI can infer safety and tolerability signals for early
LLMs assist in drug repositioning and repurposing. intervention, improving clinical trial success rates.

Volume 2 Issue 2 (2025) 36 doi: 10.36922/aih.4423