How can you optimize AI models for drug discovery?

Data quality and quantity are the key. Current state-of-the-art libraries for small molecule drug discovery address one or the other, not both. Macroscopic High Throughput (HTS) Screening libraries are diverse and drug-like but lack the scale. DNA Encoded Libraries (DEL) have the scale but lack the quality - library members are often not drug-like; they are greasy and flat. Data quality and quantity stem from the experiments. Soon, the most successful experiment platforms will rule the AI field if their data stays proprietary.

Standardization of chemical structures such as transforming different chemical representation into canonical form is crucial data preparation step when dealing with chemical data. This process is indispensable for ensuring the comparability of structures while preserving chemical uniqueness, thereby facilitating the removal of duplicates. Additional standardization actions involve the elimination of fragments, removal of salt and solvent, and addressing tautomerization. These measures collectively contribute to enhancing the quality and coherence of chemical datasets in the context of research and analysis.

For structuring and preparing data for drug discovery, focus on the following Data Quality and Consistency: Ensure data accuracy, completeness, and standardization. Remove duplicates and handle missing values. Data Integration: Combine diverse data sources (chemical, biological, clinical) for a holistic view. Annotation and Metadata: Properly annotate datasets with metadata for context and traceability. Feature Selection: Identify and select relevant features that contribute to predictive models. Data Partitioning: Divide data into training, validation, and test sets for model development. Compliance and Security: Adhere to regulatory standards for data privacy and security.

The kingpin of AI is the data we work upon, as the choice and refinement of data directly impacts the rendered insights. For drug discovery too, the scientific communities have culminated in the notion?of personalized therapies to medically intervene, as the idea of a "one size fits all, panacea" is a passe. The "targeted" approach would mandatorily encompass a wide cross-section of data spanning variegated physiologies, and hence incubating all necessary and sufficient attributes to train models capable of identifying the required biomarkers, with fair specificities. As a rudimentary measure in preparation, the manifestation of adequate biological signatures must be ensured in the data, at par with the regular data preprocessing paradigms.

Identify and create relevant features from raw data that could significantly influence the predictive power of your AI models. In drug discovery, features might include molecular properties, patient demographics, or historical response data. Combine data from multiple sources for a more comprehensive dataset. This can include integrating genomic data with electronic health records or experimental results from different types of assays. Develop strategies for dealing with missing data, which is common in biological datasets. Techniques like imputation can be used, but it's crucial to understand how these methods might affect model outcomes.

Choosing the correct model is not always straightforward and can involve quite a bit of guesswork. One way to improve your chances of getting the right one, is to make a shortlist of 10-20 of the most suitable ones for your use case, and to test them out (validate in the step below). This works really well when the time to train a model is low, which can be the case if your datasets aren't huge.

Choose the right type of model for specific problem drug discovery problem. Deep learning models like convolutional neural networks (CNNs) or recurrent neural networks (RNNs) may be suitable for certain tasks, while traditional machine learning models like random forests or support vector machines might be better for others. In my experience ensemble methods to combine predictions from multiple models led to improved model accuracy. Incorporating the feedback of end users or decision-makers, such as medicinal chemists or molecular architects, during the initial phases of model development is instrumental in ensuring the successful integration of the model into production.

Model selection should reflect the subject matter. While empirical approaches, such as neural networks, work occasionally, they provide little insight into the mechanisms. Without understanding the underlying mechanisms, it is hard to move towards improvement. Molecular dynamics on actual target structures and pattern recognition across small molecule structures are your best bets. Without solid metrics, i.e. predicted binding energy, the experiment team will have low motivation to test your candidate molecules.

Split your dataset into training and validation sets to evaluate the model's performance. Techniques like k-fold cross-validation to assess the model's generalization across different data subsets is often helpful.

In my experience, the only way to test the extrapolation is to do the experiment. The ultimate validation is the new data. So, keep close connections in the experiment department. The sooner you can churn the cycle of prediction-validation, the faster you can get to the best drug candidates.

I was once helping a client to create a model for prediction of a batch process for biologics manufacturing. We had all of the data about the process at our disposal, even including environmental conditions -- a data scientist's dream! However, there were only 100 batches ever made, which limited our ability to select model parameters and to validate the results. We settled on "eave-one-out cross validation (LOOCV), where we essentially made 100 models, each leaving one batch out. We then used that batch for judging prediction accuracy. We plotted each prediction vs the actual result to get a read on model performance; it was pretty good all things considered! I've been a big fan of LOOCV since, for models based on limited datasets.

Tailor loss functions and evaluation metrics to suit the specific needs of drug discovery. For instance, in scenarios where the dataset is imbalanced, using metrics like AUC-ROC or precision-recall curves can be more insightful than traditional accuracy metrics. Utilize automated tools like Grid Search, Random Search, Bayesian Optimization, or advanced methods like Hyperband and Population Based Training for systematic hyperparameter optimization using platforms like Google's Vizier or tools like Hyperopt Understanding why a model makes certain predictions is crucial, especially in a high-stakes field like drug discovery. Techniques like SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations)

When training a model with person-level healthcare data, ensure you have the right guardrails in place to avoid a patient’s identification / re-identification in the model output. Train the model so that it doesn’t reveal or infer protected health information about any particular individual. For language based models, design prompts and filtering mechanisms that automatically detect and prevent any queries or responses that could lead to reidentification. This includes not only direct identifiers but also combinations of indirect identifiers that could be pieced together. Always good to have a legal and ethical oversight mechanism to ensure that the model use aligns with broader legal and ethical standards related to patient privacy.