Inspiration
What if an AI could not only predict how a drug interacts with a protein, but also reason about why that interaction changes when cancer mutates?
Cancer treatment is a constantly evolving molecular challenge.
Consider EGFR-driven lung cancer. A targeted drug might successfully inhibit a cancer-driving protein, but a mutation can change the protein's behaviour, allowing the cancer to develop resistance. Scientists must then determine what changed, why the treatment stopped working, and which alternative molecules might remain effective.
Today, we have powerful molecular foundation models capable of predicting protein structures and molecular interactions. We also have language models capable of reasoning over scientific literature.
But these capabilities are largely disconnected.
We were inspired by a simple hypothesis: molecular prediction and scientific language reasoning should not have to exist in isolation.
Could a specialised language model learn the biochemical relationships underlying drug resistance? Could it complement a molecular predictor, and could aligning these two forms of intelligence eventually accelerate drug discovery?
That question became MutationScientist.
What it does
MutationScientist aligns molecular prediction with mutation-aware scientific reasoning to investigate cancer drug resistance and support targeted drug discovery.
Our system brings together two complementary foundation models:
1. Boltz-2 — Understanding molecular interactions
We adapt Boltz-2's affinity prediction capabilities to EGFR–inhibitor complexes. Given a mutated EGFR protein sequence and a ligand, the molecular model investigates protein–ligand interactions and produces structure- and affinity-related predictions.
2. Qwen3.5-4B — Understanding the science behind those interactions
We fine-tune a language foundation model on EGFR-specific scientific instructions to interpret mutation mechanisms, drug resistance, molecular relationships, and experimental findings.
The language model is trained across five scientific reasoning dimensions:
- Causality: Why does a mutation change molecular behaviour?
- Effect: What biochemical or functional consequences follow?
- Drug resistance: Why might an inhibitor lose effectiveness?
- Structure–activity relationships: How do molecular properties relate to drug activity?
- Counterfactual reasoning: What alternative inhibitors or experiments might help investigate the mechanism?
An evidence-retrieval layer connects these capabilities to biochemical measurements, mutation annotations, and published research.
For example, when investigating an EGFR resistance mutation, MutationScientist can connect molecular interaction predictions with the mutation's known biological mechanism, relevant experimental evidence, and hypotheses for further investigation.
The goal is to move from predicting molecular interactions to understanding their scientific implications.
How we built it
We developed MutationScientist as a dual-model research system, with a shared mutation-aware dataset connecting molecular modelling and language modelling.
Step 1: Building the research dataset
We assembled an EGFR-focused scientific dataset using information from ChEMBL, BindingDB, UniProt, RCSB PDB, PubChem, COSMIC, and published literature.
The resulting molecular dataset contains:
- 18,189 experimental measurement records
- 6,843 protein–ligand complexes
- 16 EGFR mutation variants
- Protein sequences, ligand SMILES, multiple-sequence alignments, affinity labels, and scientific provenance
We also constructed 1,341 scientific instruction examples for language-model specialisation, spanning mutation understanding, experimental reasoning, inhibitor comparison, resistance mechanisms, and research hypotheses.
Shared mutation and compound identifiers allow the two modelling pipelines to work with connected biological evidence.
Step 2: Specialising Boltz-2
We used the Boltz-2 affinity fine-tuning implementation to adapt molecular prediction to our EGFR–inhibitor dataset.
Our pipeline prepares mutant protein sequences, ligand representations, MSAs, and experimental affinity labels for molecular training and inference.
The reported molecular optimisation loss decreased from 0.3412 to 0.0595, an 82.56% reduction.
Step 3: Fine-tuning the scientific language model
We adapted Qwen3.5-4B through supervised instruction tuning using our scientific dataset.
Rather than relying entirely on a general-purpose model's existing knowledge, we trained it to model EGFR-specific scientific responses involving causal mechanisms, experimental interpretation, and drug resistance.
On our held-out language-modelling evaluation, loss decreased from 2.23768 to 0.19399, a 91.33% reduction.
Step 4: Evaluating scientific specialisation
We developed a semantic evaluation workflow that compares scientific responses at the level of individual factual claims rather than exact wording.
Using 215 evaluation prompts, we compared the pretrained model's responses against fine-tuned reference responses, measuring factual precision, recall, and F1.
This allowed us to quantify how significantly domain specialisation changed the model's scientific outputs.
Step 5: Connecting the system
We connected the research components through a shared evidence layer and application interface.
Our implementation includes Python-based data preparation and model services, scientific evidence retrieval, molecular prediction workflows, and an interactive application. We also explored a voice-based scientific explainer using Anam.
The result is a research workflow that brings molecular prediction, scientific interpretation, and evidence together.
Challenges we ran into
Our biggest challenge was making two very different types of AI work within the same scientific problem.
Molecular models learn from protein sequences, ligand representations, and biochemical labels. Language models learn from tokens, scientific instructions, and contextual explanations. Connecting these representations required a dataset that could support both approaches while preserving their underlying biological relationships.
We also faced several practical challenges.
Scientific data fragmentation. Mutation records, compound structures, binding measurements, and published evidence were distributed across multiple sources. We had to reconcile identifiers, measurement conventions, and protein–ligand records into a consistent dataset.
Limited GPU resources. Boltz-2 molecular training is computationally demanding, particularly when handling long protein sequences and molecular representations. We worked with NVIDIA T4 GPUs and an A30 environment, adapting our workflow around memory, storage, and compute constraints.
Model specialisation. General-purpose language models can produce fluent scientific explanations, but fluency alone does not ensure mutation-specific reasoning. We needed to construct scientific training examples and design an evaluation that could capture differences in factual content.
Evaluating reasoning. Scientific answers can be expressed correctly in many different ways. Simple word-overlap metrics are insufficient, so we implemented claim-level semantic comparisons to examine factual precision and coverage.
Working through these constraints became an important part of our research.
Accomplishments that we're proud of
1. Demonstrating substantial domain adaptation in both models.
We observed an 82.56% reduction in reported Boltz-2 optimisation loss and a 91.33% reduction in held-out language-model loss after specialisation.
These are different optimisation measurements, but together they demonstrate the impact of adapting foundation models to a focused molecular research domain.
2. Quantifying how differently a general-purpose LLM reasons about specialised science.
Against our fine-tuned scientific reference responses, the pretrained language model achieved:
| Metric | Pretrained model |
|---|---|
| Factual precision | 55.61% |
| Factual recall | 46.95% |
| Factual F1 | 48.05% |
The fine-tuned responses served as the reference for this comparison. The results reveal substantial differences in scientific content coverage, motivating further investigation of domain-specialised reasoning.
3. Building a shared molecular–language research dataset.
We connected thousands of protein–ligand examples with mutation-aware scientific instructions, making it possible to explore relationships between molecular prediction and language-based scientific understanding.
4. Making scientific reasoning evidence-centred.
Our system links mutation mechanisms and model outputs with measurements, scientific references, and experimental context.
5. Developing a working research prototype under hackathon constraints.
Despite limited compute, demanding molecular preprocessing, and a fragmented biomedical data ecosystem, we brought together specialised models, a connected dataset, evaluation workflows, and an application prototype.
What we learned
Our most important learning is that scientific specialisation can substantially change what a foundation model learns to predict.
A molecular model and a language model approach chemistry from fundamentally different perspectives.
One learns representations useful for predicting molecular structures and interactions. The other learns statistical and semantic relationships expressed through scientific language.
Yet both can be adapted to the same mutation–inhibitor research domain.
Our experiments showed that instruction tuning significantly improved the language model's fit to EGFR-specific scientific responses, while molecular affinity-head adaptation reduced its reported optimisation loss.
This suggests an exciting direction: molecular understanding and scientific reasoning may be complementary learning objectives, rather than completely separate AI capabilities.
We also learned that scientific AI depends on more than model size. Dataset design, domain-specific supervision, biological provenance, and evaluation methodology can be just as important.
Most importantly, we learned to distinguish generating a convincing explanation from producing a useful, evidence-grounded scientific hypothesis.
What's next for MutationScientist
Our long-term vision is a jointly optimised molecular–language foundation model for drug discovery.
The current system connects separately adapted molecular and language models. Our next research direction is to explore explicit cross-modal representation alignment, where molecular prediction and scientific reasoning can inform one another during learning.
We want to extend MutationScientist in four directions:
1. Joint molecular–language optimisation
Investigate shared latent representations and training objectives that connect molecular properties with scientific reasoning, rather than only combining their outputs.
2. Mutation-aware molecular generation
Explore candidate generation using chemical representations such as SMILES, followed by molecular evaluation and scientific interpretation to prioritise compounds for experimental testing.
3. Counterfactual resistance analysis
Ask what happens when a mutation changes, an inhibitor is modified, or a different resistance mechanism emerges, then use molecular predictions and retrieved evidence to investigate the possibilities.
4. Independent scientific validation
Expand our evaluations to experimentally grounded molecular benchmarks, independently reviewed scientific answers, and mutation-specific inhibitor comparisons.
Ultimately, we envision a scientific AI system that can predict molecular behaviour, reason about biochemical mechanisms, and help scientists decide which hypotheses and candidate molecules deserve investigation next.
MutationScientist is our first step toward bridging the language of molecules with the language of scientific discovery.