This 24-Year-Old From Bengaluru Is Betting Against Expensive AI Drug Discovery

For years, the biggest breakthroughs in AI-driven bioscience have come from Silicon Valley, London, or research labs backed by Big Tech. Google DeepMind gave the world AlphaFold. David Baker's lab changed protein design. Companies like Anthropic are now throwing AI agents into drug discovery.

India was largely watching. But a 24-year-old founder from Bengaluru wants India to drive such breakthroughs.

Anindyadeep Sannigrahi founded LiteFold just under a year ago. Now, his startup has released LiteMol-1—among India’s first multi-molecule foundation models for molecular design. 

But unlike AlphaFold or the growing list of structure-based AI models, LiteFold is taking a different bet altogether. Instead of building a bigger, more expensive model that reasons over complex 3D protein structures, LiteMol-1 tries to make molecular design cheaper, faster, and native for AI agents.

“Everybody is focusing on structure because it's intuitive,” the founder tells AIM. “We thought, what if we looked at biomolecule design from a totally different viewpoint?”

Making Drug Discovery Cheaper

AlphaFold changed structural biology by predicting the 3D structure of proteins from their amino acid sequences, a problem that had taken scientists years to solve through laboratory work. That breakthrough inspired Sannigrahi and his Delaware-based co-founder Cory Kornowicz to enter the field. 

"I was super motivated when AlphaFold got the Nobel Prize," Sannigrahi says. "Maybe this is the first time and the right time where we can truly invest in AI and understand biology."

But AlphaFold also created a new default in AI drug discovery: think in structures.

Models like RFdiffusion, BoltzGen, and O-Design now generate protein and molecular structures around biological targets. Pharmaceutical companies increasingly use these systems to search through millions of possible molecules.

LiteFold believes that approach is becoming computationally expensive.

"Doing work on a structure paradigm is just very expensive and very time-consuming," Sannigrahi notes, as drug discovery is essentially a funnelling process—using advanced computation to filter out millions of failures.

More importantly, binding a molecule to a target isn't the only goal.

"You may get a molecule that binds. But what about toxicity? Absorption? Will it even reach your bloodstream if you take it orally?"

That's where LiteMol-1 starts to diverge from traditional models. It isn’t designed to generate just one kind of biomolecule. The model can generate small molecules, peptides, cyclic peptides, macrocycles, depsipeptides, PROTACs, and peptides containing non-canonical amino acids—all using a single set of weights. It can also edit existing molecules instead of generating them from scratch.

But the bigger change is how LiteFold thinks about optimisation.

Most molecular models largely optimise around binding affinity. LiteMol-1 is built to optimise across multiple therapeutic objectives simultaneously: toxicity, ADME properties, drug-likeness, membrane permeability, selectivity, synthesisability, and more.

Instead of trying to find one "perfect" molecule, LiteFold uses Monte Carlo Tree Search—a technique to optimise decisions in highly complex spaces—to keep multiple promising candidates alive, each representing different trade-offs. "Molecular design is fundamentally not a single-objective optimisation problem," Sannigrahi remarks.

Sequence Over Structure

Sannigrahi believes drug discovery is moving towards autonomous research loops where AI systems repeatedly generate molecules, evaluate them, modify them, and generate again. For those systems, structure becomes a surprisingly bad interface.

"Structure is very intuitive for human beings. But it's very hard for agents because it's a huge amount of context," he explains.

Instead of outputting massive 3D structural files, LiteMol-1 works with compact molecular sequences like SMILES (Simplified Molecular Input Line Entry System) that represent complex chemical structures.

"For us, what we are trying to output is not a structure. It's a sequence." That dramatically reduces the amount of information an AI agent has to process during each reasoning cycle.

LiteFold says the model becomes an "infinite molecular canvas". The model proposes possibilities. The agent inspects them, edits them, compares them, regenerates them, and only invokes expensive structure prediction or docking when absolutely necessary.

It also changes the economics. "We just shrunk the token cost by a huge amount," Sannigrahi says. "Our model can run on an NVIDIA L40."

That's a notable contrast to traditional high throughput screening of AI-assisted drug discovery that requires hundreds of GPUs and tens of thousands of dollars worth of compute for a single workflow. In fact, pharmaceutical company Eli Lilly built an AI supercomputer equipped with more than 1,000 NVIDIA Blackwell B300 GPUs dedicated to automated molecule design.

The Problem With AI Drug Discovery

But this isn't a sequence-versus-structure debate. "There are areas where sequence models work extremely well, and there are areas where structure models are essential," he says.

LiteFold still relies on structural models like AlphaFold for understanding molecular interactions and dynamics. "If two molecules are interacting, structure becomes very important because that's where you understand bond patterns and intermolecular forces."

The company's philosophy is simpler: use sequences when they make search cheaper and use structures when they make biology clearer. The real opportunity lies in combining both.

Sannigrahi was surprisingly critical of the current excitement around AI in drug discovery. "In terms of accuracy, we are not really quite out there yet. I don't think it's good enough to just say we are doing great and everything."

According to him, AI has become good at generating molecules, but it's still weak at the two hardest problems in biology—understanding disease and predicting what happens after a molecule enters a biological system.

The industry is generating more molecules against the same known targets instead of discovering better biological targets in the first place. "The speed is there. But they're not really hitting the target."

The India Advantage

The next bottleneck isn't compute. It's biology. Sannigrahi believes biology already has enough data to build better foundation models. The problem is that researchers aren't representing that data in ways modern AI systems can learn from efficiently.

"We are just not seeing the right representation." Sequence data is abundant. Structural data is limited and expensive to produce.

If AI can learn richer biological representations from sequences, it could unlock entirely different foundation models for disease understanding, molecular generation, and prediction. That's where LiteFold wants to spend its research effort.

"We want to understand biology," Sannigrahi notes. Then comes forecasting biological outcomes before starting any experiments.

Sannigrahi believes India may have an advantage here because of its massive gene pool.

"I am super excited about the amount of genetic diversity India has." India's population contains enormous genomic diversity shaped by geography, migration, and endogamy. That is valuable for understanding disease targets and therapeutic responses.

Then comes manufacturing.

Companies like Biocon, Sun Pharma and Dr Reddy's have already built India's pharmaceutical manufacturing ecosystem. Clinical trials have also become more accessible.

Put those together, and Sannigrahi believes India has an opportunity that hasn't been fully exploited. "Biotech companies in India should start their own therapeutic programmes," he suggests. 

He believes the next breakthroughs will come from combining AI with organoids, automated laboratories, and biological circuits that generate richer experimental data.

LiteFold is already building an AI co-scientist called Rosalind, named after Rosalind Franklin whose work led to the discovery of the double-helix structure of DNA. The AI co-scientist automates research workflows and assists scientists during molecular design. But even there, Sannigrahi insists humans stay in the loop.

LiteMol-1 is still an early research model. It hasn't discovered a blockbuster drug. It hasn't shortened clinical trials or cured cancer.

Sannigrahi isn't claiming any of that. He thinks meaningful breakthroughs are still five years away.

But LiteMol-1 is doing important work and attempting to ask a question that Big Tech hasn’t even thought of. What if the future of AI drug discovery isn't a bigger model?