Ashish Makani / engineer & researcher / India

Machine learning.
Medicine.
Biology.

Building tools at the intersection of ML and medicine.

Explore the projects ↗A little about me ↓
tp53(ashish)tp53 / named after the tumor suppressor gene

Hi, I’m Ashish

I go by tp53 online, named after the tumor suppressor gene that’s central to cancer research. To learn more about tp53, I highly recommend this very accessible (even to laymen!) substack post by Stetson Thacker on tp53, the tumor suppressor to rule them all, often also christened “the guardian of the genome”.

I’m an engineer passionate about accelerating drug discovery and building tools at the intersection of machine learning and medicine. My focus is on cancer and oncology - from genomics and protein design to clinical trials and treatment planning.

What I’m working on

My mission is to impact medicine and disease through technology:

Building tools for drug discovery - Developing systems to accelerate the discovery of new therapeutics, especially for cancers where good treatments don’t exist today.

Building tools for healthcare - Using AI and biomedical data to help clinicians make better decisions faster, and to help patients become more engaged in their care.

See my projects for details on computational protein design, bio ML challenges, and regulatory analysis work.

News

Feb 19, 2025

Our project Onco-Seg (with Dr. Anurag Agrawal & Dr. Anjali Agrawal) and Ashoka University were mentioned by Alexandr Wang, head of Meta Superintelligence Labs, during his closed-door keynote at the India AI Impact Summit. That was pretty surreal to watch. Clip · Post

Feb 18, 2025

The Virtual Tumor Board was covered by the Times of India in their Times Techies: AI for Social Impact feature. Grateful for the coverage—it's encouraging to see interest in using AI for cancer care access.

New — Jun 2026
Preprint v1 — Results

RFantibody vs Biohub ESMFold2: A Neutral-Oracle Face-off

De novo binder design head-to-head — VHHs vs ESMC/ESMFold2 minibinders & scFvs

Under ESMFold2, Biohub's binders crush our RFantibody VHHs on all 10 shared cancer targets (median ipTM 0.85 vs 0.13) — but that is largely evaluation circularity: ESMFold2 is Biohub's own design objective. Re-scored with a neutral Boltz-2 oracle that neither platform optimized against, the gap halves (RFantibody recovers 0.13→0.37; Biohub falls to 0.69–0.74), though Biohub still leads — modestly — on 10/10.

14
antigens
224
Biohub designs
Boltz-2
neutral oracle
10/10
Biohub leads
View Results & Preprint →
New
Open Source Tool

rfab-harness

Campaign Orchestration for RFAntibody

One YAML config, one CLI command, full antibody design campaign. Wraps the 3-stage RFAntibody pipeline (RFdiffusion → ProteinMPNN → RF2) with target prep, multi-GPU parallelization, automated filtering, and 21 pre-built configs for cancer & rare disease targets.

View Project →
New — Feb 2026
Preprint v3 — Results

Dogfooding rfab-harness on 10 Cancer Targets

De novo VHH nanobody design against MPNST, DIPG, neuroblastoma, GBM, and pancreatic cancer

We pointed rfab-harness at 10 of the hardest cancer targets with no effective antibody therapy. 4,085 designs scored across 10 parallel GPU campaigns in ~75 minutes. 135 candidates survived stringent dual-threshold filtering. Pass rates varied 60-fold — from 0.3% (EGFRvIII) to 19.8% (CEACAM5) — revealing that computational design difficulty is wildly target-dependent.

4,085
designs scored
135
passed filters
3.3%
global pass rate
60×
range in difficulty
View Results & Preprint →
New
AI + Oncology

Onco-TTT

AI-Powered Cancer Hypothesis Generation

Type a plain-English oncology question, get: zero-shot NER (GLiNER2) → knowledge graph → literature → 6-dimensional validation (DepMap, cBioPortal, GTEx, ClinicalTrials.gov). 13 interactive views.

View Project →
New
Analysis

DADB-v1.0 Benchmark

A Therapeutic Decathlon for De Novo Antibody Design

The first comprehensive benchmark measuring what actually matters: binding, structure, developability, and immunogenicity. Featuring 5 AI platforms with first-ever human immunogenicity data.

🆕 Available Online: 29 January 2026

Read Proposal →
Featured
AI + Healthcare

Virtual Tumor Board

7 AI Specialists. 256 Guidelines. One Consensus.

Multi-agent AI that simulates a real tumor board meeting. Surgical, medical, radiation oncology, palliative care, radiology, pathology, and genetics specialists deliberate on cancer cases—grounded in NCCN, ESMO, ASTRO guidelines with Indian healthcare context.

Try Live Demo →
New
Research

Does Developability Come For Free?

FLAb Dataset Analysis: Testing Emergent Developability in De Novo Antibody Design

Analysis of 160 FLAb datasets revealing how training data bias explains "free" developability. ESM-2 benchmarks, immunogenicity prediction (0.74 AUROC), and why humanness alone fails.

View Analysis →
Analysis

Antibodies from Thin Air

Four AI Platforms Rewriting Cancer Drug Discovery

A comparison of JAM-2, Chai-2, Origin-1, and RFAntibody—de novo antibody design platforms achieving 15-40% hit rates and sub-nanomolar GPCR binders. Includes oncology target analysis and a primer on antibody biology.

Read Analysis →

View all antibody research →

New

Onco-Seg

Adapting SAM3 for Medical Image Segmentation

Multi-modal AI for tumor & organ segmentation across CT, MRI, ultrasound, and 5 more modalities. Trained on 98K+ cases using parameter-efficient LoRA fine-tuning.

View Project →
Onco-Seg demo

Background

I trained as an engineer during undergrad and got my Master’s in Computer Engineering at Purdue. Over the past several years, personal experiences and encounters with patients and doctors have inspired me to pursue a career at the intersection of tech and medicine.

Check out my curated reading list on cancer research, drug discovery, and rare diseases.

When I’m not working, I enjoy long walks, chai, cycling, standup comedy, and spending time with family.