From genome
to intelligence.
We apply frontier AI models to interpret the regulatory architecture of the genome, the vast majority of DNA that does not code for protein, yet governs nearly everything a cell does.
Research
Open Questions
Active areas of inquiry. Each investigation begins with a defined hypothesis and a bounded dataset.
Genomic regulation and protein structure remain, by the field's own admission, incompletely mapped. The questions below represent specific, answerable gaps within that space, not exploratory browsing. Each is pursued to the point of a defensible result.
Silent variants, reconsidered
Variants classified as biologically silent are assumed to leave no functional trace. We are testing whether that classification holds, or whether it predates tools now capable of detecting a regulatory footprint these variants do leave.
Regulatory and structural alignment
When a single variant alters both a gene's expression and the shape of its resulting protein, do the two effects move together, or can one be severe while the other is negligible? We are measuring the correlation directly.
Shared regulatory grammar across unrelated genes
Genes with no known functional relationship may still share recurring patterns in how their control regions are structured. This question has not been systematically tested because no existing framework compares regulatory architecture across unrelated genes at scale.
Generalization across underrepresented populations
Genomic prediction models are trained predominantly on well-represented reference data. We are examining how reliably these models generalize to genetic variation from populations underrepresented in that training data.
Model Releases
Closed-form and deep learning models developed by Afrium BioIntelligence, evaluated on standardized public benchmarks.
AminoGrapes
Structure-Weighted Evolutionary Scoring for Viral Protein Fitness Prediction
Zero-shot fitness scoring combining sequence-weighted multiple sequence alignments with AlphaFold2 residue solvent burial (Burial × Evolutionary Log-Odds). Outperforms Meta's ESM2-650M on 25 of 31 viral assays in ProteinGym v1.3 (0.430 vs 0.269 mean Spearman, +0.162 gain, p = 0.0008).
The Lab
The Lab
Computational genomics, structured around specific inquiry.
Core Focus
What the lab does
Afrium BioIntelligence Lab works at the intersection of two classes of genomic AI model. One predicts how a change in DNA sequence propagates through gene regulation, altering expression, splicing, and chromatin accessibility, differently across cell and tissue types. The other predicts the three-dimensional structure a protein assumes once translated.
Used in sequence, these two prediction layers trace a single genetic variant from cause to consequence, from a base-pair change to its regulatory effect to the shape of the resulting protein. This is a path that, run experimentally, could occupy a laboratory for months to resolve for even one variant.
The lab operates remotely by design, contributors work from across Pakistan and internationally, coordinated as a single distributed research effort rather than a fixed physical lab.
Input Locus Variant
Single nucleotide polymorphism (SNP) or insertion/deletion identified in non-coding regulatory domain.
Layer 1 · Regulatory Dynamics (AlphaGenome)
Simultaneous inference across transcription factor binding, chromatin state, and alternative splicing profiles.
Layer 2 · Structural Conformation (AlphaFold)
Evaluation of downstream translated protein fold, atomic coordinates, and phenotypic stability consequences.
Interactive Locus Inference Demonstration
Computation & Scale
Infrastructure
The lab's work depends on a recent shift in what genomic computation can do. Where prediction once meant one model run per variant, current infrastructure allows regulatory and structural predictions to be generated and compared across enormous portions of the genome simultaneously.
Models gene expression, splicing, and chromatin effects directly from raw DNA sequence, across distinct cell and tissue contexts.
Predicts the three-dimensional fold a given amino acid sequence will assume.
Precomputed impact scores spanning the full space of possible single-letter genomic changes.
Research Protocol
Method
Each question begins with a defined hypothesis and a bounded dataset, never an open-ended search. Regulatory predictions are generated across the relevant sequence space, scored against the specific hypothesis, and where applicable, cross-checked against predicted structural outcomes for the same variants. Results that contradict the hypothesis are treated as findings in their own right, not discarded.
Work is documented as it progresses. A result is written up when it is defensible, not when it is convenient. Negative results are reported with the same rigor as positive ones.
Context & Foundation
About
Biological Ground Truth
The genome, briefly
Roughly 2% of the human genome directly encodes protein, sequences cells translate into the molecules that carry out biological work. The remaining 98% does not build anything directly. It regulates: switching genes on and off, tuning how much protein gets made, determining when and where this happens. Most genetic variation linked to disease and human traits lies within this regulatory space.
Once a protein is built, its function depends on the three-dimensional shape it folds into, assembled from a one-dimensional sequence in fractions of a second. A single altered fold can be the difference between a functioning protein and a harmful one.
Significance
Why this matters
Most of the genome's regulatory language remains unread. Every variant correctly interpreted, every regulatory pattern mapped, narrows that gap by a small, exact amount.
Institution
Origin
Afrium BioIntelligence Lab operates as part of Afrium, an undergraduate AI research organization based in Islamabad, Pakistan. The lab's work sits alongside Afrium's broader programs in applied AI research, extending that work into computational genomics specifically.
Track Record
Built on a track record
Afrium BioIntelligence Lab follows the completion of Afrium's Class of ML, Class of Research, and ASoC programs, run across 47+ universities nationwide within Afrium's first year. BioIntelligence Lab extends that foundation into computational genomics.