Well, well, well — halfway through 2025 already! Mixed emotions, mixed economy. We’ve all become hyper-aware of everything. The flood of unnecessary information on social media reels is, at times, more damaging than the trauma dumps passed down by older generations.
We get misinformation faster than ever, yet we’ve become slower at making decisions — confused and overwhelmed by the chaos around us. The sluggish economy hasn’t helped either; it’s made us collectively dull. There’s no real culture of thriving or growth anymore — just little dopamine hits from scrolling through reels.
AI’s role in social media is brutal. But in life sciences? It’s nothing short of heroic.As you may know, life sciences, biotech, and drug discovery are super hot right now — especially after COVID. Faster drug discovery has become a top priority for many big pharma leaders.
Drug discovery is a time sink. On average, it takes 10+ years and billions of dollars to bring a single drug to market. Attrition rates are high, clinical trial failures are frequent, and many diseases still have no viable treatment.
The traditional pipeline — screening, target validation, hit optimization, preclinical, clinical — has hit a wall.
So what happens when we inject AI into this process?
Let’s just say the industry might finally get a speed upgrade that doesn’t compromise accuracy.
From Wet Lab Bottlenecks to In Silico Acceleration
Lydia The, Christoph Sandler, and Alex Devereson (McKinsey) make a strong case: AI isn’t just a tool to automate existing processes — it’s a paradigm shift in how we design, test, and translate new drugs.
Here’s how:
- Multi-modal data ingestion: AI can integrate disparate data types — EMRs, genomic sequences, wearable sensors, clinical trials, and literature — to create a high-dimensional patient representation.
- Generative chemistry: Foundation models for molecules (like AlphaFold or ESMFold) are enabling de novo drug design. These models predict protein structures, generate novel ligands, and simulate binding affinities — all in silico.
- Digital twins for biology: Imagine a model simulating disease progression in virtual patients. These “digital twins” can reduce reliance on early-stage animal models or poorly stratified human trials.
- Adaptive clinical trial design: ML enables real-time feedback loops during trials — optimizing dose cohorts, predicting adverse events, and adjusting protocols dynamically.
Why Is This Not Mainstream Yet?
Because most organizations get stuck in what McKinsey aptly calls pilot purgatory.
- They try AI.
- They see localized success.
- But they never scale.
Reasons?
- Lack of data infrastructure.
- No org-wide buy-in.
- Misalignment between R&D and data science teams.
- No centralized strategy for integrating models into decision-making workflows.
Technical Blueprint to Break the Cycle
If you’re running drug discovery at any stage — early target discovery to Phase II — you should be thinking in terms of:
- North Star Alignment
Define the problem-AI-fit, not just product-market fit. Set scientific goals (e.g., reduce time-to-lead by X%) and track them across the org. - End-to-End Data Strategy
Your AI models are only as good as the data pipelines they sit on. Start with a secure, unified data lake — supporting structured (clinical), unstructured (PDFs, publications), and imaging data. - Model Deployment Infrastructure
Don’t keep models in notebooks. Productionize with containerized services (Docker, K8s, Airflow), CI/CD for model updates, and robust MLOps for monitoring and drift detection. - Trust Frameworks
Explainability matters. Especially in high-stakes settings. Implement SHAP, counterfactuals, or prototype-based explainers alongside predictive outputs. - Tangible ROI in ≤3 Months
Choose use cases with high data readiness and measurable outcomes. Examples: AI-augmented compound screening, trial dropout prediction, or biomarker-based patient stratification.
AI Won’t Replace Scientists — But It Will Replace Manual Pipetting
This isn’t about automating scientists out of the lab. It’s about allowing them to move beyond repetitive, manual work and focus on hypothesis generation, clinical validation, and scientific insight.
The result?
- Faster pipelines.
- Lower failure rates.
- More personalized therapeutics.
- Higher confidence in early-stage bets.
But the question is — HOW?
Let’s walk through a simplified, hands-on example — from molecule generation to activity prediction — using tools like RDKit and PyTorch.
Step 1: Generate Molecules with RDKit
Let’s start with the basics: generate and screen small molecules.
from rdkit import Chem
from rdkit.Chem import Draw
from rdkit.Chem import Descriptors
# Example SMILES (simplified molecular-input line-entry system)
smiles_list = [
"CC(=O)OC1=CC=CC=C1C(=O)O", # Aspirin
"CN1C=NC2=C1C(=O)N(C(=O)N2C)C", # Caffeine
"CC(C)CC1=CC=C(C=C1)C(C)C(=O)O" # Ibuprofen
]
mols = [Chem.MolFromSmiles(smi) for smi in smiles_list]
Draw.MolsToGridImage(mols, molsPerRow=3)
Now, let’s compute some simple descriptors (e.g., molecular weight).
for smi in smiles_list:
mol = Chem.MolFromSmiles(smi)
mw = Descriptors.MolWt(mol)
logp = Descriptors.MolLogP(mol)
print(f"{smi}: MW = {mw:.2f}, LogP = {logp:.2f}")
Step 2: Predict Bioactivity with PyTorch
We’ll mock a small neural net to predict a molecule’s activity score (e.g., against a protein target).
import torch
import torch.nn as nn
import torch.nn.functional as F
# Dummy feature extractor: mol weight and logP
X = torch.tensor([
[180.16, 1.19], # Aspirin
[194.19, -0.07], # Caffeine
[206.29, 3.50] # Ibuprofen
])
y = torch.tensor([0.9, 0.1, 0.6]) # Fake bioactivity scores
# Define a simple MLP
class BioactivityNet(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(2, 16)
self.fc2 = nn.Linear(16, 1)
def forward(self, x):
x = F.relu(self.fc1(x))
return torch.sigmoid(self.fc2(x))
model = BioactivityNet()
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
loss_fn = nn.MSELoss()
# Train the model
for epoch in range(200):
pred = model(X).squeeze()
loss = loss_fn(pred, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Inference
predictions = model(X).squeeze().detach().numpy()
print("Predicted bioactivities:", predictions)
This is of course oversimplified — but it mimics how you’d start building predictive models for things like IC50, pIC50, or logBB.
Step 3: Scale with MLOps + Pipelines
In real-world AI drug discovery, you’d need to:
- Preprocess molecular data (via SMILES, Mol2Vec, GraphConv)
- Train deep learning models (GNNs, VAEs, Transformers)
- Serve predictions via API
- Monitor drift and model confidence
Here’s a quick sketch of what a production pipeline might look like:
# Run with Airflow / Prefect / Kubeflow
fetch_raw_molecules --> featurize.py \
--> train_model.py \
--> evaluate_model.py \
--> push_to_registry.py \
--> deploy_model.sh
Bonus: Use GNNs (Graph Neural Networks)
SMILES is just one format. Molecules are naturally graphs — atoms as nodes, bonds as edges. Here’s where GNNs shine:
- Use tools like PyTorch Geometric, DeepChem, or DGL
- Convert RDKit mols into graph tensors
- Train GCNs or message-passing networks for better bioactivity predictions
pip install torch-geometric
Example models: ChemProp, Junction Tree VAE
Code Is the New Molecule
Biotech is crossing a threshold — from intuition-based research to inference-driven iteration. The playbook:
- Build pipelines that fuse chemistry and computation
- Shift drug discovery from wet lab → in silico → wet lab
- Use AI to prioritize what’s worth testing in vitro
Once this loop is tight enough, drug discovery won’t take 12 years. It’ll feel like running a model update.
And that’s a future worth debugging toward.
Conclusion
The transformation is already underway. But without:
- infrastructure,
- cross-functional integration,
- and strategic alignment —
you’ll just keep running pilots forever.
If biotech companies want to move from science-as-art to science-as-code, AI isn’t optional. It’s the new substrate.
Follow a tech philosopher’s perspective on AI and the world, while navigating an AI career.