Hey 
I’m Sartiz.

Five of us built this for our undergraduate thesis. The problem is easy to say out loud. Protein language models will happily tell you what a protein probably does, but they will not tell you why, and a biologist cannot act on an answer they are not allowed to check.

So we gave the model something to reason from. We took the 250,000 proteins in Prot2Text, clustered them down to 87,384 with MMseqs2 so nothing in the set was a near copy of anything else, then attached an evidence pack to every entry: Pfam domains, DeepLoc localisation, DeepTMHMM topology, SignalP peptides. GPT OSS 20B turned each pack into four short reasoning steps with a confidence score, and a validator quietly binned anything that cited evidence the protein did not actually have.

Training runs twice. First a projector we call MSPP 128 learns to squeeze a frozen ESM 2 into 128 tokens a language model can actually read. Then LLaMA 3.1 8B learns, through QLoRA on two T4 cards, to write the reasoning first and the answer second.

My favourite result is the least glamorous one. Reasoning before answering beat answering alone, 0.777 to 0.821 BERT F1 across sixteen matched variants. And one clear paragraph beat a tidy four step chain, which is not what any of us had bet on.

Supervisor

Dr. Farig Yousuf Sadeque

Co-supervisor

Dr. Swakkhar Shatabda

Dataset

87,384 proteins

Where

BRAC University

Pancreatic ductal adenocarcinoma is usually found late and modelled thinly, one modality at a time. HydraNet gives each modality its own head: a 3D-CNN over 32×224×224 tumour slabs, a gated residual network over fifteen genomic principal components plus clinical fields.

Then, instead of concatenating them, the genomic profile queries the image. Scaled dot product attention decides which radiomic features matter given this patient’s molecular context. Three heads come out the other side: survival, necrosis percentage, tumour site, trained together so the auxiliary tasks regularise the hard one.

Twenty six patients. A saliency map for every prediction, because a number without a reason is not usable in a clinic.

Tumour site

83.3% accuracy

Necrosis

7.6% MAE

Survival

66.7% accuracy

Cohort

26 patients

803 engineering students, fourteen universities, 47 questions each. We scored everyone with PHQ 9, drew the line at ten, then deleted the PHQ 9 answers from the training data so the model could not simply read the answer key back to us.

Far fewer students landed on the depressed side, and that kind of imbalance quietly ruins a classifier. The usual fix is to generate the rows you are missing. What nobody agrees on is how to tell whether the rows you generated deserve to be believed, so we wrote a score for it: distribution similarity, correlation preservation, class balance, feature importance, privacy. SMOTE won on evidence rather than habit, at a KS statistic of 0.016 against CTGAN at 0.143.

Then eight classical models and three deep ones. XGBoost came out ahead at 96.25% accuracy. The leaderboard was never the point though. SHAP, LIME and ALE all pointed at the same handful of features, and a year later I went back and found out something about when we had collected the answers.

Best model

XGBoost · F1 0.9651

Explainability

SHAP · LIME · ALE

Supervisor

Annajiat Alim Rasel

Most clinical RL work hands the agent a reward function someone typed out on a Tuesday. This one learns the reward from expert judgement instead: pairwise trajectory preferences, fitted with a Bradley-Terry model, in a true dynamical environment where actions change the future state.

Then a Lagrangian penalty on top of PPO, so intervention frequency is a constraint rather than an afterthought. The interesting outcome isn’t the return. It’s that the policy refuses to collapse. No 0%. No 100%. It hesitates, which is what a good clinician does.

Preferences

1,989 pairs

In range (test)

59.7%

Intervention rate

54.2%

Method

Constrained PPO

I led the system architecture. Crop disease detection from photographs, fused with a yield and growth stage model, tuned with Bayesian optimisation, fed by NASA satellite APIs and more than ten thousand processed images.

The part I am still proud of is the last mile. You give it soil type, watering frequency, fertiliser, sunlight hours, temperature and humidity. It gives you back a growth stage, and then it turns that stage into a sentence a farmer can act on this week. A prediction nobody can read is not a decision support system.

NASA Space Apps

1 of 947 finalists

Field

9,900+ teams

Stack

PyTorch · FastAPI